Welcome to Foolbox

Foolbox is a Python toolbox to create adversarial examples that fool neural networks.

It comes with support for many frameworks to build models including

  • TensorFlow
  • PyTorch
  • Keras
  • JAX
  • MXNet
  • Theano
  • Lasagne

and it is easy to extend to other frameworks.

In addition, it comes with a large collection of adversarial attacks, both gradient-based attacks as well as black-box attacks. See foolbox.attacks for details.

The source code and a minimal working example can be found on GitHub.

Installation

Foolbox is a Python package to create adversarial examples. It supports Python 3.5 and newer (try Foolbox 1.x if you still need to use Python 2.7).

Stable release

You can install the latest stable release of Foolbox from PyPI using pip:

pip install foolbox

Make sure that pip installs packages for Python 3, otherwise you might need to use pip3 instead of pip.

Pre-release versions

You can install the latest stable release of Foolbox from PyPI using pip:

pip install foolbox --pre

Make sure that pip installs packages for Python 3, otherwise you might need to use pip3 instead of pip.

Development version

Alternatively, you can install the latest development version of Foolbox from GitHub. We try to keep the master branch stable, so this version should usually work fine. Feel free to open an issue on GitHub if you encounter any problems.

pip install https://github.com/bethgelab/foolbox/archive/master.zip

Contributing to Foolbox

If you would like to contribute the development of Foolbox, install it in editable mode:

git clone https://github.com/bethgelab/foolbox.git
cd foolbox
pip install --editable .

To contribute your changes, you will need to fork the Foolbox repository on GitHub. You can than add it as a remote:

git remote add fork git@github.com/<your-github-name>/foolbox.git

You can now commit your changes, push them to your fork and create a pull-request to contribute them to Foolbox. See Running Tests for more information on the necessary tools and conventions.

Tutorial

This tutorial will show you how an adversarial attack can be used to find adversarial examples for a model.

Creating a model

For the tutorial, we will target VGG19 implemented in TensorFlow, but it is straight forward to apply the same to other models or other frameworks such as Theano or PyTorch.

import tensorflow as tf

images = tf.placeholder(tf.float32, (None, 224, 224, 3))
preprocessed = vgg_preprocessing(images)
logits = vgg19(preprocessed)

To turn a model represented as a standard TensorFlow graph into a model that can be attacked by the Adversarial Toolbox, all we have to do is to create a new TensorFlowModel instance:

from foolbox.models import TensorFlowModel

model = TensorFlowModel(images, logits, bounds=(0, 255))

Specifying the criterion

To run an adversarial attack, we need to specify the type of adversarial we are looking for. This can be done using the Criterion class.

from foolbox.criteria import TargetClassProbability

target_class = 22
criterion = TargetClassProbability(target_class, p=0.99)

Running the attack

Finally, we can create and apply the attack:

from foolbox.attacks import LBFGSAttack

attack = LBFGSAttack(model, criterion)
images, labels = foolbox.utils.samples(dataset='imagenet', batchsize=16, data_format='channels_last', bounds=(0, 255))
adversarial = attack(image, label=label)

Visualizing the adversarial examples

To plot the adversarial example we can use matplotlib:

import matplotlib.pyplot as plt

plt.subplot(1, 3, 1)
plt.imshow(image)

plt.subplot(1, 3, 2)
plt.imshow(adversarial)

plt.subplot(1, 3, 3)
plt.imshow(adversarial - image)

External Resources

If you would like to share your Foolbox tutorial or example code, please let us know by opening an issue or pull-request on GitHub and we would be happy to add it to this list.

Examples

Here you can find a collection of examples how Foolbox models can be created using different deep learning frameworks and some full-blown attack examples at the end.

Running an attack

Running a batch attack against a PyTorch model

import foolbox
import numpy as np
import torchvision.models as models

# instantiate model (supports PyTorch, Keras, TensorFlow (Graph and Eager), MXNet and many more)
model = models.resnet18(pretrained=True).eval()
preprocessing = dict(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], axis=-3)
fmodel = foolbox.models.PyTorchModel(model, bounds=(0, 1), num_classes=1000, preprocessing=preprocessing)

# get a batch of images and labels and print the accuracy
images, labels = foolbox.utils.samples(dataset='imagenet', batchsize=16, data_format='channels_first', bounds=(0, 1))
print(np.mean(fmodel.forward(images).argmax(axis=-1) == labels))
# -> 0.9375

# apply the attack
attack = foolbox.attacks.FGSM(fmodel)
adversarials = attack(images, labels)
# if the i'th image is misclassfied without a perturbation, then adversarials[i] will be the same as images[i]
# if the attack fails to find an adversarial for the i'th image, then adversarials[i] will all be np.nan

# Foolbox guarantees that all returned adversarials are in fact in adversarials
print(np.mean(fmodel.forward(adversarials).argmax(axis=-1) == labels))
# -> 0.0

# ---

# In rare cases, it can happen that attacks return adversarials that are so close to the decision boundary,
# that they actually might end up on the other (correct) side if you pass them through the model again like
# above to get the adversarial class. This is because models are not numerically deterministic (on GPU, some
# operations such as `sum` are non-deterministic by default) and indepedent between samples (an input might
# be classified differently depending on the other inputs in the same batch).

# You can always get the actual adversarial class that was observed for that sample by Foolbox by
# passing `unpack=False` to get the actual `Adversarial` objects:
attack = foolbox.attacks.FGSM(fmodel, distance=foolbox.distances.Linf)
adversarials = attack(images, labels, unpack=False)

adversarial_classes = np.asarray([a.adversarial_class for a in adversarials])
print(labels)
print(adversarial_classes)
print(np.mean(adversarial_classes == labels))  # will always be 0.0

# The `Adversarial` objects also provide a `distance` attribute. Note that the distances
# can be 0 (misclassified without perturbation) and inf (attack failed).
distances = np.asarray([a.distance.value for a in adversarials])
print("{:.1e}, {:.1e}, {:.1e}".format(distances.min(), np.median(distances), distances.max()))
print("{} of {} attacks failed".format(sum(adv.distance.value == np.inf for adv in adversarials), len(adversarials)))
print("{} of {} inputs misclassified without perturbation".format(sum(adv.distance.value == 0 for adv in adversarials), len(adversarials)))

Running an attack on single sample against a Keras model

import foolbox
import keras
import numpy as np
from keras.applications.resnet50 import ResNet50

# instantiate model
keras.backend.set_learning_phase(0)
kmodel = ResNet50(weights='imagenet')
preprocessing = dict(flip_axis=-1, mean=np.array([104, 116, 123]))  # RGB to BGR and mean subtraction
fmodel = foolbox.models.KerasModel(kmodel, bounds=(0, 255), preprocessing=preprocessing)

# get source image and label
image, label = foolbox.utils.imagenet_example()

# apply attack on source image
attack = foolbox.v1.attacks.FGSM(fmodel)
adversarial = attack(image, label)
# if the attack fails, adversarial will be None and a warning will be printed

Creating a model

Keras: ResNet50

import keras
import numpy as np
import foolbox

keras.backend.set_learning_phase(0)
kmodel = keras.applications.resnet50.ResNet50(weights='imagenet')
preprocessing = dict(flip_axis=-1, mean=np.array([104, 116, 123]))  # RGB to BGR and mean subtraction
model = foolbox.models.KerasModel(kmodel, bounds=(0, 255), preprocessing=preprocessing)

image, label = foolbox.utils.imagenet_example()
print(np.argmax(model.forward_one(image)), label)

PyTorch: ResNet18

You might be interested in checking out the full PyTorch example at the end of this document.

import torchvision.models as models
import numpy as np
import foolbox

# instantiate the model
resnet18 = models.resnet18(pretrained=True).cuda().eval()  # for CPU, remove cuda()
mean = np.array([0.485, 0.456, 0.406]).reshape((3, 1, 1))
std = np.array([0.229, 0.224, 0.225]).reshape((3, 1, 1))
model = foolbox.models.PyTorchModel(resnet18, bounds=(0, 1), num_classes=1000, preprocessing=(mean, std))

image, label = foolbox.utils.imagenet_example(data_format='channels_first')
image = image / 255
print(np.argmax(model.forward_one(image)), label)

TensorFlow: VGG19

First, create the model in TensorFlow.

import tensorflow as tf
from tensorflow.contrib.slim.nets import vgg
import numpy as np
import foolbox

images = tf.placeholder(tf.float32, shape=(None, 224, 224, 3))
preprocessed = images - [123.68, 116.78, 103.94]
logits, _ = vgg.vgg_19(preprocessed, is_training=False)
restorer = tf.train.Saver(tf.trainable_variables())

image, _ = foolbox.utils.imagenet_example()

Then transform it into a Foolbox model using one of these four options:

Option 1

This option is recommended if you want to keep the code as short as possible. It makes use of the TensorFlow session created by Foolbox internally if no default session is set.

with foolbox.models.TensorFlowModel(images, logits, (0, 255)) as model:
    restorer.restore(model.session, '/path/to/vgg_19.ckpt')
    print(np.argmax(model.forward_one(image)))
Option 2

This option is recommended if you want to create the TensorFlow session yourself.

with tf.Session() as session:
    restorer.restore(session, '/path/to/vgg_19.ckpt')
    model = foolbox.models.TensorFlowModel(images, logits, (0, 255))
    print(np.argmax(model.forward_one(image)))
Option 3

This option is recommended if you want to avoid nesting context managers, e.g. during interactive development.

session = tf.InteractiveSession()
restorer.restore(session, '/path/to/vgg_19.ckpt')
model = foolbox.models.TensorFlowModel(images, logits, (0, 255))
print(np.argmax(model.forward_one(image)))
session.close()
Option 4

This is possible, but usually one of the other options should be preferred.

session = tf.Session()
with session.as_default():
    restorer.restore(session, '/path/to/vgg_19.ckpt')
    model = foolbox.models.TensorFlowModel(images, logits, (0, 255))
    print(np.argmax(model.forward_one(image)))
session.close()

Applying an attack

Once you created a Foolbox model (see the previous section), you can apply an attack.

FGSM (GradientSignAttack)

# create a model (see previous section)
fmodel = ...

# get source image and label
image, label = foolbox.utils.imagenet_example()

# apply attack on source image
attack  = foolbox.v1.attacks.FGSM(fmodel)
adversarial = attack(image, label)

Creating an untargeted adversarial for a PyTorch model

import foolbox
import torch
import torchvision.models as models
import numpy as np

# instantiate the model
resnet18 = models.resnet18(pretrained=True).eval()
if torch.cuda.is_available():
    resnet18 = resnet18.cuda()
mean = np.array([0.485, 0.456, 0.406]).reshape((3, 1, 1))
std = np.array([0.229, 0.224, 0.225]).reshape((3, 1, 1))
fmodel = foolbox.models.PyTorchModel(
    resnet18, bounds=(0, 1), num_classes=1000, preprocessing=(mean, std))

# get source image and label
image, label = foolbox.utils.imagenet_example(data_format='channels_first')
image = image / 255.  # because our model expects values in [0, 1]

print('label', label)
print('predicted class', np.argmax(fmodel.forward_one(image)))

# apply attack on source image
attack = foolbox.v1.attacks.FGSM(fmodel)
adversarial = attack(image, label)

print('adversarial class', np.argmax(fmodel.forward_one(adversarial)))

outputs

label 282
predicted class 282
adversarial class 281

To plot image and adversarial, don’t forget to move the channel axis to the end before passing them to matplotlib’s imshow, e.g. using np.transpose(image, (1, 2, 0)).

Creating a targeted adversarial for the Keras ResNet model

import foolbox
from foolbox.models import KerasModel
from foolbox.attacks import LBFGSAttack
from foolbox.criteria import TargetClassProbability
import numpy as np
import keras
from keras.applications.resnet50 import ResNet50
from keras.applications.resnet50 import preprocess_input
from keras.applications.resnet50 import decode_predictions

keras.backend.set_learning_phase(0)
kmodel = ResNet50(weights='imagenet')
preprocessing = dict(flip_axis=-1, mean=np.array([104, 116, 123]))  # RGB to BGR and mean subtraction
fmodel = KerasModel(kmodel, bounds=(0, 255), preprocessing=preprocessing)

image, label = foolbox.utils.imagenet_example()

# run the attack
attack = LBFGSAttack(model=fmodel, criterion=TargetClassProbability(781, p=.5))
adversarial = attack(image, label)

# show results
print(np.argmax(fmodel.forward_one(adversarial)))
print(foolbox.utils.softmax(fmodel.forward_one(adversarial))[781])
preds = kmodel.predict(preprocess_input(adversarial[np.newaxis].copy()))
print("Top 5 predictions (adversarial: ", decode_forward_one(preds, top=5))

outputs

781
0.832095
Top 5 predictions (adversarial:  [[('n04149813', 'scoreboard', 0.83013469), ('n03196217', 'digital_clock', 0.030192226), ('n04152593', 'screen', 0.016133979), ('n04141975', 'scale', 0.011708578), ('n03782006', 'monitor', 0.0091574294)]]

Advanced

The Adversarial class provides an advanced way to specify the adversarial example that should be found by an attack and provides detailed information about the created adversarial. In addition, it provides a way to improve a previously found adversarial example by re-running an attack.

from foolbox.v1 import Adversarial
from foolbox.v1.attacks import LBFGSAttack
from foolbox.models import TenosrFlowModel
from foolbox.criteria import TargetClassProbability

Implicit

model = TensorFlowModel(inputs, logits, bounds=(0, 255))
criterion = TargetClassProbability('ostrich', p=0.99)
attack = LBFGSAttack(model, criterion)

Running the attack by passing an input and a label will implicitly create an Adversarial instance. By passing unpack=False we tell the attack to return the Adversarial instance rather than a numpy array.

adversarial = attack(image, label=label, unpack=False)

We can then get the actual adversarial input using the image attribute:

adversarial_image = adversarial.perturbed

Explicit

model = TensorFlowModel(images, logits, bounds=(0, 255))
criterion = TargetClassProbability('ostrich', p=0.99)
attack = LBFGSAttack()

We can also create the Adversarial instance ourselves and then pass it to the attack.

adversarial = Adversarial(model, criterion, image, label)
attack(adversarial)

Again, we can get the image using the image attribute:

adversarial_image = adversarial.perturbed

This approach gives us more flexibility and allows us to specify a different distance measure:

distance = MeanAbsoluteDistance
adversarial = Adversarial(model, criterion, image, label, distance=distance)

Model Zoo

This tutorial will show you how the model zoo can be used to run your attack against a robust model.

Downloading a model

For this tutorial, we will download the Analysis by Synthesis model implemented in PyTorch and run a FGSM (GradienSignAttack) against it.

from foolbox import zoo

# download the model
model = zoo.get_model(url="https://github.com/bethgelab/AnalysisBySynthesis")

# read image and label
image = ...
label = ...

# apply attack on source image
attack  = foolbox.attacks.FGSM(model)
adversarial = attack(image, label)

Development

To install Foolbox in editable mode, see the installation instructions under Contributing to Foolbox.

Running Tests

pytest

To run the tests, you need to have pytest and pytest-cov installed. Afterwards, you can simply run pytest in the root folder of the project. Some tests will require TensorFlow, PyTorch and the other frameworks, so to run all tests, you need to have all of them installed. Note however that this can take quite long (Foolbox has many tests) and installing all frameworks with the correct versions is difficult due to conflicting dependencies. You can also open a pull-request and then we will run all the tests using travis.

Style Guide

We use Black to format all code in a consistent and PEP-8 conform way. All pull-requests are checked using both black and flake8. Simply install black and run black . after all your changes or ideally even on each commit using pre-commit.

New Adversarial Attacks

Foolbox makes it easy to develop new adversarial attacks that can be applied to arbitrary models.

To implement an attack, simply subclass the Attack class, implement the __call__() method and decorate it with the call_decorator(). The call_decorator() will make sure that your __call__() implementation will be called with an instance of the Adversarial class. You can use this instance to ask for model predictions and gradients, get the original image and its label and more. In addition, the Adversarial instance automatically keeps track of the best adversarial amongst all the inputs tested by the attack. That way, the implementation of the attack can focus on the attack logic.

To implement an attack that can make use of the batch support introduced in Foolbox 2.0, implement the as_generator() method and decorate it with the generator_decorator(). All model calls using the Adversarial object should use yield.

FAQ

How does Foolbox handle inputs that are misclassified without any perturbation?
The attacks will not be run and instead the unperturbed input is returned as an adversarial with distance 0 to the clean input.
What happens if an attack fails?
The attack will return None and the distance will be np.inf.
Why is the returned adversarial not misclassified by my model?
Most likely you have a discrepancy between how you evaluate your model and how you told Foolbox to evaluate it. For example, you might not be using the same preprocessing. Compare the output of the predictions method of the Foolbox model instance with your model’s output (logits). This problem can also be caused by non-deterministic models. Make sure that your model is not stochastic and always returns the same output when given the same input. In rare cases it can also be that a seemlingly deterministic model becomes numerically stochastic around the decision boundary (e.g. because of non-deterministic floating point reduce_sum operations). You can always check adversarial.output and adversarial.adversarial_class to see the output Foolbox got from your model when deciding that this was an adversarial.
Why are the gradients multiplied by the bounds (max_ - min_)?
This scaling is meant to make hyperparameters such as the epsilon for FGSM independent of the bounds. epsilon = 0.1 thus means that you perturb the input by 10% relative to the max - max range (which could for example go from 0 to 1 or from 0 to 255).

foolbox.models

Provides classes to wrap existing models in different framworks so that they provide a unified API to the attacks.

Models

Model Base class to provide attacks with a unified interface to models.
DifferentiableModel Base class for differentiable models.
TensorFlowModel Creates a Model instance from existing TensorFlow tensors.
TensorFlowEagerModel Creates a Model instance from a TensorFlow model using eager execution.
PyTorchModel Creates a Model instance from a PyTorch module.
KerasModel Creates a Model instance from a Keras model.
TheanoModel Creates a Model instance from existing Theano tensors.
LasagneModel Creates a Model instance from a Lasagne network.
MXNetModel Creates a Model instance from existing MXNet symbols and weights.
MXNetGluonModel Creates a Model instance from an existing MXNet Gluon Block.
JAXModel Creates a Model instance from a JAX predict function.
CaffeModel

Wrappers

ModelWrapper Base class for models that wrap other models.
DifferentiableModelWrapper Base class for models that wrap other models and provide gradient methods.
ModelWithoutGradients Turns a model into a model without gradients.
ModelWithEstimatedGradients Turns a model into a model with gradients estimated by the given gradient estimator.
CompositeModel Combines predictions of a (black-box) model with the gradient of a (substitute) model.
EnsembleAveragedModel Reduces stochastic effects in networks by averaging both forward and backward

Detailed description

class foolbox.models.Model(bounds, channel_axis, preprocessing=(0, 1))[source]

Base class to provide attacks with a unified interface to models.

The Model class represents a model and provides a unified interface to its predictions. Subclasses must implement forward and num_classes.

Model instances can be used as context managers and subclasses can require this to allocate and release resources.

Parameters:
bounds : tuple

Tuple of lower and upper bound for the pixel values, usually (0, 1) or (0, 255).

channel_axis : int

The index of the axis that represents color channels.

preprocessing: dict or tuple

Can be a tuple with two elements representing mean and standard deviation or a dict with keys “mean” and “std”. The two elements should be floats or numpy arrays. “mean” is subtracted from the input, the result is then divided by “std”. If “mean” and “std” are 1-dimensional arrays, an additional (negative) “axis” key can be given such that “mean” and “std” will be broadcasted to that axis (typically -1 for “channels_last” and -3 for “channels_first”, but might be different when using e.g. 1D convolutions). Finally, a (negative) “flip_axis” can be specified. This axis will be flipped (before “mean” is subtracted), e.g. to convert RGB to BGR.
forward(self, inputs)[source]

Takes a batch of inputs and returns the logits predicted by the underlying model.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

See also

forward_one()
forward_one(self, x)[source]

Takes a single input and returns the logits predicted by the underlying model.

Parameters:
x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension).
Returns:
numpy.ndarray

Predicted logits with shape (number of classes,).

See also

forward()
num_classes(self)[source]

Determines the number of classes.

Returns:
int

The number of classes for which the model creates predictions.
class foolbox.models.DifferentiableModel(bounds, channel_axis, preprocessing=(0, 1))[source]

Base class for differentiable models.

The DifferentiableModel class can be used as a base class for models that can support gradient backpropagation. Subclasses must implement gradient and backward.

A differentiable model does not necessarily provide reasonable values for the gradient, the gradient can be wrong. It only guarantees that the relevant methods can be called.

backward(self, gradient, inputs)[source]

Backpropagates the gradient of some loss w.r.t. the logits through the underlying model and returns the gradient of that loss w.r.t to the inputs.

Parameters:
gradient : numpy.ndarray

Gradient of some loss w.r.t. the logits with shape (batch size, number of classes).

inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

The gradient of the respective loss w.r.t the inputs.

See also

backward_one()
gradient()
backward_one(self, gradient, x)[source]

Backpropagates the gradient of some loss w.r.t. the logits through the underlying model and returns the gradient of that loss w.r.t to the input.

Parameters:
gradient : numpy.ndarray

Gradient of some loss w.r.t. the logits with shape (number of classes,).

x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension).
Returns:
numpy.ndarray

The gradient of the respective loss w.r.t the input.

See also

backward()
forward_and_gradient(self, inputs, labels)[source]

Takes inputs and labels and returns both the logits predicted by the underlying model and the gradients of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Inputs with shape as expected by the model (with the batch dimension).

labels : numpy.ndarray

Array of the class label of the inputs as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
forward_and_gradient_one(self, x, label)[source]

Takes a single input and label and returns both the logits predicted by the underlying model and the gradient of the cross-entropy loss w.r.t. the input.

Defaults to individual calls to forward_one and gradient_one but can be overriden by subclasses to provide a more efficient implementation.

Parameters:
x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension).

label : int

Class label of the input as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
gradient(self, inputs, labels)[source]

Takes a batch of inputs and labels and returns the gradient of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).
Returns:
gradient : numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the inputs.

See also

gradient_one()
backward()
gradient_one(self, x, label)[source]

Takes a single input and label and returns the gradient of the cross-entropy loss w.r.t. the input.

Parameters:
x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension).

label : int

Class label of the input as an integer in [0, number of classes).
Returns:
numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

gradient()
class foolbox.models.TensorFlowModel(inputs, logits, bounds, channel_axis=3, preprocessing=(0, 1))[source]

Creates a Model instance from existing TensorFlow tensors.

Parameters:
inputs : tensorflow.Tensor

The input to the model, usually a tensorflow.placeholder.

logits : tensorflow.Tensor

The predictions of the model, before the softmax.

bounds : tuple

Tuple of lower and upper bound for the pixel values, usually (0, 1) or (0, 255).

channel_axis : int

The index of the axis that represents color channels.

preprocessing: dict or tuple

Can be a tuple with two elements representing mean and standard deviation or a dict with keys “mean” and “std”. The two elements should be floats or numpy arrays. “mean” is subtracted from the input, the result is then divided by “std”. If “mean” and “std” are 1-dimensional arrays, an additional (negative) “axis” key can be given such that “mean” and “std” will be broadcasted to that axis (typically -1 for “channels_last” and -3 for “channels_first”, but might be different when using e.g. 1D convolutions). Finally, a (negative) “flip_axis” can be specified. This axis will be flipped (before “mean” is subtracted), e.g. to convert RGB to BGR.
backward(self, gradient, inputs)[source]

Backpropagates the gradient of some loss w.r.t. the logits through the underlying model and returns the gradient of that loss w.r.t to the inputs.

Parameters:
gradient : numpy.ndarray

Gradient of some loss w.r.t. the logits with shape (batch size, number of classes).

inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

The gradient of the respective loss w.r.t the inputs.

See also

backward_one()
gradient()
forward(self, inputs)[source]

Takes a batch of inputs and returns the logits predicted by the underlying model.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

See also

forward_one()
forward_and_gradient(self, inputs, labels)[source]

Takes inputs and labels and returns both the logits predicted by the underlying model and the gradients of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Inputs with shape as expected by the model (with the batch dimension).

labels : numpy.ndarray

Array of the class label of the inputs as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
forward_and_gradient_one(self, x, label)[source]

Takes a single input and label and returns both the logits predicted by the underlying model and the gradient of the cross-entropy loss w.r.t. the input.

Defaults to individual calls to forward_one and gradient_one but can be overriden by subclasses to provide a more efficient implementation.

Parameters:
x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension).

label : int

Class label of the input as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
classmethod from_keras(model, bounds, input_shape=None, channel_axis='auto', preprocessing=(0, 1))[source]

Alternative constructor for a TensorFlowModel that accepts a tf.keras.Model instance.

Parameters:
model : tensorflow.keras.Model

A tensorflow.keras.Model that accepts a single input tensor and returns a single output tensor representing logits.

bounds : tuple

Tuple of lower and upper bound for the pixel values, usually (0, 1) or (0, 255).

input_shape : tuple

The shape of a single input, e.g. (28, 28, 1) for MNIST. If None, tries to get the the shape from the model’s input_shape attribute.

channel_axis : int or ‘auto’

The index of the axis that represents color channels. If ‘auto’, will be set automatically based on keras.backend.image_data_format()

preprocessing: dict or tuple

Can be a tuple with two elements representing mean and standard deviation or a dict with keys “mean” and “std”. The two elements should be floats or numpy arrays. “mean” is subtracted from the input, the result is then divided by “std”. If “mean” and “std” are 1-dimensional arrays, an additional (negative) “axis” key can be given such that “mean” and “std” will be broadcasted to that axis (typically -1 for “channels_last” and -3 for “channels_first”, but might be different when using e.g. 1D convolutions). Finally, a (negative) “flip_axis” can be specified. This axis will be flipped (before “mean” is subtracted), e.g. to convert RGB to BGR.
gradient(self, inputs, labels)[source]

Takes a batch of inputs and labels and returns the gradient of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).
Returns:
gradient : numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the inputs.

See also

gradient_one()
backward()
num_classes(self)[source]

Determines the number of classes.

Returns:
int

The number of classes for which the model creates predictions.
class foolbox.models.TensorFlowEagerModel(model, bounds, num_classes=None, channel_axis=3, preprocessing=(0, 1))[source]

Creates a Model instance from a TensorFlow model using eager execution.

Parameters:
model : a TensorFlow eager model

The TensorFlow eager model that should be attacked. It will be called with input tensors and should return logits.

bounds : tuple

Tuple of lower and upper bound for the pixel values, usually (0, 1) or (0, 255).

num_classes : int

If None, will try to infer it from the model’s output shape.

channel_axis : int

The index of the axis that represents color channels.

preprocessing: dict or tuple

Can be a tuple with two elements representing mean and standard deviation or a dict with keys “mean” and “std”. The two elements should be floats or numpy arrays. “mean” is subtracted from the input, the result is then divided by “std”. If “mean” and “std” are 1-dimensional arrays, an additional (negative) “axis” key can be given such that “mean” and “std” will be broadcasted to that axis (typically -1 for “channels_last” and -3 for “channels_first”, but might be different when using e.g. 1D convolutions). Finally, a (negative) “flip_axis” can be specified. This axis will be flipped (before “mean” is subtracted), e.g. to convert RGB to BGR.
backward(self, gradient, inputs)[source]

Backpropagates the gradient of some loss w.r.t. the logits through the underlying model and returns the gradient of that loss w.r.t to the inputs.

Parameters:
gradient : numpy.ndarray

Gradient of some loss w.r.t. the logits with shape (batch size, number of classes).

inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

The gradient of the respective loss w.r.t the inputs.

See also

backward_one()
gradient()
forward(self, inputs)[source]

Takes a batch of inputs and returns the logits predicted by the underlying model.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

See also

forward_one()
forward_and_gradient(self, inputs, labels)[source]

Takes inputs and labels and returns both the logits predicted by the underlying model and the gradients of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Inputs with shape as expected by the model (with the batch dimension).

labels : numpy.ndarray

Array of the class label of the inputs as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
forward_and_gradient_one(self, x, label)[source]

Takes a single input and label and returns both the logits predicted by the underlying model and the gradient of the cross-entropy loss w.r.t. the input.

Defaults to individual calls to forward_one and gradient_one but can be overriden by subclasses to provide a more efficient implementation.

Parameters:
x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension).

label : int

Class label of the input as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
gradient(self, inputs, labels)[source]

Takes a batch of inputs and labels and returns the gradient of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).
Returns:
gradient : numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the inputs.

See also

gradient_one()
backward()
num_classes(self)[source]

Determines the number of classes.

Returns:
int

The number of classes for which the model creates predictions.
class foolbox.models.PyTorchModel(model, bounds, num_classes, channel_axis=1, device=None, preprocessing=(0, 1))[source]

Creates a Model instance from a PyTorch module.

Parameters:
model : torch.nn.Module

The PyTorch model that should be attacked. It should predict logits or log-probabilities, i.e. predictions without the softmax.

bounds : tuple

Tuple of lower and upper bound for the pixel values, usually (0, 1) or (0, 255).

num_classes : int

Number of classes for which the model will output predictions.

channel_axis : int

The index of the axis that represents color channels.

device : string

A string specifying the device to do computation on. If None, will default to “cuda:0” if torch.cuda.is_available() or “cpu” if not.

preprocessing: dict or tuple

Can be a tuple with two elements representing mean and standard deviation or a dict with keys “mean” and “std”. The two elements should be floats or numpy arrays. “mean” is subtracted from the input, the result is then divided by “std”. If “mean” and “std” are 1-dimensional arrays, an additional (negative) “axis” key can be given such that “mean” and “std” will be broadcasted to that axis (typically -1 for “channels_last” and -3 for “channels_first”, but might be different when using e.g. 1D convolutions). Finally, a (negative) “flip_axis” can be specified. This axis will be flipped (before “mean” is subtracted), e.g. to convert RGB to BGR.
backward(self, gradient, inputs)[source]

Backpropagates the gradient of some loss w.r.t. the logits through the underlying model and returns the gradient of that loss w.r.t to the inputs.

Parameters:
gradient : numpy.ndarray

Gradient of some loss w.r.t. the logits with shape (batch size, number of classes).

inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

The gradient of the respective loss w.r.t the inputs.

See also

backward_one()
gradient()
forward(self, inputs)[source]

Takes a batch of inputs and returns the logits predicted by the underlying model.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

See also

forward_one()
forward_and_gradient(self, inputs, labels)[source]

Takes inputs and labels and returns both the logits predicted by the underlying model and the gradients of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Inputs with shape as expected by the model (with the batch dimension).

labels : numpy.ndarray

Array of the class label of the inputs as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
forward_and_gradient_one(self, x, label)[source]

Takes a single input and label and returns both the logits predicted by the underlying model and the gradient of the cross-entropy loss w.r.t. the input.

Defaults to individual calls to forward_one and gradient_one but can be overriden by subclasses to provide a more efficient implementation.

Parameters:
x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension).

label : int

Class label of the input as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
gradient(self, inputs, labels)[source]

Takes a batch of inputs and labels and returns the gradient of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).
Returns:
gradient : numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the inputs.

See also

gradient_one()
backward()
num_classes(self)[source]

Determines the number of classes.

Returns:
int

The number of classes for which the model creates predictions.
class foolbox.models.JAXModel(predict, bounds, num_classes, channel_axis=3, preprocessing=(0, 1))[source]

Creates a Model instance from a JAX predict function.

Parameters:
predict : function

The JAX-compatible function that takes a batch of inputs as and returns a batch of predictions (logits); use functools.partial(predict, params) to pass params if necessary

bounds : tuple

Tuple of lower and upper bound for the pixel values, usually (0, 1) or (0, 255).

num_classes : int

Number of classes for which the model will output predictions.

channel_axis : int

The index of the axis that represents color channels.

preprocessing: dict or tuple

Can be a tuple with two elements representing mean and standard deviation or a dict with keys “mean” and “std”. The two elements should be floats or numpy arrays. “mean” is subtracted from the input, the result is then divided by “std”. If “mean” and “std” are 1-dimensional arrays, an additional (negative) “axis” key can be given such that “mean” and “std” will be broadcasted to that axis (typically -1 for “channels_last” and -3 for “channels_first”, but might be different when using e.g. 1D convolutions). Finally, a (negative) “flip_axis” can be specified. This axis will be flipped (before “mean” is subtracted), e.g. to convert RGB to BGR.
backward(self, gradient, inputs)[source]

Backpropagates the gradient of some loss w.r.t. the logits through the underlying model and returns the gradient of that loss w.r.t to the inputs.

Parameters:
gradient : numpy.ndarray

Gradient of some loss w.r.t. the logits with shape (batch size, number of classes).

inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

The gradient of the respective loss w.r.t the inputs.

See also

backward_one()
gradient()
forward(self, inputs)[source]

Takes a batch of inputs and returns the logits predicted by the underlying model.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

See also

forward_one()
forward_and_gradient(self, inputs, labels)[source]

Takes inputs and labels and returns both the logits predicted by the underlying model and the gradients of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Inputs with shape as expected by the model (with the batch dimension).

labels : numpy.ndarray

Array of the class label of the inputs as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
gradient(self, inputs, labels)[source]

Takes a batch of inputs and labels and returns the gradient of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).
Returns:
gradient : numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the inputs.

See also

gradient_one()
backward()
num_classes(self)[source]

Determines the number of classes.

Returns:
int

The number of classes for which the model creates predictions.
class foolbox.models.KerasModel(model, bounds, channel_axis='auto', preprocessing=(0, 1), predicts='probabilities')[source]

Creates a Model instance from a Keras model.

Parameters:
model : keras.models.Model

The Keras model that should be attacked.

bounds : tuple

Tuple of lower and upper bound for the pixel values, usually (0, 1) or (0, 255).

channel_axis : int or ‘auto’

The index of the axis that represents color channels. If ‘auto’, will be set automatically based on keras.backend.image_data_format()

preprocessing: dict or tuple

Can be a tuple with two elements representing mean and standard deviation or a dict with keys “mean” and “std”. The two elements should be floats or numpy arrays. “mean” is subtracted from the input, the result is then divided by “std”. If “mean” and “std” are 1-dimensional arrays, an additional (negative) “axis” key can be given such that “mean” and “std” will be broadcasted to that axis (typically -1 for “channels_last” and -3 for “channels_first”, but might be different when using e.g. 1D convolutions). Finally, a (negative) “flip_axis” can be specified. This axis will be flipped (before “mean” is subtracted), e.g. to convert RGB to BGR.

predicts : str

Specifies whether the Keras model predicts logits or probabilities. Logits are preferred, but probabilities are the default.
backward(self, gradient, inputs)[source]

Backpropagates the gradient of some loss w.r.t. the logits through the underlying model and returns the gradient of that loss w.r.t to the inputs.

Parameters:
gradient : numpy.ndarray

Gradient of some loss w.r.t. the logits with shape (batch size, number of classes).

inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

The gradient of the respective loss w.r.t the inputs.

See also

backward_one()
gradient()
forward(self, inputs)[source]

Takes a batch of inputs and returns the logits predicted by the underlying model.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

See also

forward_one()
forward_and_gradient(self, inputs, labels)[source]

Takes inputs and labels and returns both the logits predicted by the underlying model and the gradients of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Inputs with shape as expected by the model (with the batch dimension).

labels : numpy.ndarray

Array of the class label of the inputs as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
forward_and_gradient_one(self, x, label)[source]

Takes a single input and label and returns both the logits predicted by the underlying model and the gradient of the cross-entropy loss w.r.t. the input.

Defaults to individual calls to forward_one and gradient_one but can be overriden by subclasses to provide a more efficient implementation.

Parameters:
x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension).

label : int

Class label of the input as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
gradient(self, inputs, labels)[source]

Takes a batch of inputs and labels and returns the gradient of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).
Returns:
gradient : numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the inputs.

See also

gradient_one()
backward()
num_classes(self)[source]

Determines the number of classes.

Returns:
int

The number of classes for which the model creates predictions.
class foolbox.models.TheanoModel(inputs, logits, bounds, num_classes, channel_axis=1, preprocessing=[0, 1])[source]

Creates a Model instance from existing Theano tensors.

Parameters:
inputs : theano.tensor

The input to the model.

logits : theano.tensor

The predictions of the model, before the softmax.

bounds : tuple

Tuple of lower and upper bound for the pixel values, usually (0, 1) or (0, 255).

num_classes : int

Number of classes for which the model will output predictions.

channel_axis : int

The index of the axis that represents color channels.

preprocessing: dict or tuple

Can be a tuple with two elements representing mean and standard deviation or a dict with keys “mean” and “std”. The two elements should be floats or numpy arrays. “mean” is subtracted from the input, the result is then divided by “std”. If “mean” and “std” are 1-dimensional arrays, an additional (negative) “axis” key can be given such that “mean” and “std” will be broadcasted to that axis (typically -1 for “channels_last” and -3 for “channels_first”, but might be different when using e.g. 1D convolutions). Finally, a (negative) “flip_axis” can be specified. This axis will be flipped (before “mean” is subtracted), e.g. to convert RGB to BGR.
backward(self, gradient, inputs)[source]

Backpropagates the gradient of some loss w.r.t. the logits through the underlying model and returns the gradient of that loss w.r.t to the inputs.

Parameters:
gradient : numpy.ndarray

Gradient of some loss w.r.t. the logits with shape (batch size, number of classes).

inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

The gradient of the respective loss w.r.t the inputs.

See also

backward_one()
gradient()
forward(self, inputs)[source]

Takes a batch of inputs and returns the logits predicted by the underlying model.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

See also

forward_one()
forward_and_gradient(self, inputs, labels)[source]

Takes inputs and labels and returns both the logits predicted by the underlying model and the gradients of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Inputs with shape as expected by the model (with the batch dimension).

labels : numpy.ndarray

Array of the class label of the inputs as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
forward_and_gradient_one(self, x, label)[source]

Takes a single input and label and returns both the logits predicted by the underlying model and the gradient of the cross-entropy loss w.r.t. the input.

Defaults to individual calls to forward_one and gradient_one but can be overriden by subclasses to provide a more efficient implementation.

Parameters:
x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension).

label : int

Class label of the input as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
gradient(self, inputs, labels)[source]

Takes a batch of inputs and labels and returns the gradient of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).
Returns:
gradient : numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the inputs.

See also

gradient_one()
backward()
num_classes(self)[source]

Determines the number of classes.

Returns:
int

The number of classes for which the model creates predictions.
class foolbox.models.LasagneModel(input_layer, logits_layer, bounds, channel_axis=1, preprocessing=(0, 1))[source]

Creates a Model instance from a Lasagne network.

Parameters:
input_layer : lasagne.layers.Layer

The input to the model.

logits_layer : lasagne.layers.Layer

The output of the model, before the softmax.

bounds : tuple

Tuple of lower and upper bound for the pixel values, usually (0, 1) or (0, 255).

channel_axis : int

The index of the axis that represents color channels.

preprocessing: dict or tuple

Can be a tuple with two elements representing mean and standard deviation or a dict with keys “mean” and “std”. The two elements should be floats or numpy arrays. “mean” is subtracted from the input, the result is then divided by “std”. If “mean” and “std” are 1-dimensional arrays, an additional (negative) “axis” key can be given such that “mean” and “std” will be broadcasted to that axis (typically -1 for “channels_last” and -3 for “channels_first”, but might be different when using e.g. 1D convolutions). Finally, a (negative) “flip_axis” can be specified. This axis will be flipped (before “mean” is subtracted), e.g. to convert RGB to BGR.
class foolbox.models.MXNetModel(data, logits, args, ctx, num_classes, bounds, channel_axis=1, aux_states=None, preprocessing=(0, 1))[source]

Creates a Model instance from existing MXNet symbols and weights.

Parameters:
data : mxnet.symbol.Variable

The input to the model.

logits : mxnet.symbol.Symbol

The predictions of the model, before the softmax.

args : dictionary mapping str to mxnet.nd.array

The parameters of the model.

ctx : mxnet.context.Context

The device, e.g. mxnet.cpu() or mxnet.gpu().

num_classes : int

The number of classes.

bounds : tuple

Tuple of lower and upper bound for the pixel values, usually (0, 1) or (0, 255).

channel_axis : int

The index of the axis that represents color channels.

aux_states : dictionary mapping str to mxnet.nd.array

The states of auxiliary parameters of the model.

preprocessing: dict or tuple

Can be a tuple with two elements representing mean and standard deviation or a dict with keys “mean” and “std”. The two elements should be floats or numpy arrays. “mean” is subtracted from the input, the result is then divided by “std”. If “mean” and “std” are 1-dimensional arrays, an additional (negative) “axis” key can be given such that “mean” and “std” will be broadcasted to that axis (typically -1 for “channels_last” and -3 for “channels_first”, but might be different when using e.g. 1D convolutions). Finally, a (negative) “flip_axis” can be specified. This axis will be flipped (before “mean” is subtracted), e.g. to convert RGB to BGR.
backward(self, gradient, inputs)[source]

Backpropagates the gradient of some loss w.r.t. the logits through the underlying model and returns the gradient of that loss w.r.t to the inputs.

Parameters:
gradient : numpy.ndarray

Gradient of some loss w.r.t. the logits with shape (batch size, number of classes).

inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

The gradient of the respective loss w.r.t the inputs.

See also

backward_one()
gradient()
forward(self, inputs)[source]

Takes a batch of inputs and returns the logits predicted by the underlying model.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

See also

forward_one()
forward_and_gradient(self, inputs, labels)[source]

Takes inputs and labels and returns both the logits predicted by the underlying model and the gradients of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Inputs with shape as expected by the model (with the batch dimension).

labels : numpy.ndarray

Array of the class label of the inputs as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
forward_and_gradient_one(self, x, label)[source]

Takes a single input and label and returns both the logits predicted by the underlying model and the gradient of the cross-entropy loss w.r.t. the input.

Defaults to individual calls to forward_one and gradient_one but can be overriden by subclasses to provide a more efficient implementation.

Parameters:
x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension).

label : int

Class label of the input as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
gradient(self, inputs, labels)[source]

Takes a batch of inputs and labels and returns the gradient of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).
Returns:
gradient : numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the inputs.

See also

gradient_one()
backward()
num_classes(self)[source]

Determines the number of classes.

Returns:
int

The number of classes for which the model creates predictions.
class foolbox.models.MXNetGluonModel(block, bounds, num_classes, ctx=None, channel_axis=1, preprocessing=(0, 1))[source]

Creates a Model instance from an existing MXNet Gluon Block.

Parameters:
block : mxnet.gluon.Block

The Gluon Block representing the model to be run.

ctx : mxnet.context.Context

The device, e.g. mxnet.cpu() or mxnet.gpu().

num_classes : int

The number of classes.

bounds : tuple

Tuple of lower and upper bound for the pixel values, usually (0, 1) or (0, 255).

channel_axis : int

The index of the axis that represents color channels.

preprocessing: dict or tuple

Can be a tuple with two elements representing mean and standard deviation or a dict with keys “mean” and “std”. The two elements should be floats or numpy arrays. “mean” is subtracted from the input, the result is then divided by “std”. If “mean” and “std” are 1-dimensional arrays, an additional (negative) “axis” key can be given such that “mean” and “std” will be broadcasted to that axis (typically -1 for “channels_last” and -3 for “channels_first”, but might be different when using e.g. 1D convolutions). Finally, a (negative) “flip_axis” can be specified. This axis will be flipped (before “mean” is subtracted), e.g. to convert RGB to BGR.
backward(self, gradient, inputs)[source]

Backpropagates the gradient of some loss w.r.t. the logits through the underlying model and returns the gradient of that loss w.r.t to the inputs.

Parameters:
gradient : numpy.ndarray

Gradient of some loss w.r.t. the logits with shape (batch size, number of classes).

inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

The gradient of the respective loss w.r.t the inputs.

See also

backward_one()
gradient()
forward(self, inputs)[source]

Takes a batch of inputs and returns the logits predicted by the underlying model.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

See also

forward_one()
forward_and_gradient(self, inputs, labels)[source]

Takes inputs and labels and returns both the logits predicted by the underlying model and the gradients of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Inputs with shape as expected by the model (with the batch dimension).

labels : numpy.ndarray

Array of the class label of the inputs as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
forward_and_gradient_one(self, x, label)[source]

Takes a single input and label and returns both the logits predicted by the underlying model and the gradient of the cross-entropy loss w.r.t. the input.

Defaults to individual calls to forward_one and gradient_one but can be overriden by subclasses to provide a more efficient implementation.

Parameters:
x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension).

label : int

Class label of the input as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
gradient(self, inputs, labels)[source]

Takes a batch of inputs and labels and returns the gradient of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).
Returns:
gradient : numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the inputs.

See also

gradient_one()
backward()
num_classes(self)[source]

Determines the number of classes.

Returns:
int

The number of classes for which the model creates predictions.
class foolbox.models.CaffeModel(net, bounds, channel_axis=1, preprocessing=(0, 1), data_blob_name='data', label_blob_name='label', output_blob_name='output')[source]
backward(self, gradient, inputs)[source]

Backpropagates the gradient of some loss w.r.t. the logits through the underlying model and returns the gradient of that loss w.r.t to the inputs.

Parameters:
gradient : numpy.ndarray

Gradient of some loss w.r.t. the logits with shape (batch size, number of classes).

inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

The gradient of the respective loss w.r.t the inputs.

See also

backward_one()
gradient()
forward(self, inputs)[source]

Takes a batch of inputs and returns the logits predicted by the underlying model.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

See also

forward_one()
forward_and_gradient(self, inputs, labels)[source]

Takes inputs and labels and returns both the logits predicted by the underlying model and the gradients of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Inputs with shape as expected by the model (with the batch dimension).

labels : numpy.ndarray

Array of the class label of the inputs as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
forward_and_gradient_one(self, x, label)[source]

Takes a single input and label and returns both the logits predicted by the underlying model and the gradient of the cross-entropy loss w.r.t. the input.

Defaults to individual calls to forward_one and gradient_one but can be overriden by subclasses to provide a more efficient implementation.

Parameters:
x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension).

label : int

Class label of the input as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
gradient(self, inputs, labels)[source]

Takes a batch of inputs and labels and returns the gradient of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).
Returns:
gradient : numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the inputs.

See also

gradient_one()
backward()
num_classes(self)[source]

Determines the number of classes.

Returns:
int

The number of classes for which the model creates predictions.
class foolbox.models.ModelWrapper(model)[source]

Base class for models that wrap other models.

This base class can be used to implement model wrappers that turn models into new models, for example by preprocessing the input or modifying the gradient.

Parameters:
model : Model

The model that is wrapped.
forward(self, inputs)[source]

Takes a batch of inputs and returns the logits predicted by the underlying model.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

See also

forward_one()
num_classes(self)[source]

Determines the number of classes.

Returns:
int

The number of classes for which the model creates predictions.
class foolbox.models.DifferentiableModelWrapper(model)[source]

Base class for models that wrap other models and provide gradient methods.

This base class can be used to implement model wrappers that turn models into new models, for example by preprocessing the input or modifying the gradient.

Parameters:
model : Model

The model that is wrapped.
backward(self, gradient, inputs)[source]

Backpropagates the gradient of some loss w.r.t. the logits through the underlying model and returns the gradient of that loss w.r.t to the inputs.

Parameters:
gradient : numpy.ndarray

Gradient of some loss w.r.t. the logits with shape (batch size, number of classes).

inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

The gradient of the respective loss w.r.t the inputs.

See also

backward_one()
gradient()
forward_and_gradient(self, x, label)[source]

Takes inputs and labels and returns both the logits predicted by the underlying model and the gradients of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Inputs with shape as expected by the model (with the batch dimension).

labels : numpy.ndarray

Array of the class label of the inputs as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
forward_and_gradient_one(self, x, label)[source]

Takes a single input and label and returns both the logits predicted by the underlying model and the gradient of the cross-entropy loss w.r.t. the input.

Defaults to individual calls to forward_one and gradient_one but can be overriden by subclasses to provide a more efficient implementation.

Parameters:
x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension).

label : int

Class label of the input as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
gradient(self, inputs, labels)[source]

Takes a batch of inputs and labels and returns the gradient of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).
Returns:
gradient : numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the inputs.

See also

gradient_one()
backward()
class foolbox.models.ModelWithoutGradients(model)[source]

Turns a model into a model without gradients.

class foolbox.models.ModelWithEstimatedGradients(model, gradient_estimator)[source]

Turns a model into a model with gradients estimated by the given gradient estimator.

Parameters:
model : Model

The model that is wrapped.

gradient_estimator : GradientEstimatorBase

GradientEstimator object that can estimate gradients for single and batched samples.
backward(self, gradient, inputs)[source]

Backpropagates the gradient of some loss w.r.t. the logits through the underlying model and returns the gradient of that loss w.r.t to the inputs.

Parameters:
gradient : numpy.ndarray

Gradient of some loss w.r.t. the logits with shape (batch size, number of classes).

inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

The gradient of the respective loss w.r.t the inputs.

See also

backward_one()
gradient()
forward_and_gradient(self, inputs, labels)[source]

Takes inputs and labels and returns both the logits predicted by the underlying model and the gradients of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Inputs with shape as expected by the model (with the batch dimension).

labels : numpy.ndarray

Array of the class label of the inputs as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
forward_and_gradient_one(self, x, label)[source]

Takes a single input and label and returns both the logits predicted by the underlying model and the gradient of the cross-entropy loss w.r.t. the input.

Defaults to individual calls to forward_one and gradient_one but can be overriden by subclasses to provide a more efficient implementation.

Parameters:
x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension).

label : int

Class label of the input as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
gradient(self, inputs, labels)[source]

Takes a batch of inputs and labels and returns the gradient of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).
Returns:
gradient : numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the inputs.

See also

gradient_one()
backward()
gradient_one(self, x, label)[source]

Takes a single input and label and returns the gradient of the cross-entropy loss w.r.t. the input.

Parameters:
x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension).

label : int

Class label of the input as an integer in [0, number of classes).
Returns:
numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

gradient()
class foolbox.models.CompositeModel(forward_model, backward_model)[source]

Combines predictions of a (black-box) model with the gradient of a (substitute) model.

Parameters:
forward_model : Model

The model that should be fooled and will be used for predictions.

backward_model : Model

The model that provides the gradients.
backward(self, gradient, inputs)[source]

Backpropagates the gradient of some loss w.r.t. the logits through the underlying model and returns the gradient of that loss w.r.t to the inputs.

Parameters:
gradient : numpy.ndarray

Gradient of some loss w.r.t. the logits with shape (batch size, number of classes).

inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

The gradient of the respective loss w.r.t the inputs.

See also

backward_one()
gradient()
forward(self, inputs)[source]

Takes a batch of inputs and returns the logits predicted by the underlying model.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

See also

forward_one()
forward_and_gradient(self, inputs, labels)[source]

Takes inputs and labels and returns both the logits predicted by the underlying model and the gradients of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Inputs with shape as expected by the model (with the batch dimension).

labels : numpy.ndarray

Array of the class label of the inputs as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
forward_and_gradient_one(self, x, label)[source]

Takes a single input and label and returns both the logits predicted by the underlying model and the gradient of the cross-entropy loss w.r.t. the input.

Defaults to individual calls to forward_one and gradient_one but can be overriden by subclasses to provide a more efficient implementation.

Parameters:
x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension).

label : int

Class label of the input as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
gradient(self, inputs, labels)[source]

Takes a batch of inputs and labels and returns the gradient of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).
Returns:
gradient : numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the inputs.

See also

gradient_one()
backward()
num_classes(self)[source]

Determines the number of classes.

Returns:
int

The number of classes for which the model creates predictions.
class foolbox.models.EnsembleAveragedModel(model, ensemble_size)[source]
Reduces stochastic effects in networks by averaging both forward and backward
calculations of the network by creating an ensemble of the same model and averaging over multiple runs (i.e. instances in the ensemble) as described in [R75f1c0e135b2-1].
Parameters:
model : Model

The model that is wrapped.

ensemble_size : int

Number of networks in the ensemble over which the predictions/gradients will be averaged.

References

[R75f1c0e135b2-1](1, 2) Roland S. Zimmermann, “Comment on ‘Adv-BNN: Improved Adversarial Defense through Robust Bayesian Neural Network’”, https://arxiv.org/abs/1907.00895
backward(self, gradient, inputs)[source]

Backpropagates the gradient of some loss w.r.t. the logits through the underlying model and returns the gradient of that loss w.r.t to the inputs.

Parameters:
gradient : numpy.ndarray

Gradient of some loss w.r.t. the logits with shape (batch size, number of classes).

inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

The gradient of the respective loss w.r.t the inputs.

See also

backward_one()
gradient()
forward(self, x)[source]

Takes a batch of inputs and returns the logits predicted by the underlying model.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

See also

forward_one()
forward_and_gradient(self, x, label)[source]

Takes inputs and labels and returns both the logits predicted by the underlying model and the gradients of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Inputs with shape as expected by the model (with the batch dimension).

labels : numpy.ndarray

Array of the class label of the inputs as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
forward_and_gradient_one(self, x, label)[source]

Takes a single input and label and returns both the logits predicted by the underlying model and the gradient of the cross-entropy loss w.r.t. the input.

Defaults to individual calls to forward_one and gradient_one but can be overriden by subclasses to provide a more efficient implementation.

Parameters:
x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension).

label : int

Class label of the input as an integer in [0, number of classes).
Returns:
numpy.ndarray

Predicted logits with shape (batch size, number of classes).

numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the input.

See also

forward_one()
gradient_one()
gradient(self, inputs, labels)[source]

Takes a batch of inputs and labels and returns the gradient of the cross-entropy loss w.r.t. the inputs.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).
Returns:
gradient : numpy.ndarray

The gradient of the cross-entropy loss w.r.t. the inputs.

See also

gradient_one()
backward()

foolbox.criteria

Provides classes that define what is adversarial.

Criteria

We provide criteria for untargeted and targeted adversarial attacks.

Misclassification Defines adversarials as inputs for which the predicted class is not the original class.
TopKMisclassification Defines adversarials as inputs for which the original class is not one of the top k predicted classes.
OriginalClassProbability Defines adversarials as inputs for which the probability of the original class is below a given threshold.
ConfidentMisclassification Defines adversarials as inputs for which the probability of any class other than the original is above a given threshold.
TargetClass Defines adversarials as inputs for which the predicted class is the given target class.
TargetClassProbability Defines adversarials as inputs for which the probability of a given target class is above a given threshold.

Examples

Untargeted criteria:

>>> from foolbox.criteria import Misclassification
>>> criterion1 = Misclassification()

>>> from foolbox.criteria import TopKMisclassification
>>> criterion2 = TopKMisclassification(k=5)

Targeted criteria:

>>> from foolbox.criteria import TargetClass
>>> criterion3 = TargetClass(22)

>>> from foolbox.criteria import TargetClassProbability
>>> criterion4 = TargetClassProbability(22, p=0.99)

Criteria can be combined to create a new criterion:

>>> criterion5 = criterion2 & criterion3

Detailed description

class foolbox.criteria.Criterion[source]

Base class for criteria that define what is adversarial.

The Criterion class represents a criterion used to determine if predictions for an image are adversarial given a reference label. It should be subclassed when implementing new criteria. Subclasses must implement is_adversarial.

is_adversarial(self, predictions, label)[source]

Decides if predictions for an image are adversarial given a reference label.

Parameters:
predictions : numpy.ndarray

A vector with the pre-softmax predictions for some image.

label : int

The label of the unperturbed reference image.
Returns:
bool

True if an image with the given predictions is an adversarial example when the ground-truth class is given by label, False otherwise.
name(self)[source]

Returns a human readable name that uniquely identifies the criterion with its hyperparameters.

Returns:
str

Human readable name that uniquely identifies the criterion with its hyperparameters.

Notes

Defaults to the class name but subclasses can provide more descriptive names and must take hyperparameters into account.

class foolbox.criteria.Misclassification[source]

Defines adversarials as inputs for which the predicted class is not the original class.

See also

TopKMisclassification

Notes

Uses numpy.argmax to break ties.

is_adversarial(self, predictions, label)[source]

Decides if predictions for an image are adversarial given a reference label.

Parameters:
predictions : numpy.ndarray

A vector with the pre-softmax predictions for some image.

label : int

The label of the unperturbed reference image.
Returns:
bool

True if an image with the given predictions is an adversarial example when the ground-truth class is given by label, False otherwise.
name(self)[source]

Returns a human readable name that uniquely identifies the criterion with its hyperparameters.

Returns:
str

Human readable name that uniquely identifies the criterion with its hyperparameters.

Notes

Defaults to the class name but subclasses can provide more descriptive names and must take hyperparameters into account.

class foolbox.criteria.ConfidentMisclassification(p)[source]

Defines adversarials as inputs for which the probability of any class other than the original is above a given threshold.

Parameters:
p : float

The threshold probability. If the probability of any class other than the original is at least p, the image is considered an adversarial. It must satisfy 0 <= p <= 1.
is_adversarial(self, predictions, label)[source]

Decides if predictions for an image are adversarial given a reference label.

Parameters:
predictions : numpy.ndarray

A vector with the pre-softmax predictions for some image.

label : int

The label of the unperturbed reference image.
Returns:
bool

True if an image with the given predictions is an adversarial example when the ground-truth class is given by label, False otherwise.
name(self)[source]

Returns a human readable name that uniquely identifies the criterion with its hyperparameters.

Returns:
str

Human readable name that uniquely identifies the criterion with its hyperparameters.

Notes

Defaults to the class name but subclasses can provide more descriptive names and must take hyperparameters into account.

class foolbox.criteria.TopKMisclassification(k)[source]

Defines adversarials as inputs for which the original class is not one of the top k predicted classes.

For k = 1, the Misclassification class provides a more efficient implementation.

Parameters:
k : int

Number of top predictions to which the reference label is compared to.

See also

Misclassification
Provides a more effcient implementation for k = 1.

Notes

Uses numpy.argsort to break ties.

is_adversarial(self, predictions, label)[source]

Decides if predictions for an image are adversarial given a reference label.

Parameters:
predictions : numpy.ndarray

A vector with the pre-softmax predictions for some image.

label : int

The label of the unperturbed reference image.
Returns:
bool

True if an image with the given predictions is an adversarial example when the ground-truth class is given by label, False otherwise.
name(self)[source]

Returns a human readable name that uniquely identifies the criterion with its hyperparameters.

Returns:
str

Human readable name that uniquely identifies the criterion with its hyperparameters.

Notes

Defaults to the class name but subclasses can provide more descriptive names and must take hyperparameters into account.

class foolbox.criteria.TargetClass(target_class)[source]

Defines adversarials as inputs for which the predicted class is the given target class.

Parameters:
target_class : int

The target class that needs to be predicted for an image to be considered an adversarial.

Notes

Uses numpy.argmax to break ties.

is_adversarial(self, predictions, label)[source]

Decides if predictions for an image are adversarial given a reference label.

Parameters:
predictions : numpy.ndarray

A vector with the pre-softmax predictions for some image.

label : int

The label of the unperturbed reference image.
Returns:
bool

True if an image with the given predictions is an adversarial example when the ground-truth class is given by label, False otherwise.
name(self)[source]

Returns a human readable name that uniquely identifies the criterion with its hyperparameters.

Returns:
str

Human readable name that uniquely identifies the criterion with its hyperparameters.

Notes

Defaults to the class name but subclasses can provide more descriptive names and must take hyperparameters into account.

class foolbox.criteria.OriginalClassProbability(p)[source]

Defines adversarials as inputs for which the probability of the original class is below a given threshold.

This criterion alone does not guarantee that the class predicted for the adversarial image is not the original class (unless p < 1 / number of classes). Therefore, it should usually be combined with a classifcation criterion.

Parameters:
p : float

The threshold probability. If the probability of the original class is below this threshold, the image is considered an adversarial. It must satisfy 0 <= p <= 1.
is_adversarial(self, predictions, label)[source]

Decides if predictions for an image are adversarial given a reference label.

Parameters:
predictions : numpy.ndarray

A vector with the pre-softmax predictions for some image.

label : int

The label of the unperturbed reference image.
Returns:
bool

True if an image with the given predictions is an adversarial example when the ground-truth class is given by label, False otherwise.
name(self)[source]

Returns a human readable name that uniquely identifies the criterion with its hyperparameters.

Returns:
str

Human readable name that uniquely identifies the criterion with its hyperparameters.

Notes

Defaults to the class name but subclasses can provide more descriptive names and must take hyperparameters into account.

class foolbox.criteria.TargetClassProbability(target_class, p)[source]

Defines adversarials as inputs for which the probability of a given target class is above a given threshold.

If the threshold is below 0.5, this criterion does not guarantee that the class predicted for the adversarial image is not the original class. In that case, it should usually be combined with a classification criterion.

Parameters:
target_class : int

The target class for which the predicted probability must be above the threshold probability p, otherwise the image is not considered an adversarial.

p : float

The threshold probability. If the probability of the target class is above this threshold, the image is considered an adversarial. It must satisfy 0 <= p <= 1.
is_adversarial(self, predictions, label)[source]

Decides if predictions for an image are adversarial given a reference label.

Parameters:
predictions : numpy.ndarray

A vector with the pre-softmax predictions for some image.

label : int

The label of the unperturbed reference image.
Returns:
bool

True if an image with the given predictions is an adversarial example when the ground-truth class is given by label, False otherwise.
name(self)[source]

Returns a human readable name that uniquely identifies the criterion with its hyperparameters.

Returns:
str

Human readable name that uniquely identifies the criterion with its hyperparameters.

Notes

Defaults to the class name but subclasses can provide more descriptive names and must take hyperparameters into account.

foolbox.zoo

Get Model

foolbox.zoo.get_model(url, module_name='foolbox_model', **kwargs)[source]

Provides utilities to download foolbox-compatible robust models to easily test attacks against them by simply providing a git-URL.

Examples

Instantiate a model:

>>> from foolbox import zoo
>>> url = "https://github.com/bveliqi/foolbox-zoo-dummy.git"
>>> model = zoo.get_model(url)  # doctest: +SKIP

Only works with a foolbox-zoo compatible repository. I.e. models need to have a foolbox_model.py file with a create()-function, which returns a foolbox-wrapped model.

Using the kwargs parameter it is possible to input an arbitrary number of parameters to this methods call. These parameters are forwarded to the instantiated model.

Example repositories:

Parameters:
  • url – URL to the git repository
  • module_name – the name of the module to import
  • kwargs – Optional set of parameters that will be used by the to be instantiated model.
Returns:

a foolbox-wrapped model instance

Fetch Weights

foolbox.zoo.fetch_weights(weights_uri, unzip=False)[source]

Provides utilities to download and extract packages containing model weights when creating foolbox-zoo compatible repositories, if the weights are not part of the repository itself.

Examples

Download and unzip weights:

>>> from foolbox import zoo
>>> url = 'https://github.com/MadryLab/mnist_challenge_models/raw/master/secret.zip'  # noqa F501
>>> weights_path = zoo.fetch_weights(url, unzip=True)
Parameters:
  • weights_uri – the URI to fetch the weights from
  • unzip – should be True if the file to be downloaded is a zipped package
Returns:

local path where the weights have been downloaded and potentially unzipped to

foolbox.distances

Provides classes to measure the distance between inputs.

Distances

MeanSquaredDistance Calculates the mean squared error between two inputs.
MeanAbsoluteDistance Calculates the mean absolute error between two inputs.
Linfinity Calculates the L-infinity norm of the difference between two inputs.
L0 Calculates the L0 norm of the difference between two inputs.
ElasticNet Calculates the Elastic-Net distance between two inputs.

Aliases

MSE alias of foolbox.distances.MeanSquaredDistance
MAE alias of foolbox.distances.MeanAbsoluteDistance
Linf alias of foolbox.distances.Linfinity
EN Creates a class definition that assigns ElasticNet a fixed l1_factor.

Base class

To implement a new distance, simply subclass the Distance class and implement the _calculate() method.

Distance Base class for distances.

Detailed description

class foolbox.distances.Distance(reference=None, other=None, bounds=None, value=None)[source]

Base class for distances.

This class should be subclassed when implementing new distances. Subclasses must implement _calculate.

class foolbox.distances.MeanSquaredDistance(reference=None, other=None, bounds=None, value=None)[source]

Calculates the mean squared error between two inputs.

class foolbox.distances.MeanAbsoluteDistance(reference=None, other=None, bounds=None, value=None)[source]

Calculates the mean absolute error between two inputs.

class foolbox.distances.Linfinity(reference=None, other=None, bounds=None, value=None)[source]

Calculates the L-infinity norm of the difference between two inputs.

class foolbox.distances.L0(reference=None, other=None, bounds=None, value=None)[source]

Calculates the L0 norm of the difference between two inputs.

foolbox.distances.MSE[source]

alias of foolbox.distances.MeanSquaredDistance

foolbox.distances.MAE[source]

alias of foolbox.distances.MeanAbsoluteDistance

foolbox.distances.Linf[source]

alias of foolbox.distances.Linfinity

foolbox.attacks

Gradient-based attacks

class foolbox.attacks.GradientAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Perturbs the input with the gradient of the loss w.r.t. the input, gradually increasing the magnitude until the input is misclassified.

Does not do anything if the model does not have a gradient.

as_generator(self, a, epsilons=1000, max_epsilon=1)[source]

Perturbs the input with the gradient of the loss w.r.t. the input, gradually increasing the magnitude until the input is misclassified.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

epsilons : int or Iterable[float]

Either Iterable of step sizes in the gradient direction or number of step sizes between 0 and max_epsilon that should be tried.

max_epsilon : float

Largest step size if epsilons is not an iterable.
class foolbox.attacks.GradientSignAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Adds the sign of the gradient to the input, gradually increasing the magnitude until the input is misclassified. This attack is often referred to as Fast Gradient Sign Method and was introduced in [R20d0064ee4c9-1].

Does not do anything if the model does not have a gradient.

References

[R20d0064ee4c9-1](1, 2) Ian J. Goodfellow, Jonathon Shlens, Christian Szegedy, “Explaining and Harnessing Adversarial Examples”, https://arxiv.org/abs/1412.6572
as_generator(self, a, epsilons=1000, max_epsilon=1)[source]

Adds the sign of the gradient to the input, gradually increasing the magnitude until the input is misclassified.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

epsilons : int or Iterable[float]

Either Iterable of step sizes in the direction of the sign of the gradient or number of step sizes between 0 and max_epsilon that should be tried.

max_epsilon : float

Largest step size if epsilons is not an iterable.
foolbox.attacks.FGSM[source]

alias of foolbox.attacks.gradient.GradientSignAttack

class foolbox.attacks.LinfinityBasicIterativeAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

The Basic Iterative Method introduced in [R37dbc8f24aee-1].

This attack is also known as Projected Gradient Descent (PGD) (without random start) or FGMS^k.

References

[R37dbc8f24aee-1](1, 2)

Alexey Kurakin, Ian Goodfellow, Samy Bengio, “Adversarial examples in the physical world”,

See also

ProjectedGradientDescentAttack

as_generator(self, a, binary_search=True, epsilon=0.3, stepsize=0.05, iterations=10, random_start=False, return_early=True)[source]

Simple iterative gradient-based attack known as Basic Iterative Method, Projected Gradient Descent or FGSM^k.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

binary_search : bool or int

Whether to perform a binary search over epsilon and stepsize, keeping their ratio constant and using their values to start the search. If False, hyperparameters are not optimized. Can also be an integer, specifying the number of binary search steps (default 20).

epsilon : float

Limit on the perturbation size; if binary_search is True, this value is only for initialization and automatically adapted.

stepsize : float

Step size for gradient descent; if binary_search is True, this value is only for initialization and automatically adapted.

iterations : int

Number of iterations for each gradient descent run.

random_start : bool

Start the attack from a random point rather than from the original input.

return_early : bool

Whether an individual gradient descent run should stop as soon as an adversarial is found.
foolbox.attacks.BasicIterativeMethod[source]

alias of foolbox.attacks.iterative_projected_gradient.LinfinityBasicIterativeAttack

foolbox.attacks.BIM[source]

alias of foolbox.attacks.iterative_projected_gradient.LinfinityBasicIterativeAttack

class foolbox.attacks.L1BasicIterativeAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Modified version of the Basic Iterative Method that minimizes the L1 distance.

See also

LinfinityBasicIterativeAttack

as_generator(self, a, binary_search=True, epsilon=0.3, stepsize=0.05, iterations=10, random_start=False, return_early=True)[source]

Simple iterative gradient-based attack known as Basic Iterative Method, Projected Gradient Descent or FGSM^k.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

binary_search : bool or int

Whether to perform a binary search over epsilon and stepsize, keeping their ratio constant and using their values to start the search. If False, hyperparameters are not optimized. Can also be an integer, specifying the number of binary search steps (default 20).

epsilon : float

Limit on the perturbation size; if binary_search is True, this value is only for initialization and automatically adapted.

stepsize : float

Step size for gradient descent; if binary_search is True, this value is only for initialization and automatically adapted.

iterations : int

Number of iterations for each gradient descent run.

random_start : bool

Start the attack from a random point rather than from the original input.

return_early : bool

Whether an individual gradient descent run should stop as soon as an adversarial is found.
class foolbox.attacks.L2BasicIterativeAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Modified version of the Basic Iterative Method that minimizes the L2 distance.

See also

LinfinityBasicIterativeAttack

as_generator(self, a, binary_search=True, epsilon=0.3, stepsize=0.05, iterations=10, random_start=False, return_early=True)[source]

Simple iterative gradient-based attack known as Basic Iterative Method, Projected Gradient Descent or FGSM^k.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

binary_search : bool or int

Whether to perform a binary search over epsilon and stepsize, keeping their ratio constant and using their values to start the search. If False, hyperparameters are not optimized. Can also be an integer, specifying the number of binary search steps (default 20).

epsilon : float

Limit on the perturbation size; if binary_search is True, this value is only for initialization and automatically adapted.

stepsize : float

Step size for gradient descent; if binary_search is True, this value is only for initialization and automatically adapted.

iterations : int

Number of iterations for each gradient descent run.

random_start : bool

Start the attack from a random point rather than from the original input.

return_early : bool

Whether an individual gradient descent run should stop as soon as an adversarial is found.
class foolbox.attacks.ProjectedGradientDescentAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

The Projected Gradient Descent Attack introduced in [R367e8e10528a-1] without random start.

When used without a random start, this attack is also known as Basic Iterative Method (BIM) or FGSM^k.

References

[R367e8e10528a-1](1, 2) Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, Adrian Vladu, “Towards Deep Learning Models Resistant to Adversarial Attacks”, https://arxiv.org/abs/1706.06083

See also

LinfinityBasicIterativeAttack and RandomStartProjectedGradientDescentAttack

as_generator(self, a, binary_search=True, epsilon=0.3, stepsize=0.01, iterations=40, random_start=False, return_early=True)[source]

Simple iterative gradient-based attack known as Basic Iterative Method, Projected Gradient Descent or FGSM^k.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

binary_search : bool or int

Whether to perform a binary search over epsilon and stepsize, keeping their ratio constant and using their values to start the search. If False, hyperparameters are not optimized. Can also be an integer, specifying the number of binary search steps (default 20).

epsilon : float

Limit on the perturbation size; if binary_search is True, this value is only for initialization and automatically adapted.

stepsize : float

Step size for gradient descent; if binary_search is True, this value is only for initialization and automatically adapted.

iterations : int

Number of iterations for each gradient descent run.

random_start : bool

Start the attack from a random point rather than from the original input.

return_early : bool

Whether an individual gradient descent run should stop as soon as an adversarial is found.
foolbox.attacks.ProjectedGradientDescent[source]

alias of foolbox.attacks.iterative_projected_gradient.ProjectedGradientDescentAttack

foolbox.attacks.PGD[source]

alias of foolbox.attacks.iterative_projected_gradient.ProjectedGradientDescentAttack

class foolbox.attacks.RandomStartProjectedGradientDescentAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

The Projected Gradient Descent Attack introduced in [Re6066bc39e14-1] with random start.

References

[Re6066bc39e14-1](1, 2) Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, Adrian Vladu, “Towards Deep Learning Models Resistant to Adversarial Attacks”, https://arxiv.org/abs/1706.06083

See also

ProjectedGradientDescentAttack

as_generator(self, a, binary_search=True, epsilon=0.3, stepsize=0.01, iterations=40, random_start=True, return_early=True)[source]

Simple iterative gradient-based attack known as Basic Iterative Method, Projected Gradient Descent or FGSM^k.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

binary_search : bool or int

Whether to perform a binary search over epsilon and stepsize, keeping their ratio constant and using their values to start the search. If False, hyperparameters are not optimized. Can also be an integer, specifying the number of binary search steps (default 20).

epsilon : float

Limit on the perturbation size; if binary_search is True, this value is only for initialization and automatically adapted.

stepsize : float

Step size for gradient descent; if binary_search is True, this value is only for initialization and automatically adapted.

iterations : int

Number of iterations for each gradient descent run.

random_start : bool

Start the attack from a random point rather than from the original input.

return_early : bool

Whether an individual gradient descent run should stop as soon as an adversarial is found.
foolbox.attacks.RandomProjectedGradientDescent[source]

alias of foolbox.attacks.iterative_projected_gradient.RandomStartProjectedGradientDescentAttack

foolbox.attacks.RandomPGD[source]

alias of foolbox.attacks.iterative_projected_gradient.RandomStartProjectedGradientDescentAttack

class foolbox.attacks.AdamL1BasicIterativeAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Modified version of the Basic Iterative Method that minimizes the L1 distance using the Adam optimizer.

See also

LinfinityBasicIterativeAttack

as_generator(self, a, binary_search=True, epsilon=0.3, stepsize=0.05, iterations=10, random_start=False, return_early=True)[source]

Simple iterative gradient-based attack known as Basic Iterative Method, Projected Gradient Descent or FGSM^k.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

binary_search : bool or int

Whether to perform a binary search over epsilon and stepsize, keeping their ratio constant and using their values to start the search. If False, hyperparameters are not optimized. Can also be an integer, specifying the number of binary search steps (default 20).

epsilon : float

Limit on the perturbation size; if binary_search is True, this value is only for initialization and automatically adapted.

stepsize : float

Step size for gradient descent; if binary_search is True, this value is only for initialization and automatically adapted.

iterations : int

Number of iterations for each gradient descent run.

random_start : bool

Start the attack from a random point rather than from the original input.

return_early : bool

Whether an individual gradient descent run should stop as soon as an adversarial is found.
class foolbox.attacks.AdamL2BasicIterativeAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Modified version of the Basic Iterative Method that minimizes the L2 distance using the Adam optimizer.

See also

LinfinityBasicIterativeAttack

as_generator(self, a, binary_search=True, epsilon=0.3, stepsize=0.05, iterations=10, random_start=False, return_early=True)[source]

Simple iterative gradient-based attack known as Basic Iterative Method, Projected Gradient Descent or FGSM^k.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

binary_search : bool or int

Whether to perform a binary search over epsilon and stepsize, keeping their ratio constant and using their values to start the search. If False, hyperparameters are not optimized. Can also be an integer, specifying the number of binary search steps (default 20).

epsilon : float

Limit on the perturbation size; if binary_search is True, this value is only for initialization and automatically adapted.

stepsize : float

Step size for gradient descent; if binary_search is True, this value is only for initialization and automatically adapted.

iterations : int

Number of iterations for each gradient descent run.

random_start : bool

Start the attack from a random point rather than from the original input.

return_early : bool

Whether an individual gradient descent run should stop as soon as an adversarial is found.
class foolbox.attacks.AdamProjectedGradientDescentAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

The Projected Gradient Descent Attack introduced in [Re2d4f39a0205-1], [Re2d4f39a0205-2] without random start using the Adam optimizer.

When used without a random start, this attack is also known as Basic Iterative Method (BIM) or FGSM^k.

References

[Re2d4f39a0205-1](1, 2) Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, Adrian Vladu, “Towards Deep Learning Models Resistant to Adversarial Attacks”, https://arxiv.org/abs/1706.06083
[Re2d4f39a0205-2](1, 2) Nicholas Carlini, David Wagner: “Towards Evaluating the Robustness of Neural Networks”, https://arxiv.org/abs/1608.04644

See also

LinfinityBasicIterativeAttack and RandomStartProjectedGradientDescentAttack

as_generator(self, a, binary_search=True, epsilon=0.3, stepsize=0.01, iterations=40, random_start=False, return_early=True)[source]

Simple iterative gradient-based attack known as Basic Iterative Method, Projected Gradient Descent or FGSM^k.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

binary_search : bool or int

Whether to perform a binary search over epsilon and stepsize, keeping their ratio constant and using their values to start the search. If False, hyperparameters are not optimized. Can also be an integer, specifying the number of binary search steps (default 20).

epsilon : float

Limit on the perturbation size; if binary_search is True, this value is only for initialization and automatically adapted.

stepsize : float

Step size for gradient descent; if binary_search is True, this value is only for initialization and automatically adapted.

iterations : int

Number of iterations for each gradient descent run.

random_start : bool

Start the attack from a random point rather than from the original input.

return_early : bool

Whether an individual gradient descent run should stop as soon as an adversarial is found.
foolbox.attacks.AdamProjectedGradientDescent[source]

alias of foolbox.attacks.iterative_projected_gradient.AdamProjectedGradientDescentAttack

foolbox.attacks.AdamPGD[source]

alias of foolbox.attacks.iterative_projected_gradient.AdamProjectedGradientDescentAttack

class foolbox.attacks.AdamRandomStartProjectedGradientDescentAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

The Projected Gradient Descent Attack introduced in [R3210aa339085-1], [R3210aa339085-2] with random start using the Adam optimizer.

References

[R3210aa339085-1]Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, Adrian Vladu, “Towards Deep Learning Models Resistant to Adversarial Attacks”, https://arxiv.org/abs/1706.06083
[R3210aa339085-2]Nicholas Carlini, David Wagner: “Towards Evaluating the Robustness of Neural Networks”, https://arxiv.org/abs/1608.04644

See also

ProjectedGradientDescentAttack

as_generator(self, a, binary_search=True, epsilon=0.3, stepsize=0.01, iterations=40, random_start=True, return_early=True)[source]

Simple iterative gradient-based attack known as Basic Iterative Method, Projected Gradient Descent or FGSM^k.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

binary_search : bool or int

Whether to perform a binary search over epsilon and stepsize, keeping their ratio constant and using their values to start the search. If False, hyperparameters are not optimized. Can also be an integer, specifying the number of binary search steps (default 20).

epsilon : float

Limit on the perturbation size; if binary_search is True, this value is only for initialization and automatically adapted.

stepsize : float

Step size for gradient descent; if binary_search is True, this value is only for initialization and automatically adapted.

iterations : int

Number of iterations for each gradient descent run.

random_start : bool

Start the attack from a random point rather than from the original input.

return_early : bool

Whether an individual gradient descent run should stop as soon as an adversarial is found.
foolbox.attacks.AdamRandomProjectedGradientDescent[source]

alias of foolbox.attacks.iterative_projected_gradient.AdamRandomStartProjectedGradientDescentAttack

foolbox.attacks.AdamRandomPGD[source]

alias of foolbox.attacks.iterative_projected_gradient.AdamRandomStartProjectedGradientDescentAttack

class foolbox.attacks.MomentumIterativeAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

The Momentum Iterative Method attack introduced in [R86d363e1fb2f-1]. It’s like the Basic Iterative Method or Projected Gradient Descent except that it uses momentum.

References

[R86d363e1fb2f-1]Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, Jianguo Li, “Boosting Adversarial Attacks with Momentum”, https://arxiv.org/abs/1710.06081
as_generator(self, a, binary_search=True, epsilon=0.3, stepsize=0.06, iterations=10, decay_factor=1.0, random_start=False, return_early=True)[source]

Momentum-based iterative gradient attack known as Momentum Iterative Method.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

binary_search : bool

Whether to perform a binary search over epsilon and stepsize, keeping their ratio constant and using their values to start the search. If False, hyperparameters are not optimized. Can also be an integer, specifying the number of binary search steps (default 20).

epsilon : float

Limit on the perturbation size; if binary_search is True, this value is only for initialization and automatically adapted.

stepsize : float

Step size for gradient descent; if binary_search is True, this value is only for initialization and automatically adapted.

iterations : int

Number of iterations for each gradient descent run.

decay_factor : float

Decay factor used by the momentum term.

random_start : bool

Start the attack from a random point rather than from the original input.

return_early : bool

Whether an individual gradient descent run should stop as soon as an adversarial is found.
foolbox.attacks.MomentumIterativeMethod[source]

alias of foolbox.attacks.iterative_projected_gradient.MomentumIterativeAttack

class foolbox.attacks.DeepFoolAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Simple and close to optimal gradient-based adversarial attack.

Implementes DeepFool introduced in [Rb4dd02640756-1].

References

[Rb4dd02640756-1]Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Pascal Frossard, “DeepFool: a simple and accurate method to fool deep neural networks”, https://arxiv.org/abs/1511.04599
as_generator(self, a, steps=100, subsample=10, p=None)[source]

Simple and close to optimal gradient-based adversarial attack.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

steps : int

Maximum number of steps to perform.

subsample : int

Limit on the number of the most likely classes that should be considered. A small value is usually sufficient and much faster.

p : int or float

Lp-norm that should be minimzed, must be 2 or np.inf.
class foolbox.attacks.NewtonFoolAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Implements the NewtonFool Attack.

The attack was introduced in [R6a972939b320-1].

References

[R6a972939b320-1]Uyeong Jang et al., “Objective Metrics and Gradient Descent Algorithms for Adversarial Examples in Machine Learning”, https://dl.acm.org/citation.cfm?id=3134635
as_generator(self, a, max_iter=100, eta=0.01)[source]
Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

max_iter : int

The maximum number of iterations.

eta : float

the eta coefficient
class foolbox.attacks.DeepFoolL2Attack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]
as_generator(self, a, steps=100, subsample=10)[source]

Simple and close to optimal gradient-based adversarial attack.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

steps : int

Maximum number of steps to perform.

subsample : int

Limit on the number of the most likely classes that should be considered. A small value is usually sufficient and much faster.

p : int or float

Lp-norm that should be minimzed, must be 2 or np.inf.
class foolbox.attacks.DeepFoolLinfinityAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]
as_generator(self, a, steps=100, subsample=10)[source]

Simple and close to optimal gradient-based adversarial attack.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

steps : int

Maximum number of steps to perform.

subsample : int

Limit on the number of the most likely classes that should be considered. A small value is usually sufficient and much faster.

p : int or float

Lp-norm that should be minimzed, must be 2 or np.inf.
class foolbox.attacks.ADefAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Adversarial attack that distorts the image, i.e. changes the locations of pixels. The algorithm is described in [Rf241e6d2664d-1], a Repository with the original code can be found in [Rf241e6d2664d-2]. References ———- .. [Rf241e6d2664d-1] Rima Alaifari, Giovanni S. Alberti, and Tandri Gauksson:

“ADef: an Iterative Algorithm to Construct Adversarial Deformations”, https://arxiv.org/abs/1804.07729
as_generator(self, a, max_iter=100, smooth=1.0, subsample=10)[source]
Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

max_iter : int > 0

Maximum number of iterations (default max_iter = 100).

smooth : float >= 0

Width of the Gaussian kernel used for smoothing. (default is smooth = 0 for no smoothing).

subsample : int >= 2

Limit on the number of the most likely classes that should be considered. A small value is usually sufficient and much faster. (default subsample = 10)
class foolbox.attacks.SaliencyMapAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Implements the Saliency Map Attack.

The attack was introduced in [R08e06ca693ba-1].

References

[R08e06ca693ba-1]Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z. Berkay Celik, Ananthram Swami, “The Limitations of Deep Learning in Adversarial Settings”, https://arxiv.org/abs/1511.07528
as_generator(self, a, max_iter=2000, num_random_targets=0, fast=True, theta=0.1, max_perturbations_per_pixel=7)[source]

Implements the Saliency Map Attack.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

max_iter : int

The maximum number of iterations to run.

num_random_targets : int

Number of random target classes if no target class is given by the criterion.

fast : bool

Whether to use the fast saliency map calculation.

theta : float

perturbation per pixel relative to [min, max] range.

max_perturbations_per_pixel : int

Maximum number of times a pixel can be modified.
class foolbox.attacks.IterativeGradientAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Like GradientAttack but with several steps for each epsilon.

as_generator(self, a, epsilons=100, max_epsilon=1, steps=10)[source]

Like GradientAttack but with several steps for each epsilon.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

epsilons : int or Iterable[float]

Either Iterable of step sizes in the gradient direction or number of step sizes between 0 and max_epsilon that should be tried.

max_epsilon : float

Largest step size if epsilons is not an iterable.

steps : int

Number of iterations to run.
class foolbox.attacks.IterativeGradientSignAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Like GradientSignAttack but with several steps for each epsilon.

as_generator(self, a, epsilons=100, max_epsilon=1, steps=10)[source]

Like GradientSignAttack but with several steps for each epsilon.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

epsilons : int or Iterable[float]

Either Iterable of step sizes in the direction of the sign of the gradient or number of step sizes between 0 and max_epsilon that should be tried.

max_epsilon : float

Largest step size if epsilons is not an iterable.

steps : int

Number of iterations to run.
class foolbox.attacks.CarliniWagnerL2Attack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

The L2 version of the Carlini & Wagner attack.

This attack is described in [Rc2cb572b91c5-1]. This implementation is based on the reference implementation by Carlini [Rc2cb572b91c5-2]. For bounds ≠ (0, 1), it differs from [Rc2cb572b91c5-2] because we normalize the squared L2 loss with the bounds.

References

[Rc2cb572b91c5-1]Nicholas Carlini, David Wagner: “Towards Evaluating the Robustness of Neural Networks”, https://arxiv.org/abs/1608.04644
[Rc2cb572b91c5-2](1, 2) https://github.com/carlini/nn_robust_attacks
as_generator(self, a, binary_search_steps=5, max_iterations=1000, confidence=0, learning_rate=0.005, initial_const=0.01, abort_early=True)[source]

The L2 version of the Carlini & Wagner attack.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

binary_search_steps : int

The number of steps for the binary search used to find the optimal tradeoff-constant between distance and confidence.

max_iterations : int

The maximum number of iterations. Larger values are more accurate; setting it too small will require a large learning rate and will produce poor results.

confidence : int or float

Confidence of adversarial examples: a higher value produces adversarials that are further away, but more strongly classified as adversarial.

learning_rate : float

The learning rate for the attack algorithm. Smaller values produce better results but take longer to converge.

initial_const : float

The initial tradeoff-constant to use to tune the relative importance of distance and confidence. If binary_search_steps is large, the initial constant is not important.

abort_early : bool

If True, Adam will be aborted if the loss hasn’t decreased for some time (a tenth of max_iterations).
static best_other_class(logits, exclude)[source]

Returns the index of the largest logit, ignoring the class that is passed as exclude.

classmethod loss_function(const, a, x, logits, reconstructed_original, confidence, min_, max_)[source]

Returns the loss and the gradient of the loss w.r.t. x, assuming that logits = model(x).

class foolbox.attacks.EADAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Gradient based attack which uses an elastic-net regularization [1]. This implementation is based on the attacks description [1] and its reference implementation [2].

References

[Rf0e4124daa63-1]Pin-Yu Chen (*), Yash Sharma (*), Huan Zhang, Jinfeng Yi, Cho-Jui Hsieh, “EAD: Elastic-Net Attacks to Deep Neural Networks via Adversarial Examples”, https://arxiv.org/abs/1709.04114
[Rf0e4124daa63-2]Pin-Yu Chen (*), Yash Sharma (*), Huan Zhang, Jinfeng Yi, Cho-Jui Hsieh, “Reference Implementation of ‘EAD: Elastic-Net Attacks to Deep Neural Networks via Adversarial Examples’”, https://github.com/ysharma1126/EAD_Attack/blob/master/en_attack.py
as_generator(self, a, binary_search_steps=5, max_iterations=1000, confidence=0, initial_learning_rate=0.01, regularization=0.01, initial_const=0.01, abort_early=True)[source]

The L2 version of the Carlini & Wagner attack.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

binary_search_steps : int

The number of steps for the binary search used to find the optimal tradeoff-constant between distance and confidence.

max_iterations : int

The maximum number of iterations. Larger values are more accurate; setting it too small will require a large learning rate and will produce poor results.

confidence : int or float

Confidence of adversarial examples: a higher value produces adversarials that are further away, but more strongly classified as adversarial.

initial_learning_rate : float

The initial learning rate for the attack algorithm. Smaller values produce better results but take longer to converge. During the attack a square-root decay in the learning rate is performed.

initial_const : float

The initial tradeoff-constant to use to tune the relative importance of distance and confidence. If binary_search_steps is large, the initial constant is not important.

regularization : float

The L1 regularization parameter (also called beta). A value of 0 corresponds to the attacks.CarliniWagnerL2Attack attack.

abort_early : bool

If True, Adam will be aborted if the loss hasn’t decreased for some time (a tenth of max_iterations).
static best_other_class(logits, exclude)[source]

Returns the index of the largest logit, ignoring the class that is passed as exclude.

classmethod loss_function(const, a, x, logits, reconstructed_original, confidence, min_, max_)[source]

Returns the loss and the gradient of the loss w.r.t. x, assuming that logits = model(x).

classmethod project_shrinkage_thresholding(z, x0, regularization, min_, max_)[source]

Performs the element-wise projected shrinkage-thresholding operation

class foolbox.attacks.DecoupledDirectionNormL2Attack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

The Decoupled Direction and Norm L2 adversarial attack from [R0e9d4da0ab48-1].

References

[R0e9d4da0ab48-1]Jérôme Rony, Luiz G. Hafemann, Luiz S. Oliveira, Ismail Ben Ayed,

Robert Sabourin, Eric Granger, “Decoupling Direction and Norm for Efficient Gradient-Based L2 Adversarial Attacks and Defenses”, https://arxiv.org/abs/1811.09600

as_generator(self, a, steps=100, gamma=0.05, initial_norm=1, quantize=True, levels=256)[source]

The Decoupled Direction and Norm L2 adversarial attack.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

steps : int

Number of steps for the optimization.

gamma : float, optional

Factor by which the norm will be modified. new_norm = norm * (1 + or - gamma).

init_norm : float, optional

Initial value for the norm.

quantize : bool, optional

If True, the returned adversarials will have quantized values to the specified number of levels.

levels : int, optional

Number of levels to use for quantization (e.g. 256 for 8 bit images).
class foolbox.attacks.SparseL1BasicIterativeAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Sparse version of the Basic Iterative Method that minimizes the L1 distance introduced in [R0591d14da1c3-1].

References

[R0591d14da1c3-1]Florian Tramèr, Dan Boneh, “Adversarial Training and Robustness for Multiple Perturbations”, https://arxiv.org/abs/1904.13000

See also

L1BasicIterativeAttack

as_generator(self, a, q=80.0, binary_search=True, epsilon=0.3, stepsize=0.05, iterations=10, random_start=False, return_early=True)[source]

Sparse version of a gradient-based attack that minimizes the L1 distance.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

q : float

Relative percentile to make gradients sparse (must be in [0, 100))

binary_search : bool or int

Whether to perform a binary search over epsilon and stepsize, keeping their ratio constant and using their values to start the search. If False, hyperparameters are not optimized. Can also be an integer, specifying the number of binary search steps (default 20).

epsilon : float

Limit on the perturbation size; if binary_search is True, this value is only for initialization and automatically adapted.

stepsize : float

Step size for gradient descent; if binary_search is True, this value is only for initialization and automatically adapted.

iterations : int

Number of iterations for each gradient descent run.

random_start : bool

Start the attack from a random point rather than from the original input.

return_early : bool

Whether an individual gradient descent run should stop as soon as an adversarial is found.
class foolbox.attacks.VirtualAdversarialAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Calculate an untargeted adversarial perturbation by performing a approximated second order optimization step on the KL divergence between the unperturbed predictions and the predictions for the adversarial perturbation. This attack was introduced in [Rc6516d158ac2-1].

References

[Rc6516d158ac2-1]Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, Ken Nakae, Shin Ishii, “Distributional Smoothing with Virtual Adversarial Training”, https://arxiv.org/abs/1507.00677
as_generator(self, a, xi=1e-05, iterations=1, epsilons=1000, max_epsilon=0.3)[source]
Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

xi : float

The finite difference size for performing the power method.

iterations : int

Number of iterations to perform power method to search for second order perturbation of KL divergence.

epsilons : int or Iterable[float]

Either Iterable of step sizes in the direction of the sign of the gradient or number of step sizes between 0 and max_epsilon that should be tried.

max_epsilon : float

Largest step size if epsilons is not an iterable.

Score-based attacks

class foolbox.attacks.SinglePixelAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Perturbs just a single pixel and sets it to the min or max.

as_generator(self, a, max_pixels=1000)[source]

Perturbs just a single pixel and sets it to the min or max.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, correctly classified input. If it is a numpy array, label must be passed as well. If it is an Adversarial instance, label must not be passed.

label : int

The reference label of the original input. Must be passed if input is a numpy array, must not be passed if input is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

max_pixels : int

Maximum number of pixels to try.
class foolbox.attacks.LocalSearchAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

A black-box attack based on the idea of greedy local search.

This implementation is based on the algorithm in [Rb320cee6998a-1].

References

[Rb320cee6998a-1](1, 2) Nina Narodytska, Shiva Prasad Kasiviswanathan, “Simple Black-Box Adversarial Perturbations for Deep Networks”, https://arxiv.org/abs/1612.06299
as_generator(self, a, r=1.5, p=10.0, d=5, t=5, R=150)[source]

A black-box attack based on the idea of greedy local search.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, correctly classified input. If it is a numpy array, label must be passed as well. If it is an Adversarial instance, label must not be passed.

label : int

The reference label of the original input. Must be passed if input is a numpy array, must not be passed if input is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

r : float

Perturbation parameter that controls the cyclic perturbation; must be in [0, 2]

p : float

Perturbation parameter that controls the pixel sensitivity estimation

d : int

The half side length of the neighborhood square

t : int

The number of pixels perturbed at each round

R : int

An upper bound on the number of iterations

Decision-based attacks

class foolbox.attacks.BoundaryAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

A powerful adversarial attack that requires neither gradients nor probabilities.

This is the reference implementation for the attack introduced in [Re72ca268aa55-1].

Notes

This implementation provides several advanced features:

  • ability to continue previous attacks by passing an instance of the Adversarial class
  • ability to pass an explicit starting point; especially to initialize a targeted attack
  • ability to pass an alternative attack used for initialization
  • fine-grained control over logging
  • ability to specify the batch size
  • optional automatic batch size tuning
  • optional multithreading for random number generation
  • optional multithreading for candidate point generation

References

[Re72ca268aa55-1](1, 2) Wieland Brendel (*), Jonas Rauber (*), Matthias Bethge, “Decision-Based Adversarial Attacks: Reliable Attacks Against Black-Box Machine Learning Models”, https://arxiv.org/abs/1712.04248
as_generator(self, a, iterations=5000, max_directions=25, starting_point=None, initialization_attack=None, log_every_n_steps=None, spherical_step=0.01, source_step=0.01, step_adaptation=1.5, batch_size=1, tune_batch_size=True, threaded_rnd=True, threaded_gen=True, alternative_generator=False, internal_dtype=<Mock name='mock.float64' id='139850469184792'>, loggingLevel=30)[source]

Applies the Boundary Attack.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, correctly classified input. If it is a numpy array, label must be passed as well. If it is an Adversarial instance, label must not be passed.

label : int

The reference label of the original input. Must be passed if input is a numpy array, must not be passed if input is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

iterations : int

Maximum number of iterations to run. Might converge and stop before that.

max_directions : int

Maximum number of trials per ieration.

starting_point : numpy.ndarray

Adversarial input to use as a starting point, in particular for targeted attacks.

initialization_attack : Attack

Attack to use to find a starting point. Defaults to BlendedUniformNoiseAttack.

log_every_n_steps : int

Determines verbositity of the logging.

spherical_step : float

Initial step size for the orthogonal (spherical) step.

source_step : float

Initial step size for the step towards the target.

step_adaptation : float

Factor by which the step sizes are multiplied or divided.

batch_size : int

Batch size or initial batch size if tune_batch_size is True

tune_batch_size : bool

Whether or not the batch size should be automatically chosen between 1 and max_directions.

threaded_rnd : bool

Whether the random number generation should be multithreaded.

threaded_gen : bool

Whether the candidate point generation should be multithreaded.

alternative_generator: bool

Whether an alternative implemenation of the candidate generator should be used.

internal_dtype : np.float32 or np.float64

Higher precision might be slower but is numerically more stable.

loggingLevel : int

Controls the verbosity of the logging, e.g. logging.INFO or logging.WARNING.
class foolbox.attacks.SpatialAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Adversarially chosen rotations and translations [1].

This implementation is based on the reference implementation by Madry et al.: https://github.com/MadryLab/adversarial_spatial

References

[Rdffd25498f9d-1]Logan Engstrom*, Brandon Tran*, Dimitris Tsipras*, Ludwig Schmidt, Aleksander Mądry: “A Rotation and a Translation Suffice: Fooling CNNs with Simple Transformations”, http://arxiv.org/abs/1712.02779
as_generator(self, a, do_rotations=True, do_translations=True, x_shift_limits=(-5, 5), y_shift_limits=(-5, 5), angular_limits=(-5, 5), granularity=10, random_sampling=False, abort_early=True)[source]

Adversarially chosen rotations and translations.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

do_rotations : bool

If False no rotations will be applied to the image.

do_translations : bool

If False no translations will be applied to the image.

x_shift_limits : int or (int, int)

Limits for horizontal translations in pixels. If one integer is provided the limits will be (-x_shift_limits, x_shift_limits).

y_shift_limits : int or (int, int)

Limits for vertical translations in pixels. If one integer is provided the limits will be (-y_shift_limits, y_shift_limits).

angular_limits : int or (int, int)

Limits for rotations in degrees. If one integer is provided the limits will be [-angular_limits, angular_limits].

granularity : int

Density of sampling within limits for each dimension.

random_sampling : bool

If True we sample translations/rotations randomly within limits, otherwise we use a regular grid.

abort_early : bool

If True, the attack stops as soon as it finds an adversarial.
class foolbox.attacks.PointwiseAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Starts with an adversarial and performs a binary search between the adversarial and the original for each dimension of the input individually.

References

[R739f80a24875-1]L. Schott, J. Rauber, M. Bethge, W. Brendel: “Towards the first adversarially robust neural network model on MNIST”, ICLR (2019) https://arxiv.org/abs/1805.09190
as_generator(self, a, starting_point=None, initialization_attack=None)[source]

Starts with an adversarial and performs a binary search between the adversarial and the original for each dimension of the input individually.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

starting_point : numpy.ndarray

Adversarial input to use as a starting point, in particular for targeted attacks.

initialization_attack : Attack

Attack to use to find a starting point. Defaults to SaltAndPepperNoiseAttack.
class foolbox.attacks.GaussianBlurAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Blurs the input until it is misclassified.

as_generator(self, a, epsilons=1000)[source]

Blurs the input until it is misclassified.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if input is a numpy.ndarray, must not be passed if input is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

epsilons : int or Iterable[float]

Either Iterable of standard deviations of the Gaussian blur or number of standard deviations between 0 and 1 that should be tried.
class foolbox.attacks.ContrastReductionAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Reduces the contrast of the input until it is misclassified.

as_generator(self, a, epsilons=1000)[source]

Reduces the contrast of the input until it is misclassified.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

epsilons : int or Iterable[float]

Either Iterable of contrast levels or number of contrast levels between 1 and 0 that should be tried. Epsilons are one minus the contrast level.
class foolbox.attacks.AdditiveUniformNoiseAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Adds uniform noise to the input, gradually increasing the standard deviation until the input is misclassified.

__call__(self, inputs, labels, unpack=True, individual_kwargs=None, **kwargs)[source]

Call self as a function.

__class__[source]

alias of abc.ABCMeta

__delattr__(self, name, /)[source]

Implement delattr(self, name).

__dir__()[source]

default dir() implementation

__eq__(self, value, /)[source]

Return self==value.

__format__()[source]

default object formatter

__ge__(self, value, /)[source]

Return self>=value.

__getattribute__(self, name, /)[source]

Return getattr(self, name).

__gt__(self, value, /)[source]

Return self>value.

__hash__(self, /)[source]

Return hash(self).

__init__(self, model=None, criterion=<foolbox.criteria.Misclassification object at 0x7f3174564748>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Initialize self. See help(type(self)) for accurate signature.

__le__(self, value, /)[source]

Return self<=value.

__lt__(self, value, /)[source]

Return self<value.

__ne__(self, value, /)[source]

Return self!=value.

__new__(*args, **kwargs)[source]

Create and return a new object. See help(type) for accurate signature.

__reduce__()[source]

helper for pickle

__reduce_ex__()[source]

helper for pickle

__repr__(self, /)[source]

Return repr(self).

__setattr__(self, name, value, /)[source]

Implement setattr(self, name, value).

__sizeof__()[source]

size of object in memory, in bytes

__str__(self, /)[source]

Return str(self).

__subclasshook__()[source]

Abstract classes can override this to customize issubclass().

This is invoked early on by abc.ABCMeta.__subclasscheck__(). It should return True, False or NotImplemented. If it returns NotImplemented, the normal algorithm is used. Otherwise, it overrides the normal algorithm (and the outcome is cached).

__weakref__[source]

list of weak references to the object (if defined)

as_generator(self, a, epsilons=1000)[source]

Adds uniform or Gaussian noise to the input, gradually increasing the standard deviation until the input is misclassified.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

epsilons : int or Iterable[float]

Either Iterable of noise levels or number of noise levels between 0 and 1 that should be tried.
name(self)[source]

Returns a human readable name that uniquely identifies the attack with its hyperparameters.

Returns:
str

Human readable name that uniquely identifies the attack with its hyperparameters.

Notes

Defaults to the class name but subclasses can provide more descriptive names and must take hyperparameters into account.

class foolbox.attacks.AdditiveGaussianNoiseAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Adds Gaussian noise to the input, gradually increasing the standard deviation until the input is misclassified.

__call__(self, inputs, labels, unpack=True, individual_kwargs=None, **kwargs)[source]

Call self as a function.

__class__[source]

alias of abc.ABCMeta

__delattr__(self, name, /)[source]

Implement delattr(self, name).

__dir__()[source]

default dir() implementation

__eq__(self, value, /)[source]

Return self==value.

__format__()[source]

default object formatter

__ge__(self, value, /)[source]

Return self>=value.

__getattribute__(self, name, /)[source]

Return getattr(self, name).

__gt__(self, value, /)[source]

Return self>value.

__hash__(self, /)[source]

Return hash(self).

__init__(self, model=None, criterion=<foolbox.criteria.Misclassification object at 0x7f3174564748>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Initialize self. See help(type(self)) for accurate signature.

__le__(self, value, /)[source]

Return self<=value.

__lt__(self, value, /)[source]

Return self<value.

__ne__(self, value, /)[source]

Return self!=value.

__new__(*args, **kwargs)[source]

Create and return a new object. See help(type) for accurate signature.

__reduce__()[source]

helper for pickle

__reduce_ex__()[source]

helper for pickle

__repr__(self, /)[source]

Return repr(self).

__setattr__(self, name, value, /)[source]

Implement setattr(self, name, value).

__sizeof__()[source]

size of object in memory, in bytes

__str__(self, /)[source]

Return str(self).

__subclasshook__()[source]

Abstract classes can override this to customize issubclass().

This is invoked early on by abc.ABCMeta.__subclasscheck__(). It should return True, False or NotImplemented. If it returns NotImplemented, the normal algorithm is used. Otherwise, it overrides the normal algorithm (and the outcome is cached).

__weakref__[source]

list of weak references to the object (if defined)

as_generator(self, a, epsilons=1000)[source]

Adds uniform or Gaussian noise to the input, gradually increasing the standard deviation until the input is misclassified.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

epsilons : int or Iterable[float]

Either Iterable of noise levels or number of noise levels between 0 and 1 that should be tried.
name(self)[source]

Returns a human readable name that uniquely identifies the attack with its hyperparameters.

Returns:
str

Human readable name that uniquely identifies the attack with its hyperparameters.

Notes

Defaults to the class name but subclasses can provide more descriptive names and must take hyperparameters into account.

class foolbox.attacks.SaltAndPepperNoiseAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Increases the amount of salt and pepper noise until the input is misclassified.

as_generator(self, a, epsilons=100, repetitions=10)[source]

Increases the amount of salt and pepper noise until the input is misclassified.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

epsilons : int

Number of steps to try between probability 0 and 1.

repetitions : int

Specifies how often the attack will be repeated.
class foolbox.attacks.BlendedUniformNoiseAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Blends the input with a uniform noise input until it is misclassified.

as_generator(self, a, epsilons=1000, max_directions=1000)[source]

Blends the input with a uniform noise input until it is misclassified.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

epsilons : int or Iterable[float]

Either Iterable of blending steps or number of blending steps between 0 and 1 that should be tried.

max_directions : int

Maximum number of random inputs to try.
class foolbox.attacks.HopSkipJumpAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

A powerful adversarial attack that requires neither gradients nor probabilities.

Notes

Features: * ability to switch between two types of distances: MSE and Linf. * ability to continue previous attacks by passing an instance of the

Adversarial class
  • ability to pass an explicit starting point; especially to initialize a targeted attack
  • ability to pass an alternative attack used for initialization
  • ability to specify the batch size

References

HopSkipJumpAttack was originally proposed by Chen, Jordan and Wainwright. It is a decision-based attack that requires access to output labels of a model alone. Paper link: https://arxiv.org/abs/1904.02144 The implementation in Foolbox is based on Boundary Attack.

approximate_gradient(self, decision_function, sample, num_evals, delta)[source]

Gradient direction estimation

as_generator(self, a, iterations=64, initial_num_evals=100, max_num_evals=10000, stepsize_search='geometric_progression', gamma=1.0, starting_point=None, batch_size=256, internal_dtype=<Mock name='mock.float64' id='139850469184792'>, log_every_n_steps=None, loggingLevel=30)[source]

Applies HopSkipJumpAttack.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, correctly classified input. If it is a numpy array, label must be passed as well. If it is an Adversarial instance, label must not be passed.

label : int

The reference label of the original input. Must be passed if input is a numpy array, must not be passed if input is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

iterations : int

Number of iterations to run.

initial_num_evals: int

Initial number of evaluations for gradient estimation. Larger initial_num_evals increases time efficiency, but may decrease query efficiency.

max_num_evals: int

Maximum number of evaluations for gradient estimation.

stepsize_search: str

How to search for stepsize; choices are ‘geometric_progression’, ‘grid_search’. ‘geometric progression’ initializes the stepsize by ||x_t - x||_p / sqrt(iteration), and keep decreasing by half until reaching the target side of the boundary. ‘grid_search’ chooses the optimal epsilon over a grid, in the scale of ||x_t - x||_p.

gamma: float
The binary search threshold theta is gamma / d^1.5 for

l2 attack and gamma / d^2 for linf attack.

starting_point : numpy.ndarray

Adversarial input to use as a starting point, required for targeted attacks.

batch_size : int

Batch size for model prediction.

internal_dtype : np.float32 or np.float64

Higher precision might be slower but is numerically more stable.

log_every_n_steps : int

Determines verbositity of the logging.

loggingLevel : int

Controls the verbosity of the logging, e.g. logging.INFO or logging.WARNING.
attack(self, a, iterations)[source]
iterations : int
Maximum number of iterations to run.
binary_search_batch(self, unperturbed, perturbed_inputs, decision_function)[source]

Binary search to approach the boundary.

geometric_progression_for_stepsize(self, x, update, dist, decision_function, current_iteration)[source]

Geometric progression to search for stepsize. Keep decreasing stepsize by half until reaching the desired side of the boundary.

project(self, unperturbed, perturbed_inputs, alphas)[source]

Projection onto given l2 / linf balls in a batch.

select_delta(self, dist_post_update, current_iteration)[source]

Choose the delta at the scale of distance between x and perturbed sample.

class foolbox.attacks.GenAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

The GenAttack introduced in [R996613153a1e-1].

This attack is performs a genetic search in order to find an adversarial perturbation in a black-box scenario in as few queries as possible.

References

[R996613153a1e-1](1, 2)

Moustafa Alzantot, Yash Sharma, Supriyo Chakraborty, Huan Zhang, Cho-Jui Hsieh, Mani Srivastava, “GenAttack: Practical Black-box Attacks with Gradient-Free Optimization”,
as_generator(self, a, generations=10, alpha=1.0, p=0.05, N=10, tau=0.1, search_shape=None, epsilon=0.3, binary_search=20)[source]

A black-box attack based on genetic algorithms. Can either try to find an adversarial perturbation for a fixed epsilon distance or perform a binary search over epsilon values in order to find a minimal perturbation. Parameters ———- inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
labels : numpy.ndarray
Class labels of the inputs as a vector of integers in [0, number of classes).
unpack : bool
If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.
generations : int
Number of generations, i.e. iterations, in the genetic algorithm.
alpha : float
Mutation-range.
p : float
Mutation probability.
N : int
Population size of the genetic algorithm.
tau: float
Temperature for the softmax sampling used to determine the parents of the new crossover.
search_shape : tuple (default: None)
Set this to a smaller image shape than the true shape to search in a smaller input space. The input will be scaled using a linear interpolation to match the required input shape of the model.
binary_search : bool or int
Whether to perform a binary search over epsilon and using their values to start the search. If False, hyperparameters are not optimized. Can also be an integer, specifying the number of binary search steps (default 20).
epsilon : float
Limit on the perturbation size; if binary_search is True, this value is only for initialization and automatically adapted.

Other attacks

class foolbox.attacks.BinarizationRefinementAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

For models that preprocess their inputs by binarizing the inputs, this attack can improve adversarials found by other attacks. It does os by utilizing information about the binarization and mapping values to the corresponding value in the clean input or to the right side of the threshold.

as_generator(self, a, starting_point=None, threshold=None, included_in='upper')[source]

For models that preprocess their inputs by binarizing the inputs, this attack can improve adversarials found by other attacks. It does this by utilizing information about the binarization and mapping values to the corresponding value in the clean input or to the right side of the threshold.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

starting_point : numpy.ndarray

Adversarial input to use as a starting point.

threshold : float

The treshold used by the models binarization. If none, defaults to (model.bounds()[1] - model.bounds()[0]) / 2.

included_in : str

Whether the threshold value itself belongs to the lower or upper interval.
class foolbox.attacks.PrecomputedAdversarialsAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Attacks a model using precomputed adversarial candidates.

as_generator(self, a, candidate_inputs, candidate_outputs)[source]

Attacks a model using precomputed adversarial candidates.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

candidate_inputs : numpy.ndarray

The original inputs that will be expected by this attack.

candidate_outputs : numpy.ndarray

The adversarial candidates corresponding to the inputs.
class foolbox.attacks.InversionAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Creates “negative images” by inverting the pixel values according to [R57cf8375f1ff-1].

References

[R57cf8375f1ff-1](1, 2)

Hossein Hosseini, Baicen Xiao, Mayoore Jaiswal, Radha Poovendran, “On the Limitation of Convolutional Neural Networks in Recognizing Negative Images”,
as_generator(self, a)[source]

Creates “negative images” by inverting the pixel values.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

Gradient-based attacks

GradientAttack Perturbs the input with the gradient of the loss w.r.t.
GradientSignAttack Adds the sign of the gradient to the input, gradually increasing the magnitude until the input is misclassified.
FGSM alias of foolbox.attacks.gradient.GradientSignAttack
LinfinityBasicIterativeAttack The Basic Iterative Method introduced in [R37dbc8f24aee-1].
BasicIterativeMethod alias of foolbox.attacks.iterative_projected_gradient.LinfinityBasicIterativeAttack
BIM alias of foolbox.attacks.iterative_projected_gradient.LinfinityBasicIterativeAttack
L1BasicIterativeAttack Modified version of the Basic Iterative Method that minimizes the L1 distance.
L2BasicIterativeAttack Modified version of the Basic Iterative Method that minimizes the L2 distance.
ProjectedGradientDescentAttack The Projected Gradient Descent Attack introduced in [R367e8e10528a-1] without random start.
ProjectedGradientDescent alias of foolbox.attacks.iterative_projected_gradient.ProjectedGradientDescentAttack
PGD alias of foolbox.attacks.iterative_projected_gradient.ProjectedGradientDescentAttack
RandomStartProjectedGradientDescentAttack The Projected Gradient Descent Attack introduced in [Re6066bc39e14-1] with random start.
RandomProjectedGradientDescent alias of foolbox.attacks.iterative_projected_gradient.RandomStartProjectedGradientDescentAttack
RandomPGD alias of foolbox.attacks.iterative_projected_gradient.RandomStartProjectedGradientDescentAttack
AdamL1BasicIterativeAttack Modified version of the Basic Iterative Method that minimizes the L1 distance using the Adam optimizer.
AdamL2BasicIterativeAttack Modified version of the Basic Iterative Method that minimizes the L2 distance using the Adam optimizer.
AdamProjectedGradientDescentAttack The Projected Gradient Descent Attack introduced in [Re2d4f39a0205-1], [Re2d4f39a0205-2] without random start using the Adam optimizer.
AdamProjectedGradientDescent alias of foolbox.attacks.iterative_projected_gradient.AdamProjectedGradientDescentAttack
AdamPGD alias of foolbox.attacks.iterative_projected_gradient.AdamProjectedGradientDescentAttack
AdamRandomStartProjectedGradientDescentAttack The Projected Gradient Descent Attack introduced in [R3210aa339085-1], [R3210aa339085-2] with random start using the Adam optimizer.
AdamRandomProjectedGradientDescent alias of foolbox.attacks.iterative_projected_gradient.AdamRandomStartProjectedGradientDescentAttack
AdamRandomPGD alias of foolbox.attacks.iterative_projected_gradient.AdamRandomStartProjectedGradientDescentAttack
MomentumIterativeAttack The Momentum Iterative Method attack introduced in [R86d363e1fb2f-1].
MomentumIterativeMethod alias of foolbox.attacks.iterative_projected_gradient.MomentumIterativeAttack
LBFGSAttack
DeepFoolAttack Simple and close to optimal gradient-based adversarial attack.
NewtonFoolAttack Implements the NewtonFool Attack.
DeepFoolL2Attack
DeepFoolLinfinityAttack
ADefAttack Adversarial attack that distorts the image, i.e.
SLSQPAttack
SaliencyMapAttack Implements the Saliency Map Attack.
IterativeGradientAttack Like GradientAttack but with several steps for each epsilon.
IterativeGradientSignAttack Like GradientSignAttack but with several steps for each epsilon.
CarliniWagnerL2Attack The L2 version of the Carlini & Wagner attack.
EADAttack Gradient based attack which uses an elastic-net regularization [1].
DecoupledDirectionNormL2Attack The Decoupled Direction and Norm L2 adversarial attack from [R0e9d4da0ab48-1].
SparseFoolAttack
SparseL1BasicIterativeAttack Sparse version of the Basic Iterative Method that minimizes the L1 distance introduced in [R0591d14da1c3-1].
VirtualAdversarialAttack Calculate an untargeted adversarial perturbation by performing a approximated second order optimization step on the KL divergence between the unperturbed predictions and the predictions for the adversarial perturbation.

Score-based attacks

SinglePixelAttack Perturbs just a single pixel and sets it to the min or max.
LocalSearchAttack A black-box attack based on the idea of greedy local search.
ApproximateLBFGSAttack

Decision-based attacks

BoundaryAttack A powerful adversarial attack that requires neither gradients nor probabilities.
SpatialAttack Adversarially chosen rotations and translations [1].
PointwiseAttack Starts with an adversarial and performs a binary search between the adversarial and the original for each dimension of the input individually.
GaussianBlurAttack Blurs the input until it is misclassified.
ContrastReductionAttack Reduces the contrast of the input until it is misclassified.
AdditiveUniformNoiseAttack Adds uniform noise to the input, gradually increasing the standard deviation until the input is misclassified.
AdditiveGaussianNoiseAttack Adds Gaussian noise to the input, gradually increasing the standard deviation until the input is misclassified.
SaltAndPepperNoiseAttack Increases the amount of salt and pepper noise until the input is misclassified.
BlendedUniformNoiseAttack Blends the input with a uniform noise input until it is misclassified.
BoundaryAttackPlusPlus
GenAttack The GenAttack introduced in [R996613153a1e-1].
HopSkipJumpAttack A powerful adversarial attack that requires neither gradients nor probabilities.

Other attacks

BinarizationRefinementAttack For models that preprocess their inputs by binarizing the inputs, this attack can improve adversarials found by other attacks.
PrecomputedAdversarialsAttack Attacks a model using precomputed adversarial candidates.
InversionAttack Creates “negative images” by inverting the pixel values according to [R57cf8375f1ff-1].

foolbox.adversarial

Provides a class that represents an adversarial example.

class foolbox.adversarial.Adversarial(model, criterion, unperturbed, original_class, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None, verbose=False)[source]
adversarial_class[source]

The argmax of the model predictions for the best adversarial found so far.

None if no adversarial has been found.

backward_one(self, gradient, x=None, strict=True)[source]

Interface to model.backward_one for attacks.

Parameters:
gradient : numpy.ndarray

Gradient of some loss w.r.t. the logits.

x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension).
Returns:
gradient : numpy.ndarray

The gradient w.r.t the input.

See also

gradient()
channel_axis(self, batch)[source]

Interface to model.channel_axis for attacks.

Parameters:
batch : bool

Controls whether the index of the axis for a batch of inputs (4 dimensions) or a single input (3 dimensions) should be returned.
distance[source]

The distance of the adversarial input to the original input.

forward(self, inputs, greedy=False, strict=True, return_details=False)[source]

Interface to model.forward for attacks.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the model.

greedy : bool

Whether the first adversarial should be returned.

strict : bool

Controls if the bounds for the pixel values should be checked.
forward_and_gradient(self, x, label=None, strict=True, return_details=False)[source]

Interface to model.forward_and_gradient_one for attacks.

Parameters:
x : numpy.ndarray

Multiple input with shape as expected by the model (with the batch dimension).

label : numpy.ndarray

Labels used to calculate the loss that is differentiated. Defaults to the original label.

strict : bool

Controls if the bounds for the pixel values should be checked.
forward_and_gradient_one(self, x=None, label=None, strict=True, return_details=False)[source]

Interface to model.forward_and_gradient_one for attacks.

Parameters:
x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension). Defaults to the original input.

label : int

Label used to calculate the loss that is differentiated. Defaults to the original label.

strict : bool

Controls if the bounds for the pixel values should be checked.
forward_one(self, x, strict=True, return_details=False)[source]

Interface to model.forward_one for attacks.

Parameters:
x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension).

strict : bool

Controls if the bounds for the pixel values should be checked.
gradient_one(self, x=None, label=None, strict=True)[source]

Interface to model.gradient_one for attacks.

Parameters:
x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension). Defaults to the original input.

label : int

Label used to calculate the loss that is differentiated. Defaults to the original label.

strict : bool

Controls if the bounds for the pixel values should be checked.
has_gradient(self)[source]

Returns true if _backward and _forward_backward can be called by an attack, False otherwise.

normalized_distance(self, x)[source]

Calculates the distance of a given input x to the original input.

Parameters:
x : numpy.ndarray

The input x that should be compared to the original input.
Returns:
Distance

The distance between the given input and the original input.
original_class[source]

The class of the original input (ground-truth, not model prediction).

output[source]

The model predictions for the best adversarial found so far.

None if no adversarial has been found.

perturbed[source]

The best adversarial example found so far.

reached_threshold(self)[source]

Returns True if a threshold is given and the currently best adversarial distance is smaller than the threshold.

target_class[source]

Interface to criterion.target_class for attacks.

unperturbed[source]

The original input.

foolbox.utils

foolbox.utils.softmax(logits)[source]

Transforms predictions into probability values.

Parameters:
logits : array_like

The logits predicted by the model.
Returns:
numpy.ndarray

Probability values corresponding to the logits.
foolbox.utils.crossentropy(label, logits)[source]

Calculates the cross-entropy.

Parameters:
logits : array_like

The logits predicted by the model.

label : int

The label describing the target distribution.
Returns:
float

The cross-entropy between softmax(logits) and onehot(label).
foolbox.utils.batch_crossentropy(label, logits)[source]

Calculates the cross-entropy for a batch of logits.

Parameters:
logits : array_like

The logits predicted by the model for a batch of inputs.

label : int

The label describing the target distribution.
Returns:
np.ndarray

The cross-entropy between softmax(logits[i]) and onehot(label) for all i.
foolbox.utils.binarize(x, values, threshold=None, included_in='upper')[source]

Binarizes the values of x.

Parameters:
values : tuple of two floats

The lower and upper value to which the inputs are mapped.

threshold : float

The threshold; defaults to (values[0] + values[1]) / 2 if None.

included_in : str

Whether the threshold value itself belongs to the lower or upper interval.
foolbox.utils.imagenet_example(shape=(224, 224), data_format='channels_last', bounds=(0, 255))[source]

Returns an example image and its imagenet class label.

Parameters:
shape : list of integers

The shape of the returned image.

data_format : str

“channels_first” or “channels_last”

bounds : tuple

smallest and largest allowed pixel value
Returns:
image : array_like

The example image.

label : int

The imagenet label associated with the image.

NOTE: This function is deprecated and will be removed in the future.
foolbox.utils.samples(dataset='imagenet', index=0, batchsize=1, shape=(224, 224), data_format='channels_last', bounds=(0, 255))[source]

Returns a batch of example images and the corresponding labels

Parameters:
dataset : string

The data set to load (options: imagenet, mnist, cifar10, cifar100, fashionMNIST)

index : int

For each data set 20 example images exist. The returned batch contains the images with index [index, index + 1, index + 2, …]

batchsize : int

Size of batch.

shape : list of integers

The shape of the returned image (only relevant for Imagenet).

data_format : str

“channels_first” or “channels_last”

bounds : tuple

smallest and largest allowed pixel value
Returns:
images : array_like

The batch of example images

labels : array of int

The labels associated with the images.
foolbox.utils.onehot_like(a, index, value=1)[source]

Creates an array like a, with all values set to 0 except one.

Parameters:
a : array_like

The returned one-hot array will have the same shape and dtype as this array

index : int

The index that should be set to value

value : single value compatible with a.dtype

The value to set at the given index
Returns:
numpy.ndarray

One-hot array with the given value at the given location and zeros everywhere else.

foolbox.v1.attacks

Gradient-based attacks

class foolbox.attacks.GradientAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Perturbs the input with the gradient of the loss w.r.t. the input, gradually increasing the magnitude until the input is misclassified.

Does not do anything if the model does not have a gradient.

as_generator(self, a, epsilons=1000, max_epsilon=1)[source]

Perturbs the input with the gradient of the loss w.r.t. the input, gradually increasing the magnitude until the input is misclassified.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

epsilons : int or Iterable[float]

Either Iterable of step sizes in the gradient direction or number of step sizes between 0 and max_epsilon that should be tried.

max_epsilon : float

Largest step size if epsilons is not an iterable.
class foolbox.attacks.GradientSignAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Adds the sign of the gradient to the input, gradually increasing the magnitude until the input is misclassified. This attack is often referred to as Fast Gradient Sign Method and was introduced in [R20d0064ee4c9-1].

Does not do anything if the model does not have a gradient.

References

[R20d0064ee4c9-1](1, 2) Ian J. Goodfellow, Jonathon Shlens, Christian Szegedy, “Explaining and Harnessing Adversarial Examples”, https://arxiv.org/abs/1412.6572
as_generator(self, a, epsilons=1000, max_epsilon=1)[source]

Adds the sign of the gradient to the input, gradually increasing the magnitude until the input is misclassified.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

epsilons : int or Iterable[float]

Either Iterable of step sizes in the direction of the sign of the gradient or number of step sizes between 0 and max_epsilon that should be tried.

max_epsilon : float

Largest step size if epsilons is not an iterable.
foolbox.attacks.FGSM[source]

alias of foolbox.attacks.gradient.GradientSignAttack

class foolbox.attacks.LinfinityBasicIterativeAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

The Basic Iterative Method introduced in [R37dbc8f24aee-1].

This attack is also known as Projected Gradient Descent (PGD) (without random start) or FGMS^k.

References

[R37dbc8f24aee-1](1, 2)

Alexey Kurakin, Ian Goodfellow, Samy Bengio, “Adversarial examples in the physical world”,

See also

ProjectedGradientDescentAttack

as_generator(self, a, binary_search=True, epsilon=0.3, stepsize=0.05, iterations=10, random_start=False, return_early=True)[source]

Simple iterative gradient-based attack known as Basic Iterative Method, Projected Gradient Descent or FGSM^k.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

binary_search : bool or int

Whether to perform a binary search over epsilon and stepsize, keeping their ratio constant and using their values to start the search. If False, hyperparameters are not optimized. Can also be an integer, specifying the number of binary search steps (default 20).

epsilon : float

Limit on the perturbation size; if binary_search is True, this value is only for initialization and automatically adapted.

stepsize : float

Step size for gradient descent; if binary_search is True, this value is only for initialization and automatically adapted.

iterations : int

Number of iterations for each gradient descent run.

random_start : bool

Start the attack from a random point rather than from the original input.

return_early : bool

Whether an individual gradient descent run should stop as soon as an adversarial is found.
foolbox.attacks.BasicIterativeMethod[source]

alias of foolbox.attacks.iterative_projected_gradient.LinfinityBasicIterativeAttack

foolbox.attacks.BIM[source]

alias of foolbox.attacks.iterative_projected_gradient.LinfinityBasicIterativeAttack

class foolbox.attacks.L1BasicIterativeAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Modified version of the Basic Iterative Method that minimizes the L1 distance.

See also

LinfinityBasicIterativeAttack

as_generator(self, a, binary_search=True, epsilon=0.3, stepsize=0.05, iterations=10, random_start=False, return_early=True)[source]

Simple iterative gradient-based attack known as Basic Iterative Method, Projected Gradient Descent or FGSM^k.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

binary_search : bool or int

Whether to perform a binary search over epsilon and stepsize, keeping their ratio constant and using their values to start the search. If False, hyperparameters are not optimized. Can also be an integer, specifying the number of binary search steps (default 20).

epsilon : float

Limit on the perturbation size; if binary_search is True, this value is only for initialization and automatically adapted.

stepsize : float

Step size for gradient descent; if binary_search is True, this value is only for initialization and automatically adapted.

iterations : int

Number of iterations for each gradient descent run.

random_start : bool

Start the attack from a random point rather than from the original input.

return_early : bool

Whether an individual gradient descent run should stop as soon as an adversarial is found.
class foolbox.attacks.L2BasicIterativeAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Modified version of the Basic Iterative Method that minimizes the L2 distance.

See also

LinfinityBasicIterativeAttack

as_generator(self, a, binary_search=True, epsilon=0.3, stepsize=0.05, iterations=10, random_start=False, return_early=True)[source]

Simple iterative gradient-based attack known as Basic Iterative Method, Projected Gradient Descent or FGSM^k.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

binary_search : bool or int

Whether to perform a binary search over epsilon and stepsize, keeping their ratio constant and using their values to start the search. If False, hyperparameters are not optimized. Can also be an integer, specifying the number of binary search steps (default 20).

epsilon : float

Limit on the perturbation size; if binary_search is True, this value is only for initialization and automatically adapted.

stepsize : float

Step size for gradient descent; if binary_search is True, this value is only for initialization and automatically adapted.

iterations : int

Number of iterations for each gradient descent run.

random_start : bool

Start the attack from a random point rather than from the original input.

return_early : bool

Whether an individual gradient descent run should stop as soon as an adversarial is found.
class foolbox.attacks.ProjectedGradientDescentAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

The Projected Gradient Descent Attack introduced in [R367e8e10528a-1] without random start.

When used without a random start, this attack is also known as Basic Iterative Method (BIM) or FGSM^k.

References

[R367e8e10528a-1](1, 2) Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, Adrian Vladu, “Towards Deep Learning Models Resistant to Adversarial Attacks”, https://arxiv.org/abs/1706.06083

See also

LinfinityBasicIterativeAttack and RandomStartProjectedGradientDescentAttack

as_generator(self, a, binary_search=True, epsilon=0.3, stepsize=0.01, iterations=40, random_start=False, return_early=True)[source]

Simple iterative gradient-based attack known as Basic Iterative Method, Projected Gradient Descent or FGSM^k.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

binary_search : bool or int

Whether to perform a binary search over epsilon and stepsize, keeping their ratio constant and using their values to start the search. If False, hyperparameters are not optimized. Can also be an integer, specifying the number of binary search steps (default 20).

epsilon : float

Limit on the perturbation size; if binary_search is True, this value is only for initialization and automatically adapted.

stepsize : float

Step size for gradient descent; if binary_search is True, this value is only for initialization and automatically adapted.

iterations : int

Number of iterations for each gradient descent run.

random_start : bool

Start the attack from a random point rather than from the original input.

return_early : bool

Whether an individual gradient descent run should stop as soon as an adversarial is found.
foolbox.attacks.ProjectedGradientDescent[source]

alias of foolbox.attacks.iterative_projected_gradient.ProjectedGradientDescentAttack

foolbox.attacks.PGD[source]

alias of foolbox.attacks.iterative_projected_gradient.ProjectedGradientDescentAttack

class foolbox.attacks.RandomStartProjectedGradientDescentAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

The Projected Gradient Descent Attack introduced in [Re6066bc39e14-1] with random start.

References

[Re6066bc39e14-1](1, 2) Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, Adrian Vladu, “Towards Deep Learning Models Resistant to Adversarial Attacks”, https://arxiv.org/abs/1706.06083

See also

ProjectedGradientDescentAttack

as_generator(self, a, binary_search=True, epsilon=0.3, stepsize=0.01, iterations=40, random_start=True, return_early=True)[source]

Simple iterative gradient-based attack known as Basic Iterative Method, Projected Gradient Descent or FGSM^k.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

binary_search : bool or int

Whether to perform a binary search over epsilon and stepsize, keeping their ratio constant and using their values to start the search. If False, hyperparameters are not optimized. Can also be an integer, specifying the number of binary search steps (default 20).

epsilon : float

Limit on the perturbation size; if binary_search is True, this value is only for initialization and automatically adapted.

stepsize : float

Step size for gradient descent; if binary_search is True, this value is only for initialization and automatically adapted.

iterations : int

Number of iterations for each gradient descent run.

random_start : bool

Start the attack from a random point rather than from the original input.

return_early : bool

Whether an individual gradient descent run should stop as soon as an adversarial is found.
foolbox.attacks.RandomProjectedGradientDescent[source]

alias of foolbox.attacks.iterative_projected_gradient.RandomStartProjectedGradientDescentAttack

foolbox.attacks.RandomPGD[source]

alias of foolbox.attacks.iterative_projected_gradient.RandomStartProjectedGradientDescentAttack

class foolbox.attacks.AdamL1BasicIterativeAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Modified version of the Basic Iterative Method that minimizes the L1 distance using the Adam optimizer.

See also

LinfinityBasicIterativeAttack

as_generator(self, a, binary_search=True, epsilon=0.3, stepsize=0.05, iterations=10, random_start=False, return_early=True)[source]

Simple iterative gradient-based attack known as Basic Iterative Method, Projected Gradient Descent or FGSM^k.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

binary_search : bool or int

Whether to perform a binary search over epsilon and stepsize, keeping their ratio constant and using their values to start the search. If False, hyperparameters are not optimized. Can also be an integer, specifying the number of binary search steps (default 20).

epsilon : float

Limit on the perturbation size; if binary_search is True, this value is only for initialization and automatically adapted.

stepsize : float

Step size for gradient descent; if binary_search is True, this value is only for initialization and automatically adapted.

iterations : int

Number of iterations for each gradient descent run.

random_start : bool

Start the attack from a random point rather than from the original input.

return_early : bool

Whether an individual gradient descent run should stop as soon as an adversarial is found.
class foolbox.attacks.AdamL2BasicIterativeAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Modified version of the Basic Iterative Method that minimizes the L2 distance using the Adam optimizer.

See also

LinfinityBasicIterativeAttack

as_generator(self, a, binary_search=True, epsilon=0.3, stepsize=0.05, iterations=10, random_start=False, return_early=True)[source]

Simple iterative gradient-based attack known as Basic Iterative Method, Projected Gradient Descent or FGSM^k.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

binary_search : bool or int

Whether to perform a binary search over epsilon and stepsize, keeping their ratio constant and using their values to start the search. If False, hyperparameters are not optimized. Can also be an integer, specifying the number of binary search steps (default 20).

epsilon : float

Limit on the perturbation size; if binary_search is True, this value is only for initialization and automatically adapted.

stepsize : float

Step size for gradient descent; if binary_search is True, this value is only for initialization and automatically adapted.

iterations : int

Number of iterations for each gradient descent run.

random_start : bool

Start the attack from a random point rather than from the original input.

return_early : bool

Whether an individual gradient descent run should stop as soon as an adversarial is found.
class foolbox.attacks.AdamProjectedGradientDescentAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

The Projected Gradient Descent Attack introduced in [Re2d4f39a0205-1], [Re2d4f39a0205-2] without random start using the Adam optimizer.

When used without a random start, this attack is also known as Basic Iterative Method (BIM) or FGSM^k.

References

[Re2d4f39a0205-1](1, 2) Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, Adrian Vladu, “Towards Deep Learning Models Resistant to Adversarial Attacks”, https://arxiv.org/abs/1706.06083
[Re2d4f39a0205-2](1, 2) Nicholas Carlini, David Wagner: “Towards Evaluating the Robustness of Neural Networks”, https://arxiv.org/abs/1608.04644

See also

LinfinityBasicIterativeAttack and RandomStartProjectedGradientDescentAttack

as_generator(self, a, binary_search=True, epsilon=0.3, stepsize=0.01, iterations=40, random_start=False, return_early=True)[source]

Simple iterative gradient-based attack known as Basic Iterative Method, Projected Gradient Descent or FGSM^k.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

binary_search : bool or int

Whether to perform a binary search over epsilon and stepsize, keeping their ratio constant and using their values to start the search. If False, hyperparameters are not optimized. Can also be an integer, specifying the number of binary search steps (default 20).

epsilon : float

Limit on the perturbation size; if binary_search is True, this value is only for initialization and automatically adapted.

stepsize : float

Step size for gradient descent; if binary_search is True, this value is only for initialization and automatically adapted.

iterations : int

Number of iterations for each gradient descent run.

random_start : bool

Start the attack from a random point rather than from the original input.

return_early : bool

Whether an individual gradient descent run should stop as soon as an adversarial is found.
foolbox.attacks.AdamProjectedGradientDescent[source]

alias of foolbox.attacks.iterative_projected_gradient.AdamProjectedGradientDescentAttack

foolbox.attacks.AdamPGD[source]

alias of foolbox.attacks.iterative_projected_gradient.AdamProjectedGradientDescentAttack

class foolbox.attacks.AdamRandomStartProjectedGradientDescentAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

The Projected Gradient Descent Attack introduced in [R3210aa339085-1], [R3210aa339085-2] with random start using the Adam optimizer.

References

[R3210aa339085-1]Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, Adrian Vladu, “Towards Deep Learning Models Resistant to Adversarial Attacks”, https://arxiv.org/abs/1706.06083
[R3210aa339085-2]Nicholas Carlini, David Wagner: “Towards Evaluating the Robustness of Neural Networks”, https://arxiv.org/abs/1608.04644

See also

ProjectedGradientDescentAttack

as_generator(self, a, binary_search=True, epsilon=0.3, stepsize=0.01, iterations=40, random_start=True, return_early=True)[source]

Simple iterative gradient-based attack known as Basic Iterative Method, Projected Gradient Descent or FGSM^k.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

binary_search : bool or int

Whether to perform a binary search over epsilon and stepsize, keeping their ratio constant and using their values to start the search. If False, hyperparameters are not optimized. Can also be an integer, specifying the number of binary search steps (default 20).

epsilon : float

Limit on the perturbation size; if binary_search is True, this value is only for initialization and automatically adapted.

stepsize : float

Step size for gradient descent; if binary_search is True, this value is only for initialization and automatically adapted.

iterations : int

Number of iterations for each gradient descent run.

random_start : bool

Start the attack from a random point rather than from the original input.

return_early : bool

Whether an individual gradient descent run should stop as soon as an adversarial is found.
foolbox.attacks.AdamRandomProjectedGradientDescent[source]

alias of foolbox.attacks.iterative_projected_gradient.AdamRandomStartProjectedGradientDescentAttack

foolbox.attacks.AdamRandomPGD[source]

alias of foolbox.attacks.iterative_projected_gradient.AdamRandomStartProjectedGradientDescentAttack

class foolbox.attacks.MomentumIterativeAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

The Momentum Iterative Method attack introduced in [R86d363e1fb2f-1]. It’s like the Basic Iterative Method or Projected Gradient Descent except that it uses momentum.

References

[R86d363e1fb2f-1]Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, Jianguo Li, “Boosting Adversarial Attacks with Momentum”, https://arxiv.org/abs/1710.06081
as_generator(self, a, binary_search=True, epsilon=0.3, stepsize=0.06, iterations=10, decay_factor=1.0, random_start=False, return_early=True)[source]

Momentum-based iterative gradient attack known as Momentum Iterative Method.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

binary_search : bool

Whether to perform a binary search over epsilon and stepsize, keeping their ratio constant and using their values to start the search. If False, hyperparameters are not optimized. Can also be an integer, specifying the number of binary search steps (default 20).

epsilon : float

Limit on the perturbation size; if binary_search is True, this value is only for initialization and automatically adapted.

stepsize : float

Step size for gradient descent; if binary_search is True, this value is only for initialization and automatically adapted.

iterations : int

Number of iterations for each gradient descent run.

decay_factor : float

Decay factor used by the momentum term.

random_start : bool

Start the attack from a random point rather than from the original input.

return_early : bool

Whether an individual gradient descent run should stop as soon as an adversarial is found.
foolbox.attacks.MomentumIterativeMethod[source]

alias of foolbox.attacks.iterative_projected_gradient.MomentumIterativeAttack

class foolbox.attacks.DeepFoolAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Simple and close to optimal gradient-based adversarial attack.

Implementes DeepFool introduced in [Rb4dd02640756-1].

References

[Rb4dd02640756-1]Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Pascal Frossard, “DeepFool: a simple and accurate method to fool deep neural networks”, https://arxiv.org/abs/1511.04599
as_generator(self, a, steps=100, subsample=10, p=None)[source]

Simple and close to optimal gradient-based adversarial attack.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

steps : int

Maximum number of steps to perform.

subsample : int

Limit on the number of the most likely classes that should be considered. A small value is usually sufficient and much faster.

p : int or float

Lp-norm that should be minimzed, must be 2 or np.inf.
class foolbox.attacks.NewtonFoolAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Implements the NewtonFool Attack.

The attack was introduced in [R6a972939b320-1].

References

[R6a972939b320-1]Uyeong Jang et al., “Objective Metrics and Gradient Descent Algorithms for Adversarial Examples in Machine Learning”, https://dl.acm.org/citation.cfm?id=3134635
as_generator(self, a, max_iter=100, eta=0.01)[source]
Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

max_iter : int

The maximum number of iterations.

eta : float

the eta coefficient
class foolbox.attacks.DeepFoolL2Attack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]
as_generator(self, a, steps=100, subsample=10)[source]

Simple and close to optimal gradient-based adversarial attack.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

steps : int

Maximum number of steps to perform.

subsample : int

Limit on the number of the most likely classes that should be considered. A small value is usually sufficient and much faster.

p : int or float

Lp-norm that should be minimzed, must be 2 or np.inf.
class foolbox.attacks.DeepFoolLinfinityAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]
as_generator(self, a, steps=100, subsample=10)[source]

Simple and close to optimal gradient-based adversarial attack.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

steps : int

Maximum number of steps to perform.

subsample : int

Limit on the number of the most likely classes that should be considered. A small value is usually sufficient and much faster.

p : int or float

Lp-norm that should be minimzed, must be 2 or np.inf.
class foolbox.attacks.ADefAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Adversarial attack that distorts the image, i.e. changes the locations of pixels. The algorithm is described in [Rf241e6d2664d-1], a Repository with the original code can be found in [Rf241e6d2664d-2]. References ———- .. [Rf241e6d2664d-1] Rima Alaifari, Giovanni S. Alberti, and Tandri Gauksson:

“ADef: an Iterative Algorithm to Construct Adversarial Deformations”, https://arxiv.org/abs/1804.07729
as_generator(self, a, max_iter=100, smooth=1.0, subsample=10)[source]
Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

max_iter : int > 0

Maximum number of iterations (default max_iter = 100).

smooth : float >= 0

Width of the Gaussian kernel used for smoothing. (default is smooth = 0 for no smoothing).

subsample : int >= 2

Limit on the number of the most likely classes that should be considered. A small value is usually sufficient and much faster. (default subsample = 10)
class foolbox.attacks.SaliencyMapAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Implements the Saliency Map Attack.

The attack was introduced in [R08e06ca693ba-1].

References

[R08e06ca693ba-1]Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z. Berkay Celik, Ananthram Swami, “The Limitations of Deep Learning in Adversarial Settings”, https://arxiv.org/abs/1511.07528
as_generator(self, a, max_iter=2000, num_random_targets=0, fast=True, theta=0.1, max_perturbations_per_pixel=7)[source]

Implements the Saliency Map Attack.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

max_iter : int

The maximum number of iterations to run.

num_random_targets : int

Number of random target classes if no target class is given by the criterion.

fast : bool

Whether to use the fast saliency map calculation.

theta : float

perturbation per pixel relative to [min, max] range.

max_perturbations_per_pixel : int

Maximum number of times a pixel can be modified.
class foolbox.attacks.IterativeGradientAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Like GradientAttack but with several steps for each epsilon.

as_generator(self, a, epsilons=100, max_epsilon=1, steps=10)[source]

Like GradientAttack but with several steps for each epsilon.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

epsilons : int or Iterable[float]

Either Iterable of step sizes in the gradient direction or number of step sizes between 0 and max_epsilon that should be tried.

max_epsilon : float

Largest step size if epsilons is not an iterable.

steps : int

Number of iterations to run.
class foolbox.attacks.IterativeGradientSignAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Like GradientSignAttack but with several steps for each epsilon.

as_generator(self, a, epsilons=100, max_epsilon=1, steps=10)[source]

Like GradientSignAttack but with several steps for each epsilon.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

epsilons : int or Iterable[float]

Either Iterable of step sizes in the direction of the sign of the gradient or number of step sizes between 0 and max_epsilon that should be tried.

max_epsilon : float

Largest step size if epsilons is not an iterable.

steps : int

Number of iterations to run.
class foolbox.attacks.CarliniWagnerL2Attack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

The L2 version of the Carlini & Wagner attack.

This attack is described in [Rc2cb572b91c5-1]. This implementation is based on the reference implementation by Carlini [Rc2cb572b91c5-2]. For bounds ≠ (0, 1), it differs from [Rc2cb572b91c5-2] because we normalize the squared L2 loss with the bounds.

References

[Rc2cb572b91c5-1]Nicholas Carlini, David Wagner: “Towards Evaluating the Robustness of Neural Networks”, https://arxiv.org/abs/1608.04644
[Rc2cb572b91c5-2](1, 2) https://github.com/carlini/nn_robust_attacks
as_generator(self, a, binary_search_steps=5, max_iterations=1000, confidence=0, learning_rate=0.005, initial_const=0.01, abort_early=True)[source]

The L2 version of the Carlini & Wagner attack.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

binary_search_steps : int

The number of steps for the binary search used to find the optimal tradeoff-constant between distance and confidence.

max_iterations : int

The maximum number of iterations. Larger values are more accurate; setting it too small will require a large learning rate and will produce poor results.

confidence : int or float

Confidence of adversarial examples: a higher value produces adversarials that are further away, but more strongly classified as adversarial.

learning_rate : float

The learning rate for the attack algorithm. Smaller values produce better results but take longer to converge.

initial_const : float

The initial tradeoff-constant to use to tune the relative importance of distance and confidence. If binary_search_steps is large, the initial constant is not important.

abort_early : bool

If True, Adam will be aborted if the loss hasn’t decreased for some time (a tenth of max_iterations).
static best_other_class(logits, exclude)[source]

Returns the index of the largest logit, ignoring the class that is passed as exclude.

classmethod loss_function(const, a, x, logits, reconstructed_original, confidence, min_, max_)[source]

Returns the loss and the gradient of the loss w.r.t. x, assuming that logits = model(x).

class foolbox.attacks.EADAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Gradient based attack which uses an elastic-net regularization [1]. This implementation is based on the attacks description [1] and its reference implementation [2].

References

[Rf0e4124daa63-1]Pin-Yu Chen (*), Yash Sharma (*), Huan Zhang, Jinfeng Yi, Cho-Jui Hsieh, “EAD: Elastic-Net Attacks to Deep Neural Networks via Adversarial Examples”, https://arxiv.org/abs/1709.04114
[Rf0e4124daa63-2]Pin-Yu Chen (*), Yash Sharma (*), Huan Zhang, Jinfeng Yi, Cho-Jui Hsieh, “Reference Implementation of ‘EAD: Elastic-Net Attacks to Deep Neural Networks via Adversarial Examples’”, https://github.com/ysharma1126/EAD_Attack/blob/master/en_attack.py
as_generator(self, a, binary_search_steps=5, max_iterations=1000, confidence=0, initial_learning_rate=0.01, regularization=0.01, initial_const=0.01, abort_early=True)[source]

The L2 version of the Carlini & Wagner attack.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

binary_search_steps : int

The number of steps for the binary search used to find the optimal tradeoff-constant between distance and confidence.

max_iterations : int

The maximum number of iterations. Larger values are more accurate; setting it too small will require a large learning rate and will produce poor results.

confidence : int or float

Confidence of adversarial examples: a higher value produces adversarials that are further away, but more strongly classified as adversarial.

initial_learning_rate : float

The initial learning rate for the attack algorithm. Smaller values produce better results but take longer to converge. During the attack a square-root decay in the learning rate is performed.

initial_const : float

The initial tradeoff-constant to use to tune the relative importance of distance and confidence. If binary_search_steps is large, the initial constant is not important.

regularization : float

The L1 regularization parameter (also called beta). A value of 0 corresponds to the attacks.CarliniWagnerL2Attack attack.

abort_early : bool

If True, Adam will be aborted if the loss hasn’t decreased for some time (a tenth of max_iterations).
static best_other_class(logits, exclude)[source]

Returns the index of the largest logit, ignoring the class that is passed as exclude.

classmethod loss_function(const, a, x, logits, reconstructed_original, confidence, min_, max_)[source]

Returns the loss and the gradient of the loss w.r.t. x, assuming that logits = model(x).

classmethod project_shrinkage_thresholding(z, x0, regularization, min_, max_)[source]

Performs the element-wise projected shrinkage-thresholding operation

class foolbox.attacks.DecoupledDirectionNormL2Attack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

The Decoupled Direction and Norm L2 adversarial attack from [R0e9d4da0ab48-1].

References

[R0e9d4da0ab48-1]Jérôme Rony, Luiz G. Hafemann, Luiz S. Oliveira, Ismail Ben Ayed,

Robert Sabourin, Eric Granger, “Decoupling Direction and Norm for Efficient Gradient-Based L2 Adversarial Attacks and Defenses”, https://arxiv.org/abs/1811.09600

as_generator(self, a, steps=100, gamma=0.05, initial_norm=1, quantize=True, levels=256)[source]

The Decoupled Direction and Norm L2 adversarial attack.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

steps : int

Number of steps for the optimization.

gamma : float, optional

Factor by which the norm will be modified. new_norm = norm * (1 + or - gamma).

init_norm : float, optional

Initial value for the norm.

quantize : bool, optional

If True, the returned adversarials will have quantized values to the specified number of levels.

levels : int, optional

Number of levels to use for quantization (e.g. 256 for 8 bit images).
class foolbox.attacks.SparseL1BasicIterativeAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Sparse version of the Basic Iterative Method that minimizes the L1 distance introduced in [R0591d14da1c3-1].

References

[R0591d14da1c3-1]Florian Tramèr, Dan Boneh, “Adversarial Training and Robustness for Multiple Perturbations”, https://arxiv.org/abs/1904.13000

See also

L1BasicIterativeAttack

as_generator(self, a, q=80.0, binary_search=True, epsilon=0.3, stepsize=0.05, iterations=10, random_start=False, return_early=True)[source]

Sparse version of a gradient-based attack that minimizes the L1 distance.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

q : float

Relative percentile to make gradients sparse (must be in [0, 100))

binary_search : bool or int

Whether to perform a binary search over epsilon and stepsize, keeping their ratio constant and using their values to start the search. If False, hyperparameters are not optimized. Can also be an integer, specifying the number of binary search steps (default 20).

epsilon : float

Limit on the perturbation size; if binary_search is True, this value is only for initialization and automatically adapted.

stepsize : float

Step size for gradient descent; if binary_search is True, this value is only for initialization and automatically adapted.

iterations : int

Number of iterations for each gradient descent run.

random_start : bool

Start the attack from a random point rather than from the original input.

return_early : bool

Whether an individual gradient descent run should stop as soon as an adversarial is found.
class foolbox.attacks.VirtualAdversarialAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Calculate an untargeted adversarial perturbation by performing a approximated second order optimization step on the KL divergence between the unperturbed predictions and the predictions for the adversarial perturbation. This attack was introduced in [Rc6516d158ac2-1].

References

[Rc6516d158ac2-1]Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, Ken Nakae, Shin Ishii, “Distributional Smoothing with Virtual Adversarial Training”, https://arxiv.org/abs/1507.00677
as_generator(self, a, xi=1e-05, iterations=1, epsilons=1000, max_epsilon=0.3)[source]
Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

xi : float

The finite difference size for performing the power method.

iterations : int

Number of iterations to perform power method to search for second order perturbation of KL divergence.

epsilons : int or Iterable[float]

Either Iterable of step sizes in the direction of the sign of the gradient or number of step sizes between 0 and max_epsilon that should be tried.

max_epsilon : float

Largest step size if epsilons is not an iterable.

Score-based attacks

class foolbox.attacks.SinglePixelAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Perturbs just a single pixel and sets it to the min or max.

as_generator(self, a, max_pixels=1000)[source]

Perturbs just a single pixel and sets it to the min or max.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, correctly classified input. If it is a numpy array, label must be passed as well. If it is an Adversarial instance, label must not be passed.

label : int

The reference label of the original input. Must be passed if input is a numpy array, must not be passed if input is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

max_pixels : int

Maximum number of pixels to try.
class foolbox.attacks.LocalSearchAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

A black-box attack based on the idea of greedy local search.

This implementation is based on the algorithm in [Rb320cee6998a-1].

References

[Rb320cee6998a-1](1, 2) Nina Narodytska, Shiva Prasad Kasiviswanathan, “Simple Black-Box Adversarial Perturbations for Deep Networks”, https://arxiv.org/abs/1612.06299
as_generator(self, a, r=1.5, p=10.0, d=5, t=5, R=150)[source]

A black-box attack based on the idea of greedy local search.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, correctly classified input. If it is a numpy array, label must be passed as well. If it is an Adversarial instance, label must not be passed.

label : int

The reference label of the original input. Must be passed if input is a numpy array, must not be passed if input is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

r : float

Perturbation parameter that controls the cyclic perturbation; must be in [0, 2]

p : float

Perturbation parameter that controls the pixel sensitivity estimation

d : int

The half side length of the neighborhood square

t : int

The number of pixels perturbed at each round

R : int

An upper bound on the number of iterations

Decision-based attacks

class foolbox.attacks.BoundaryAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

A powerful adversarial attack that requires neither gradients nor probabilities.

This is the reference implementation for the attack introduced in [Re72ca268aa55-1].

Notes

This implementation provides several advanced features:

  • ability to continue previous attacks by passing an instance of the Adversarial class
  • ability to pass an explicit starting point; especially to initialize a targeted attack
  • ability to pass an alternative attack used for initialization
  • fine-grained control over logging
  • ability to specify the batch size
  • optional automatic batch size tuning
  • optional multithreading for random number generation
  • optional multithreading for candidate point generation

References

[Re72ca268aa55-1](1, 2) Wieland Brendel (*), Jonas Rauber (*), Matthias Bethge, “Decision-Based Adversarial Attacks: Reliable Attacks Against Black-Box Machine Learning Models”, https://arxiv.org/abs/1712.04248
as_generator(self, a, iterations=5000, max_directions=25, starting_point=None, initialization_attack=None, log_every_n_steps=None, spherical_step=0.01, source_step=0.01, step_adaptation=1.5, batch_size=1, tune_batch_size=True, threaded_rnd=True, threaded_gen=True, alternative_generator=False, internal_dtype=<Mock name='mock.float64' id='139850469184792'>, loggingLevel=30)[source]

Applies the Boundary Attack.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, correctly classified input. If it is a numpy array, label must be passed as well. If it is an Adversarial instance, label must not be passed.

label : int

The reference label of the original input. Must be passed if input is a numpy array, must not be passed if input is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

iterations : int

Maximum number of iterations to run. Might converge and stop before that.

max_directions : int

Maximum number of trials per ieration.

starting_point : numpy.ndarray

Adversarial input to use as a starting point, in particular for targeted attacks.

initialization_attack : Attack

Attack to use to find a starting point. Defaults to BlendedUniformNoiseAttack.

log_every_n_steps : int

Determines verbositity of the logging.

spherical_step : float

Initial step size for the orthogonal (spherical) step.

source_step : float

Initial step size for the step towards the target.

step_adaptation : float

Factor by which the step sizes are multiplied or divided.

batch_size : int

Batch size or initial batch size if tune_batch_size is True

tune_batch_size : bool

Whether or not the batch size should be automatically chosen between 1 and max_directions.

threaded_rnd : bool

Whether the random number generation should be multithreaded.

threaded_gen : bool

Whether the candidate point generation should be multithreaded.

alternative_generator: bool

Whether an alternative implemenation of the candidate generator should be used.

internal_dtype : np.float32 or np.float64

Higher precision might be slower but is numerically more stable.

loggingLevel : int

Controls the verbosity of the logging, e.g. logging.INFO or logging.WARNING.
class foolbox.attacks.SpatialAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Adversarially chosen rotations and translations [1].

This implementation is based on the reference implementation by Madry et al.: https://github.com/MadryLab/adversarial_spatial

References

[Rdffd25498f9d-1]Logan Engstrom*, Brandon Tran*, Dimitris Tsipras*, Ludwig Schmidt, Aleksander Mądry: “A Rotation and a Translation Suffice: Fooling CNNs with Simple Transformations”, http://arxiv.org/abs/1712.02779
as_generator(self, a, do_rotations=True, do_translations=True, x_shift_limits=(-5, 5), y_shift_limits=(-5, 5), angular_limits=(-5, 5), granularity=10, random_sampling=False, abort_early=True)[source]

Adversarially chosen rotations and translations.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

do_rotations : bool

If False no rotations will be applied to the image.

do_translations : bool

If False no translations will be applied to the image.

x_shift_limits : int or (int, int)

Limits for horizontal translations in pixels. If one integer is provided the limits will be (-x_shift_limits, x_shift_limits).

y_shift_limits : int or (int, int)

Limits for vertical translations in pixels. If one integer is provided the limits will be (-y_shift_limits, y_shift_limits).

angular_limits : int or (int, int)

Limits for rotations in degrees. If one integer is provided the limits will be [-angular_limits, angular_limits].

granularity : int

Density of sampling within limits for each dimension.

random_sampling : bool

If True we sample translations/rotations randomly within limits, otherwise we use a regular grid.

abort_early : bool

If True, the attack stops as soon as it finds an adversarial.
class foolbox.attacks.PointwiseAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Starts with an adversarial and performs a binary search between the adversarial and the original for each dimension of the input individually.

References

[R739f80a24875-1]L. Schott, J. Rauber, M. Bethge, W. Brendel: “Towards the first adversarially robust neural network model on MNIST”, ICLR (2019) https://arxiv.org/abs/1805.09190
as_generator(self, a, starting_point=None, initialization_attack=None)[source]

Starts with an adversarial and performs a binary search between the adversarial and the original for each dimension of the input individually.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

starting_point : numpy.ndarray

Adversarial input to use as a starting point, in particular for targeted attacks.

initialization_attack : Attack

Attack to use to find a starting point. Defaults to SaltAndPepperNoiseAttack.
class foolbox.attacks.GaussianBlurAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Blurs the input until it is misclassified.

as_generator(self, a, epsilons=1000)[source]

Blurs the input until it is misclassified.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if input is a numpy.ndarray, must not be passed if input is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

epsilons : int or Iterable[float]

Either Iterable of standard deviations of the Gaussian blur or number of standard deviations between 0 and 1 that should be tried.
class foolbox.attacks.ContrastReductionAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Reduces the contrast of the input until it is misclassified.

as_generator(self, a, epsilons=1000)[source]

Reduces the contrast of the input until it is misclassified.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

epsilons : int or Iterable[float]

Either Iterable of contrast levels or number of contrast levels between 1 and 0 that should be tried. Epsilons are one minus the contrast level.
class foolbox.attacks.AdditiveUniformNoiseAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Adds uniform noise to the input, gradually increasing the standard deviation until the input is misclassified.

__call__(self, inputs, labels, unpack=True, individual_kwargs=None, **kwargs)[source]

Call self as a function.

__class__[source]

alias of abc.ABCMeta

__delattr__(self, name, /)[source]

Implement delattr(self, name).

__dir__()[source]

default dir() implementation

__eq__(self, value, /)[source]

Return self==value.

__format__()[source]

default object formatter

__ge__(self, value, /)[source]

Return self>=value.

__getattribute__(self, name, /)[source]

Return getattr(self, name).

__gt__(self, value, /)[source]

Return self>value.

__hash__(self, /)[source]

Return hash(self).

__init__(self, model=None, criterion=<foolbox.criteria.Misclassification object at 0x7f3174564748>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Initialize self. See help(type(self)) for accurate signature.

__le__(self, value, /)[source]

Return self<=value.

__lt__(self, value, /)[source]

Return self<value.

__ne__(self, value, /)[source]

Return self!=value.

__new__(*args, **kwargs)[source]

Create and return a new object. See help(type) for accurate signature.

__reduce__()[source]

helper for pickle

__reduce_ex__()[source]

helper for pickle

__repr__(self, /)[source]

Return repr(self).

__setattr__(self, name, value, /)[source]

Implement setattr(self, name, value).

__sizeof__()[source]

size of object in memory, in bytes

__str__(self, /)[source]

Return str(self).

__subclasshook__()[source]

Abstract classes can override this to customize issubclass().

This is invoked early on by abc.ABCMeta.__subclasscheck__(). It should return True, False or NotImplemented. If it returns NotImplemented, the normal algorithm is used. Otherwise, it overrides the normal algorithm (and the outcome is cached).

__weakref__[source]

list of weak references to the object (if defined)

as_generator(self, a, epsilons=1000)[source]

Adds uniform or Gaussian noise to the input, gradually increasing the standard deviation until the input is misclassified.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

epsilons : int or Iterable[float]

Either Iterable of noise levels or number of noise levels between 0 and 1 that should be tried.
name(self)[source]

Returns a human readable name that uniquely identifies the attack with its hyperparameters.

Returns:
str

Human readable name that uniquely identifies the attack with its hyperparameters.

Notes

Defaults to the class name but subclasses can provide more descriptive names and must take hyperparameters into account.

class foolbox.attacks.AdditiveGaussianNoiseAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Adds Gaussian noise to the input, gradually increasing the standard deviation until the input is misclassified.

__call__(self, inputs, labels, unpack=True, individual_kwargs=None, **kwargs)[source]

Call self as a function.

__class__[source]

alias of abc.ABCMeta

__delattr__(self, name, /)[source]

Implement delattr(self, name).

__dir__()[source]

default dir() implementation

__eq__(self, value, /)[source]

Return self==value.

__format__()[source]

default object formatter

__ge__(self, value, /)[source]

Return self>=value.

__getattribute__(self, name, /)[source]

Return getattr(self, name).

__gt__(self, value, /)[source]

Return self>value.

__hash__(self, /)[source]

Return hash(self).

__init__(self, model=None, criterion=<foolbox.criteria.Misclassification object at 0x7f3174564748>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Initialize self. See help(type(self)) for accurate signature.

__le__(self, value, /)[source]

Return self<=value.

__lt__(self, value, /)[source]

Return self<value.

__ne__(self, value, /)[source]

Return self!=value.

__new__(*args, **kwargs)[source]

Create and return a new object. See help(type) for accurate signature.

__reduce__()[source]

helper for pickle

__reduce_ex__()[source]

helper for pickle

__repr__(self, /)[source]

Return repr(self).

__setattr__(self, name, value, /)[source]

Implement setattr(self, name, value).

__sizeof__()[source]

size of object in memory, in bytes

__str__(self, /)[source]

Return str(self).

__subclasshook__()[source]

Abstract classes can override this to customize issubclass().

This is invoked early on by abc.ABCMeta.__subclasscheck__(). It should return True, False or NotImplemented. If it returns NotImplemented, the normal algorithm is used. Otherwise, it overrides the normal algorithm (and the outcome is cached).

__weakref__[source]

list of weak references to the object (if defined)

as_generator(self, a, epsilons=1000)[source]

Adds uniform or Gaussian noise to the input, gradually increasing the standard deviation until the input is misclassified.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

epsilons : int or Iterable[float]

Either Iterable of noise levels or number of noise levels between 0 and 1 that should be tried.
name(self)[source]

Returns a human readable name that uniquely identifies the attack with its hyperparameters.

Returns:
str

Human readable name that uniquely identifies the attack with its hyperparameters.

Notes

Defaults to the class name but subclasses can provide more descriptive names and must take hyperparameters into account.

class foolbox.attacks.SaltAndPepperNoiseAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Increases the amount of salt and pepper noise until the input is misclassified.

as_generator(self, a, epsilons=100, repetitions=10)[source]

Increases the amount of salt and pepper noise until the input is misclassified.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

epsilons : int

Number of steps to try between probability 0 and 1.

repetitions : int

Specifies how often the attack will be repeated.
class foolbox.attacks.BlendedUniformNoiseAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Blends the input with a uniform noise input until it is misclassified.

as_generator(self, a, epsilons=1000, max_directions=1000)[source]

Blends the input with a uniform noise input until it is misclassified.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

epsilons : int or Iterable[float]

Either Iterable of blending steps or number of blending steps between 0 and 1 that should be tried.

max_directions : int

Maximum number of random inputs to try.
class foolbox.attacks.HopSkipJumpAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

A powerful adversarial attack that requires neither gradients nor probabilities.

Notes

Features: * ability to switch between two types of distances: MSE and Linf. * ability to continue previous attacks by passing an instance of the

Adversarial class
  • ability to pass an explicit starting point; especially to initialize a targeted attack
  • ability to pass an alternative attack used for initialization
  • ability to specify the batch size

References

HopSkipJumpAttack was originally proposed by Chen, Jordan and Wainwright. It is a decision-based attack that requires access to output labels of a model alone. Paper link: https://arxiv.org/abs/1904.02144 The implementation in Foolbox is based on Boundary Attack.

approximate_gradient(self, decision_function, sample, num_evals, delta)[source]

Gradient direction estimation

as_generator(self, a, iterations=64, initial_num_evals=100, max_num_evals=10000, stepsize_search='geometric_progression', gamma=1.0, starting_point=None, batch_size=256, internal_dtype=<Mock name='mock.float64' id='139850469184792'>, log_every_n_steps=None, loggingLevel=30)[source]

Applies HopSkipJumpAttack.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, correctly classified input. If it is a numpy array, label must be passed as well. If it is an Adversarial instance, label must not be passed.

label : int

The reference label of the original input. Must be passed if input is a numpy array, must not be passed if input is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

iterations : int

Number of iterations to run.

initial_num_evals: int

Initial number of evaluations for gradient estimation. Larger initial_num_evals increases time efficiency, but may decrease query efficiency.

max_num_evals: int

Maximum number of evaluations for gradient estimation.

stepsize_search: str

How to search for stepsize; choices are ‘geometric_progression’, ‘grid_search’. ‘geometric progression’ initializes the stepsize by ||x_t - x||_p / sqrt(iteration), and keep decreasing by half until reaching the target side of the boundary. ‘grid_search’ chooses the optimal epsilon over a grid, in the scale of ||x_t - x||_p.

gamma: float
The binary search threshold theta is gamma / d^1.5 for

l2 attack and gamma / d^2 for linf attack.

starting_point : numpy.ndarray

Adversarial input to use as a starting point, required for targeted attacks.

batch_size : int

Batch size for model prediction.

internal_dtype : np.float32 or np.float64

Higher precision might be slower but is numerically more stable.

log_every_n_steps : int

Determines verbositity of the logging.

loggingLevel : int

Controls the verbosity of the logging, e.g. logging.INFO or logging.WARNING.
attack(self, a, iterations)[source]
iterations : int
Maximum number of iterations to run.
binary_search_batch(self, unperturbed, perturbed_inputs, decision_function)[source]

Binary search to approach the boundary.

geometric_progression_for_stepsize(self, x, update, dist, decision_function, current_iteration)[source]

Geometric progression to search for stepsize. Keep decreasing stepsize by half until reaching the desired side of the boundary.

project(self, unperturbed, perturbed_inputs, alphas)[source]

Projection onto given l2 / linf balls in a batch.

select_delta(self, dist_post_update, current_iteration)[source]

Choose the delta at the scale of distance between x and perturbed sample.

class foolbox.attacks.GenAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

The GenAttack introduced in [R996613153a1e-1].

This attack is performs a genetic search in order to find an adversarial perturbation in a black-box scenario in as few queries as possible.

References

[R996613153a1e-1](1, 2)

Moustafa Alzantot, Yash Sharma, Supriyo Chakraborty, Huan Zhang, Cho-Jui Hsieh, Mani Srivastava, “GenAttack: Practical Black-box Attacks with Gradient-Free Optimization”,
as_generator(self, a, generations=10, alpha=1.0, p=0.05, N=10, tau=0.1, search_shape=None, epsilon=0.3, binary_search=20)[source]

A black-box attack based on genetic algorithms. Can either try to find an adversarial perturbation for a fixed epsilon distance or perform a binary search over epsilon values in order to find a minimal perturbation. Parameters ———- inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.
labels : numpy.ndarray
Class labels of the inputs as a vector of integers in [0, number of classes).
unpack : bool
If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.
generations : int
Number of generations, i.e. iterations, in the genetic algorithm.
alpha : float
Mutation-range.
p : float
Mutation probability.
N : int
Population size of the genetic algorithm.
tau: float
Temperature for the softmax sampling used to determine the parents of the new crossover.
search_shape : tuple (default: None)
Set this to a smaller image shape than the true shape to search in a smaller input space. The input will be scaled using a linear interpolation to match the required input shape of the model.
binary_search : bool or int
Whether to perform a binary search over epsilon and using their values to start the search. If False, hyperparameters are not optimized. Can also be an integer, specifying the number of binary search steps (default 20).
epsilon : float
Limit on the perturbation size; if binary_search is True, this value is only for initialization and automatically adapted.

Other attacks

class foolbox.attacks.BinarizationRefinementAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

For models that preprocess their inputs by binarizing the inputs, this attack can improve adversarials found by other attacks. It does os by utilizing information about the binarization and mapping values to the corresponding value in the clean input or to the right side of the threshold.

as_generator(self, a, starting_point=None, threshold=None, included_in='upper')[source]

For models that preprocess their inputs by binarizing the inputs, this attack can improve adversarials found by other attacks. It does this by utilizing information about the binarization and mapping values to the corresponding value in the clean input or to the right side of the threshold.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

starting_point : numpy.ndarray

Adversarial input to use as a starting point.

threshold : float

The treshold used by the models binarization. If none, defaults to (model.bounds()[1] - model.bounds()[0]) / 2.

included_in : str

Whether the threshold value itself belongs to the lower or upper interval.
class foolbox.attacks.PrecomputedAdversarialsAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Attacks a model using precomputed adversarial candidates.

as_generator(self, a, candidate_inputs, candidate_outputs)[source]

Attacks a model using precomputed adversarial candidates.

Parameters:
input_or_adv : numpy.ndarray or Adversarial

The original, unperturbed input as a numpy.ndarray or an Adversarial instance.

label : int

The reference label of the original input. Must be passed if a is a numpy.ndarray, must not be passed if a is an Adversarial instance.

unpack : bool

If true, returns the adversarial input, otherwise returns the Adversarial object.

candidate_inputs : numpy.ndarray

The original inputs that will be expected by this attack.

candidate_outputs : numpy.ndarray

The adversarial candidates corresponding to the inputs.
class foolbox.attacks.InversionAttack(model=None, criterion=<foolbox.criteria.Misclassification object>, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None)[source]

Creates “negative images” by inverting the pixel values according to [R57cf8375f1ff-1].

References

[R57cf8375f1ff-1](1, 2)

Hossein Hosseini, Baicen Xiao, Mayoore Jaiswal, Radha Poovendran, “On the Limitation of Convolutional Neural Networks in Recognizing Negative Images”,
as_generator(self, a)[source]

Creates “negative images” by inverting the pixel values.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the underlying model.

labels : numpy.ndarray

Class labels of the inputs as a vector of integers in [0, number of classes).

unpack : bool

If true, returns the adversarial inputs as an array, otherwise returns Adversarial objects.

Gradient-based attacks

GradientAttack Perturbs the input with the gradient of the loss w.r.t.
GradientSignAttack Adds the sign of the gradient to the input, gradually increasing the magnitude until the input is misclassified.
FGSM alias of foolbox.v1.attacks.gradient.GradientSignAttack
LinfinityBasicIterativeAttack The Basic Iterative Method introduced in [Rbd27454db950-1].
BasicIterativeMethod alias of foolbox.v1.attacks.iterative_projected_gradient.LinfinityBasicIterativeAttack
BIM alias of foolbox.v1.attacks.iterative_projected_gradient.LinfinityBasicIterativeAttack
L1BasicIterativeAttack Modified version of the Basic Iterative Method that minimizes the L1 distance.
L2BasicIterativeAttack Modified version of the Basic Iterative Method that minimizes the L2 distance.
ProjectedGradientDescentAttack The Projected Gradient Descent Attack introduced in [R37229719ede6-1] without random start.
ProjectedGradientDescent alias of foolbox.v1.attacks.iterative_projected_gradient.ProjectedGradientDescentAttack
PGD alias of foolbox.v1.attacks.iterative_projected_gradient.ProjectedGradientDescentAttack
RandomStartProjectedGradientDescentAttack The Projected Gradient Descent Attack introduced in [R876f5a9eb8eb-1] with random start.
RandomProjectedGradientDescent alias of foolbox.v1.attacks.iterative_projected_gradient.RandomStartProjectedGradientDescentAttack
RandomPGD alias of foolbox.v1.attacks.iterative_projected_gradient.RandomStartProjectedGradientDescentAttack
AdamL1BasicIterativeAttack Modified version of the Basic Iterative Method that minimizes the L1 distance using the Adam optimizer.
AdamL2BasicIterativeAttack Modified version of the Basic Iterative Method that minimizes the L2 distance using the Adam optimizer.
AdamProjectedGradientDescentAttack The Projected Gradient Descent Attack introduced in [R78a2267bf0c5-1], [R78a2267bf0c5-2] without random start using the Adam optimizer.
AdamProjectedGradientDescent alias of foolbox.v1.attacks.iterative_projected_gradient.AdamProjectedGradientDescentAttack
AdamPGD alias of foolbox.v1.attacks.iterative_projected_gradient.AdamProjectedGradientDescentAttack
AdamRandomStartProjectedGradientDescentAttack The Projected Gradient Descent Attack introduced in [Rb42f1f35d85c-1], [Rb42f1f35d85c-2] with random start using the Adam optimizer.
AdamRandomProjectedGradientDescent alias of foolbox.v1.attacks.iterative_projected_gradient.AdamRandomStartProjectedGradientDescentAttack
AdamRandomPGD alias of foolbox.v1.attacks.iterative_projected_gradient.AdamRandomStartProjectedGradientDescentAttack
MomentumIterativeAttack The Momentum Iterative Method attack introduced in [R0c7c08fb6fc4-1].
MomentumIterativeMethod alias of foolbox.v1.attacks.iterative_projected_gradient.MomentumIterativeAttack
LBFGSAttack Uses L-BFGS-B to minimize the distance between the input and the adversarial as well as the cross-entropy between the predictions for the adversarial and the the one-hot encoded target class.
DeepFoolAttack Simple and close to optimal gradient-based adversarial attack.
NewtonFoolAttack Implements the NewtonFool Attack.
DeepFoolL2Attack
DeepFoolLinfinityAttack
ADefAttack Adversarial attack that distorts the image, i.e.
SLSQPAttack Uses SLSQP to minimize the distance between the input and the adversarial under the constraint that the input is adversarial.
SaliencyMapAttack Implements the Saliency Map Attack.
IterativeGradientAttack Like GradientAttack but with several steps for each epsilon.
IterativeGradientSignAttack Like GradientSignAttack but with several steps for each epsilon.
CarliniWagnerL2Attack The L2 version of the Carlini & Wagner attack.
EADAttack Gradient based attack which uses an elastic-net regularization [1].
DecoupledDirectionNormL2Attack The Decoupled Direction and Norm L2 adversarial attack from [R1326043d948c-1].
SparseFoolAttack A geometry-inspired and fast attack for computing sparse adversarial perturbations.
SparseL1BasicIterativeAttack
VirtualAdversarialAttack

Score-based attacks

SinglePixelAttack Perturbs just a single pixel and sets it to the min or max.
LocalSearchAttack A black-box attack based on the idea of greedy local search.
ApproximateLBFGSAttack Same as LBFGSAttack with approximate_gradient set to True.

Decision-based attacks

BoundaryAttack A powerful adversarial attack that requires neither gradients nor probabilities.
SpatialAttack Adversarially chosen rotations and translations [1].
PointwiseAttack Starts with an adversarial and performs a binary search between the adversarial and the original for each dimension of the input individually.
GaussianBlurAttack Blurs the input until it is misclassified.
ContrastReductionAttack Reduces the contrast of the input until it is misclassified.
AdditiveUniformNoiseAttack Adds uniform noise to the input, gradually increasing the standard deviation until the input is misclassified.
AdditiveGaussianNoiseAttack Adds Gaussian noise to the input, gradually increasing the standard deviation until the input is misclassified.
SaltAndPepperNoiseAttack Increases the amount of salt and pepper noise until the input is misclassified.
BlendedUniformNoiseAttack Blends the input with a uniform noise input until it is misclassified.
BoundaryAttackPlusPlus
GenAttack
HopSkipJumpAttack A powerful adversarial attack that requires neither gradients nor probabilities.

Other attacks

BinarizationRefinementAttack For models that preprocess their inputs by binarizing the inputs, this attack can improve adversarials found by other attacks.
PrecomputedAdversarialsAttack Attacks a model using precomputed adversarial candidates.
InversionAttack

foolbox.v1.adversarial

Provides a class that represents an adversarial example.

class foolbox.v1.adversarial.Adversarial(model, criterion, unperturbed, original_class, distance=<class 'foolbox.distances.MeanSquaredDistance'>, threshold=None, verbose=False)[source]

Defines an adversarial that should be found and stores the result.

The Adversarial class represents a single adversarial example for a given model, criterion and reference input. It can be passed to an adversarial attack to find the actual adversarial perturbation.

Parameters:
model : a Model instance

The model that should be fooled by the adversarial.

criterion : a Criterion instance

The criterion that determines which inputs are adversarial.

unperturbed : a numpy.ndarray

The unperturbed input to which the adversarial input should be as close as possible.

original_class : int

The ground-truth label of the unperturbed input.

distance : a Distance class

The measure used to quantify how close inputs are.

threshold : float or Distance

If not None, the attack will stop as soon as the adversarial perturbation has a size smaller than this threshold. Can be an instance of the Distance class passed to the distance argument, or a float assumed to have the same unit as the the given distance. If None, the attack will simply minimize the distance as good as possible. Note that the threshold only influences early stopping of the attack; the returned adversarial does not necessarily have smaller perturbation size than this threshold; the reached_threshold() method can be used to check if the threshold has been reached.
adversarial_class[source]

The argmax of the model predictions for the best adversarial found so far.

None if no adversarial has been found.

backward_one(self, gradient, x=None, strict=True)[source]

Interface to model.backward_one for attacks.

Parameters:
gradient : numpy.ndarray

Gradient of some loss w.r.t. the logits.

x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension).
Returns:
gradient : numpy.ndarray

The gradient w.r.t the input.

See also

gradient()
channel_axis(self, batch)[source]

Interface to model.channel_axis for attacks.

Parameters:
batch : bool

Controls whether the index of the axis for a batch of inputs (4 dimensions) or a single input (3 dimensions) should be returned.
distance[source]

The distance of the adversarial input to the original input.

forward(self, inputs, greedy=False, strict=True, return_details=False)[source]

Interface to model.forward for attacks.

Parameters:
inputs : numpy.ndarray

Batch of inputs with shape as expected by the model.

greedy : bool

Whether the first adversarial should be returned.

strict : bool

Controls if the bounds for the pixel values should be checked.
forward_and_gradient(self, x, label=None, strict=True, return_details=False)[source]

Interface to model.forward_and_gradient_one for attacks.

Parameters:
x : numpy.ndarray

Multiple input with shape as expected by the model (with the batch dimension).

label : numpy.ndarray

Labels used to calculate the loss that is differentiated. Defaults to the original label.

strict : bool

Controls if the bounds for the pixel values should be checked.
forward_and_gradient_one(self, x=None, label=None, strict=True, return_details=False)[source]

Interface to model.forward_and_gradient_one for attacks.

Parameters:
x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension). Defaults to the original input.

label : int

Label used to calculate the loss that is differentiated. Defaults to the original label.

strict : bool

Controls if the bounds for the pixel values should be checked.
forward_one(self, x, strict=True, return_details=False)[source]

Interface to model.forward_one for attacks.

Parameters:
x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension).

strict : bool

Controls if the bounds for the pixel values should be checked.
gradient_one(self, x=None, label=None, strict=True)[source]

Interface to model.gradient_one for attacks.

Parameters:
x : numpy.ndarray

Single input with shape as expected by the model (without the batch dimension). Defaults to the original input.

label : int

Label used to calculate the loss that is differentiated. Defaults to the original label.

strict : bool

Controls if the bounds for the pixel values should be checked.
has_gradient(self)[source]

Returns true if _backward and _forward_backward can be called by an attack, False otherwise.

normalized_distance(self, x)[source]

Calculates the distance of a given input x to the original input.

Parameters:
x : numpy.ndarray

The input x that should be compared to the original input.
Returns:
Distance

The distance between the given input and the original input.
original_class[source]

The class of the original input (ground-truth, not model prediction).

output[source]

The model predictions for the best adversarial found so far.

None if no adversarial has been found.

perturbed[source]

The best adversarial example found so far.

reached_threshold(self)[source]

Returns True if a threshold is given and the currently best adversarial distance is smaller than the threshold.

target_class[source]

Interface to criterion.target_class for attacks.

unperturbed[source]

The original input.

Indices and tables