foolbox

BoundaryAttack less verbose

BoundaryAttack added: a new, powerful decision-based attack that does not rely on gradients or probabilities

Performance improvements and minor changes. Easier to write fast attacks.

several bugfixes

Fixes a bug in ContrastAttack when the lower bound is non-zero (thanks to @whg1g15)

MXNet model bugfixes
improved support for other Keras backends (in particular Theano; in addition to TensorFlow)

Keras model improvements

MXNet model improvements

Preprocessing has been changed and fixed.

• Preprocessing can now be specified in all models
• Preprocessing must now be specified as subtraction and division only
• Gradients are now correctly backpropagated through the preprocessing
• Gradients are now tested numerically

fix for the Readme on PyPI

Our technical report is now available on arXiv: https://arxiv.org/abs/1707.04131

added support for a preprocessing callback to the PyTorch model

Bugfix for the PyTorch interface when using GPUs, contributed by @ducha-aiki.

Support for older versions of setuptools.

• MXNet support
• CompositeModel for blackbox transfer attack

added support for Python 2.7

Improved Distances

• value and gradient can now be accessed using properties rather than methods
• the distances are now relative to the bounds (and therefore always between 0 and 1) and much more interpretable

Models

• TensorFlow
• PyTorch
• Keras
• Theano
• Lasagne
• and more

Attacks

• GradientSignAttack (FGSM)
• IterativeGradientSignAttack
• GradientAttack
• IterativeGradientAttack
• LBFGSAttack
• DeepFoolAttack
• SaliencyMapAttack
• GaussianBlurAttack
• ContrastReductionAttack
• SinglePixelAttack
• LocalSearchAttack
• SLSQPAttack
• AdditiveUniformNoiseAttack
• AdditiveGaussianNoiseAttack
• SaltAndPepperNoiseAttack
• and more

Criteria

• Misclassification
• TopKMisclassification
• OriginalClassProbability
• TargetClass
• TargetClassProbability

Distances

• MeanSquaredDistance (MSE)
• MeanAbsoluteDistance