Version 0.8.0

New Features

• value_and_grad is now part of EagerPy and has a much nicer API (works with EagerPy tensors)
• other improvements and bugfixes

New Attacks

• added noise attacks
  • L2AdditiveGaussianNoiseAttack
  • L2AdditiveUniformNoiseAttack
  • LinfAdditiveUniformNoiseAttack
  • L2RepeatedAdditiveGaussianNoiseAttack
  • L2RepeatedAdditiveUniformNoiseAttack
  • LinfRepeatedAdditiveUniformNoiseAttack
• added GaussianBlurAttack
• added the Brendel-Bethge attack
  • L0BrendelBethgeAttack
  • L1BrendelBethgeAttack
  • L2BrendelBethgeAttack
  • LinfinityBrendelBethgeAttack
• added DeepFool
  • L2DeepFoolAttack
  • LinfDeepFoolAttack
• added DatasetAttack

Version 0.7.0

New Features

• added an fbn.plot.images() utility function
• added fbn.norms.l2()
• added a prototype of fbn.evaluate_l2()
• added support for TensorBoard (currently used by the BoundaryAttack)
• lot's of bugfixes and improvements

New Attacks

• added L2ContrastReductionAttack, BinarySearchContrastReductionAttack, LinearSearchContrastReductionAttack
• added InversionAttack
• added EADAttack
• added LinearSearchBlendedUniformNoiseAttack
• added BinarizationRefinementAttack
• added L2AdditiveGaussianNoiseAttack , L2RepeatedAdditiveGaussianNoiseAttack
• added BoundaryAttack

Version 0.6.0

New Features

• added value_and_grad to PyTorchModel, TensorFlowModel and JAXModel to differentiate arbitrary loss functions natively in all frameworks
• added the Carlini Wagner L2 attack (works natively with PyTorch, TensorFlow and JAX) using value_and_grad
• added a random_start argument to the L-inf Basic Iterative Method
• added PGD
• added atleast_kd to utils
• bug fixes

Version 0.5.0

New Features

• added the Fast Gradient Sign Method (L-infinity)
• added the Fast Gradient Method (L2)
• changed default epsilon of L2 attacks
• improved tests
• bug fixes

Version 0.4.0

New Features

• Support for JAX models

Version 0.3.0

New Features

• Real-world examples for PyTorch and TensorFlow with native performance
• fbn.utils.samples with support for PyTorch and TensorFlow
• Full support for axis and flip_axis arguments to preprocessing (in addition to mean and std)
• Faster preprocessing
• fbn.models.FoolboxModel to use classic Foolbox models with Foolbox Native
• Lot's of bugfixes

Version 0.2.0

New Features

• Support for TensorFlow models

Version 0.1.0

New Features

• Support for PyTorch models

New Attacks

• L2 Basic Iterative Method
• L-infinity Basic Iterative Method