FAB: a Fast Adaptive Boundary Attack

This is the code relative to the method introduced in

Minimally distorted Adversarial Examples with a Fast Adaptive Boundary Attack
Francesco Croce, Matthias Hein
University of Tübingen
https://arxiv.org/pdf/1907.02044.pdf

We propose a new white-box adversarial attack against neural networks-based classifiers. FAB-attack aims at changing the classification of a clean input applying a perturbation with minimal Lp-norm, for p in {1, 2, inf}. It achieves quickly good quality results, does not need the specification of a step size and tries to track the desicion boundary.

News

• The paper is accepted at ICML 2020!
• FAB attack is included in AutoAttack, a new parameter-free protocol to evaluate adversarial robustness!
• An optimized PyTorch implementation of FAB attack is now available here in Advertorch!

Running the attack

We provide here, in the folder models, classifiers on MNIST and CIFAR-10, trained with either natural training (plain), adversarial training wrt the L2-norm (l2-at) or wrt the Linf-norm (linf-at). In the folder datasets, available at the same link, we provide also the datasets in the format consistent with the scripts.

With

python test_attack.py --dataset mnist --model plain --bs 1000 --im 1000 --p linf --n_iter 100 --n_restarts 3 --eps 0.3

one would run FAB-attack on the plain model on MNIST wrt the Linf-norm (that is the attack aims at minimizing the Linf-norm of the adversarial perturbations), using 100 iterations and 3 restarts (eps defines the region where to sample the random starting points). It returns adversarial examples for the first 1000 images of the test set. More informations about the parameters are available in test_attack.py.

The FAB-attack is implemented in FAB_linf.py, FAB_l2.py and FAB_l1.py. In order to run the attack on other classifiers, it is sufficient to define a model as in utils.Model. Then, e.g., FABattack_linf(model, x_input, y_input, sess, hps) performs the Linf attack on the model.

Results

FAB-attack achieves top results in the two challenges at https://github.com/yaodongyu/TRADES and the one at https://github.com/MadryLab/cifar10_challenge.

Citations

@inproceedings{croce2020minimally,
  author    = {F. Croce and M. Hein},
  title     = {Minimally distorted Adversarial Examples with a Fast Adaptive Boundary Attack},
  booktitle = {ICML},
  year      = {2020}
}