Fast-Minimum-Norm (FMN) Attack

The Fast Minimum Norm Attack (FMN), from Fast Minimum-norm Adversarial Attacks through Adaptive Norm Constraints.

:dart: Accepted at NeurIPS 2021! Paper available at this link.

🎉 Now available also in Foolbox, SecML and Adversarial Library.

:video_game: For a quick demo example, check out this notebook.

:pencil: For a more complete example, with different datasets and robust models, check out the full example notebook.

Here is a conceptual figure of the attack. In summary, the algorithm performs normalized gradient descent and projects into an epsilon Lp-ball which is adapted to find the minimum norm adversarials.

GIF created with SecML library.

Using FMN from this repository

from src.attacks.fmn import L1FMNAttack
import foolbox as fb

model = ...  # pytorch model
fb_model = fb.models.PyTorchModel(model)
attack = L1FMNAttack()
advs, _, is_adv = attack(fb_model, images, criterion, epsilons=None)

Using FMN from Foolbox

import foolbox as fb

model = ...  # pytorch model
fb_model = fb.models.PyTorchModel(model)
attack = fb.attacks.L1FMNAttack()
advs, _, is_adv = attack(fb_model, samples, labels, epsilons=None)

Using FMN from SecML (using the Foolbox Wrapper)

import foolbox as fb
from secml.adv.attacks.evasion import CAttackEvasionFoolbox

model = ...  # pytorch model
secml_model = CClassifierPyTorch(model=model, pretrained=True, ...)  # wraps pytorch model in Secml
attack = CAttackEvasionFoolbox(secml_model, y_target=None, fb_attack_class=fb.attacks.L1FMNAttack)
y_pred, _, adv_ds, _ = attack.run(samples, labels)

Using FMN from Adversarial Library

from adv_lib.attacks import fmn

model = ...  # pytorch model
norm = 1  # will use L1 norm
results = fmn(model, inputs, labels, norm)

Preview of the results

These are results against a MNIST 9-layer ConvNet. Check out the notebooks for more examples.

:book: Cite this work

If you use FMN in your work, please cite us using the following BibTeX entry:

@article{pintor2021fast,
  title={Fast minimum-norm adversarial attacks through adaptive norm constraints},
  author={Pintor, Maura and Roli, Fabio and Brendel, Wieland and Biggio, Battista},
  journal={Advances in Neural Information Processing Systems},
  volume={34},
  year={2021}
}