Constrained attacks

Description

Constrained attacks is a framework for constraints adversarial examples unified across multiple constraints' domain. It currently supports a large diversity of constraints (linear and non-linear). We instantiated our framework with two attacks:

• MoEvA2: a multi-objective genetic based approach
• C-PGD: a gradient based approach extended from PGD (cite) to support domain constraints.

To learn more, check out our paper A Unified Framework for Adversarial Attack and Defense in Constrained Feature Space.

Installation

Using pip

pip install constrained-attacks

Dependencies

constrained-attacks requires:

• python = "~3.8"
• numpy = "^1.22.3"
• joblib = "^1.1.0"
• pymoo = "^0.5.0"
• tqdm = "^4.63.1"
• pandas = "^1.4.1"

Additional optional requirements for C-PGD are:

• tensorflow = "2.8"
• adversarial-robustness-toolbox[tensorflow] = "1.10"

Examples

You can find a usage example

• for MoEvA2: tests/attacks/moeva/test_moeva_run.py
• for C-PGD: tests/attacks/cpgd/test_pgd_run.py
• for the constraints definition: tests/attacks/moeva/url_constraints.py.

Citation

If you have used our framework for research purposes, you can cite our publication by:

BibTex:

@article{simonetto2021unified,
  title={A unified framework for adversarial attack and defense in constrained feature space},
  author={Simonetto, Thibault and Dyrmishi, Salijona and Ghamizi, Salah and Cordy, Maxime and Traon, Yves Le},
  journal={arXiv preprint arXiv:2112.01156},
  year={2021}
}