Code for Paper "Imbalanced Gradients: A Subtle Cause of Overestimated Adversarial Robustness" by Xingjun Ma, Linxi Jiang, Hanxun Huang, Zejia Weng, James Bailey, Yu-Gang Jiang
python main.py --defence [Choice from defence models] \
--attack [MD, MDMT, MDE] \
--eps 8 --bs 100
bs
as batch size.eps
as the epsilon.- Defence models evaluated in the paper are available in the defence folder.
- The following attacks are implemented
['MD', 'MDMT', 'MDE', 'PGD', 'CW', 'PGD-ODI']
, Auto Attacks aviliable at this link
If you use this code in your work, please cite the accompanying paper:
@article{ma2023imbalanced,
title={Imbalanced gradients: a subtle cause of overestimated adversarial robustness},
author={Ma, Xingjun and Jiang, Linxi and Huang, Hanxun and Weng, Zejia and Bailey, James and Jiang, Yu-Gang},
journal={Machine Learning},
pages={1--26},
year={2023},
publisher={Springer}
}
- MART: https://github.com/YisenWang/MART
- TREADES: https://github.com/yaodongyu/TRADES
- RST: https://github.com/yaircarmon/semisup-adv
- AutoAttack: https://github.com/fra31/auto-attack
- ODI-PGD https://github.com/ermongroup/ODS
- Linxi Jiang implementation on MD attack https://github.com/Jack-lx-jiang/MD_attacks