Robust Weight Perturbation for Adversarial Training

This branch is developed for adversarial training with robust weight perturbation, the related paper is as follows:

Robust Weight Perturbation for Adversarial Training[C]
Chaojian Yu, Bo Han, Mingming Gong, Li Shen, Shiming Ge, Du Bo, Tongliang Liu
IJCAI. 2022.

Requisite

This code is implemented in PyTorch, and we have tested the code under the following environment settings:

• python = 3.7.3
• torch = 1.2.0
• torchvision = 0.4.0

What is in this repository

Codes for our RWP-based adversarial training (AT-RWP) are in AT_AWP_RWP; RWP-based TRADES (TRADES-RWP) are in trades_AWP_RWP, and RWP-based RST (RST-RWP) are in RST_AWP_RWP:

• In ./AT_AWP_RWP, the codes for CIFAR-10, CIFAR-100, and SVHN are in train_cifar10.py, train_cifar100.py, train_svhn.py respectively.
• In ./trades_AWP_RWP, the codes for CIFAR-10 are in train_trades_cifar.py.
• In ./RST_AWP_RWP, the codes for CIFAR-10 are in robust_self_training.py.

Robustness evaluation codes are in Robustness Evaluation:

• The codes for robustness evaluation on FGSM, PGD-20, PGD-100, C&W are in eval_pgd.py.
• The codes for robustness evaluation on AA are in eval_aa.py.

How to use it

For AT-RWP with a PreAct ResNet-18 on CIFAR-10 under L_inf threat model (8/255), run codes as follows,

python train_cifar10.py

For TRADES-RWP with a WRN-34-10 on CIFAR10 under L_inf threat model (8/255), run codes as follows,

python train_trades_cifar.py

For RST-RWP with a WRN-28-10 on CIFAR10 under L_inf threat model (8/255), you first need to download 500K unlabeled data from TinyImages (with pseudo-labels), and then run codes as follows,

python robust_self_training.py

Citation

If you find our code helpful, please consider citing our work:

@article{yu2022robust,
  title={Robust Weight Perturbation for Adversarial Training},
  author={Yu, Chaojian and Han, Bo and Gong, Mingming and Shen, Li and Ge, Shiming and Du, Bo and Liu, Tongliang},
  journal={arXiv preprint arXiv:2205.14826},
  year={2022}
}

Reference Code

[1] AT: https://github.com/locuslab/robust_overfitting

[2] TRADES: https://github.com/yaodongyu/TRADES/

[3] RST: https://github.com/yaircarmon/semisup-adv

[4] AWP: https://github.com/csdongxian/AWP

[5] AutoAttack: https://github.com/fra31/auto-attack