huthvincent / weights-robustify

Code repository for model weights robustifying research

Base on the official code from AWP Paper (NeurIPS 2020 "Adversarial Weight Perturbation Helps Robust Generalization"). Github link https://github.com/csdongxian/AWP.

Baseline Experiments

Some notes:

• Train:
  • Specify the output path before each realization with arg --fname to avoid confounding.
  • If possible, run multiple different realizations (arg --seed) for certain configuration, to measure performance in terms of mean and std. Results would be output into the folder f'{fname}/{seed}'.
  • Default backbone architecture is PreActResNet18.
  • Model with best train robust loss would be newly saved as candidate.
• Test:
  • Default test process would be conducted after training via attack pgd-20 with target metrics natural accuracy and robust accuracy.
  • Autoattack could be launched via python AT-AWP/eval_autoattack.py for both training and test sets, with standard attacks apgd-ce, apgd-t, fab-t, and square. Note that args --norm and --epsilon shall be kept in consistency with that in training. For instance, in setting 1 of cifar-10, run python AT-AWP/eval_autoattack.py --data CIFAR10 --norm L2 --epsilon 128/255.

Cifar-10 Dataset

Setting 1

Training attack pgd-10 with L2 norm, epsilon 128/255, and step-size (pgd-alpha) 15/255. Example command as follows python AT-AWP/train_cifar10.py --norm l_2 --pgd-alpha 15 --epsilon 128

Setting 2

Training attack pgd-10 with Linf norm, epsilon 8/255, and step-size (pgd-alpha) 2/255. Example command as follows python AT-AWP/train_cifar10.py --norm l_inf --pgd-alpha 2 --epsilon 8

SVHN Dataset