search

nblt / Sub-AT

[CVPR 2022 oral] Subspace Adversarial Training
26 stars 4 forks source link

Purpose of `attack_pgd` function #4

Open ShengYun-Peng opened 2 years ago

ShengYun-Peng commented 2 years ago

Thanks authors for the great work. While reviewing the code, I am confused why overwriting the pgd attack again in utils.py though it was already imported from advertorch.attack.

Best, Anthony

nblt commented 2 years ago

Hi Anthony,

Thanks for your question. This pgd attack is specifically used for evaluating the robustness under PGD-50 attack (with 10 random starts), following previous implementation [1], for a fair comparison with previous methods (keeping the implementation the same).

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

ShengYun-Peng commented 2 years ago

Ah, I see! Thanks so much for the immediate response. Btw, have you expanded the Sub-AT on full ImageNet?

Best, Anthony

nblt commented 2 years ago

We have tried on Tiny-ImageNet, but haven't expanded on full ImageNet yet. We think that similar results can be extended to full ImageNet:)

Best, Tao