RobustBench / robustbench

RobustBench: a standardized adversarial robustness benchmark [NeurIPS 2021 Benchmarks and Datasets Track]
https://robustbench.github.io
Other
664 stars 99 forks source link

Pretrain ResNet-50 with autoattack #170

Closed AlbertoSinigaglia closed 8 months ago

AlbertoSinigaglia commented 8 months ago

Hi everybody,
First, thanks for the repo and the work you are carrying out... a very small question: do you plan also to try to train on ImageNet a ResNet-50 (or other architectures) and see if adversarial pretraining is effective? All the models you offer seem to be some sort of adversarial defense "mythbusting", but I can't find any (adversarial) pretraining using AutoAttack

dedeswim commented 8 months ago

Hi, do you mean running adversarial training using AutoAttack as an attack for generating adversarial examples at each training step? If so, then it's not really practical because of the time needed to run AutoAttack. AutoAttack is meant as an attack to evaluate robustness, rather than as an attack to do adversarial training with. Please let us know if this helps!

AlbertoSinigaglia commented 8 months ago

@dedeswim exactly, using adversarial training with AutoAttack as an adversarial defense, like what Madry did here with projected gradient descent

dedeswim commented 8 months ago

I understand. However, we don't train models ourselves, but rather benchmark and host models trained by other researchers, so we are not planning to train such model. I hope this helps!

AlbertoSinigaglia commented 8 months ago

Hi @dedeswim yup, just wanted to be sure that no model was available before training it myself, thanks anyway!

dedeswim commented 8 months ago

No problem!