Problems about PGD attack

[YuhengHuang42]

Dear author,

Thanks for your brilliant work. After reading your paper, I am interested in the performance of pixeldp_cnn under PGD attack. I notice that you have implemented one in this repo (Just as the paper mentioned in the Appendix). However, It seems that something is missing.

First of all, in the train_attack.py, the initialization of the attack:
attack = attack_class.Attack
is only for the Carlini. However Carlini has different parameters with PGD for initialization.

What's more, in the pgd.py, parameters like
self.min_pix = min_pix
self.max_pix = max_pix
are also missing.

[m-lecuyer]

Ah yes we probably messed things up when refactoring before putting the code online, sorry. It should work if you pass the correct constructor in train_attack.train_one().
I don't have all the code details in mind anymore unfortunately, but it looks like min/max_pix were replaced with boxmin/max to be consistent with the Carlini attack, so you'll have to replace that too.