Stepsize for PGD attacks

[Changgun-Choi]

Hello, Q.1 is it possible to give list of epsilon for LinfPGD attacks? I wanted to test how a number of steps affect the result in terms of PGD attacks. Q. 2 I would like to know how we could decide on stepsize which is abs_stepsize? I read the documents but seems like no explanation about it. Is there a strategy for deciding on stepsize?

My code:

eps = [0, 0.1/255, 0.3/255, 1/255, 4/255] stepsize = [i/4 for i in eps] steps= args.steps
attack = LinfPGD(rel_stepsize=0.033, abs_stepsize=stepsize, steps=steps, random_start=True)

[zimmerrol]

Hi. Sadly, I don't fully understand your first question. Regarding your second question: please refer to the literature for this - a common choice is to choose the step size such that the equation n_steps * step_size = epsilon * k for k>1 is satisfied. I believe, Madry et al. first proposed this relation with something like k=2.5.