You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Thanks for your interest in our work. Yes, this was an oversight in our paper. For Fog, Gabor, and Snow, the step size should be sqrt(0.001 / steps), since distortion size is controlled differently for these attacks. In particular, a perturbation is generated by optimizing parameters in a latent space of fixed size, scaled by distortion size epsilon, and then added to the original image. This means that the step_size should be independent of epsilon in this case. We have corrected this in the updated version of our paper at: https://arxiv.org/abs/1908.08016.
In the paper, it's mentioned that step_size = bound / sqrt(iterations). But based on the output files, e.g.
analysis/calibrations/imagenet-100/fog.out
analysis/calibrations/imagenet-100/snow.out
The step sizes seem to be a constant 0.002236 no matter what bound is used. Are these the expected setup?
The text was updated successfully, but these errors were encountered: