robust_overfitting

robust_overfitting copied to clipboard

experience on SVHN outputs anormally high confidence on adversarial accuracy

2

Hello. Thank you for opening your code and experience log. While running your code to train SVHN, I found that training SVHN gives strangely high adversarial accuracy on test set....

feather0011

I noticed that the labels of the validation set are slightly unbalanced, something like this: Counter({3: 113, 1: 112, 5: 107, 0: 100, 6: 99, 7: 98, 9: 98, 2:...

nagisa-eevee

``` elif norm == "l_2": g_norm = torch.norm(g.view(g.shape[0],-1),dim=1).view(-1,1,1,1) scaled_g = g/(g_norm + 1e-10) d = (d + scaled_g*alpha).view(d.size(0),-1).renorm(p=2,dim=0,maxnorm=epsilon).view_as(d) ``` Why is the `renorm` operation in the L2-norm PGD attack, i.e.,...

erhul