peustr / wca-net Goto Github PK

Thanks for releasing the code. I have reproduced the paper results, but when I test it on AutoAttack, the robust accuracy is close to zero, and I found that in AutoAttack, it is robust to APGD-CE and Square attack, but it performs worse in fab and apgd-t, so is it a fake defense method?

In your paper, three loss are used for optimiztion, cross-entropy loss, wca loss and L2 regularization loss.
But I found the L2 regularization loss, is not used in train.py.

• scale_tril (Tensor) – lower-triangular factor of covariance, with positive-valued diagonal (from pytorch document)
  I check the self.L and find the diagonals of this matrix have negative values. So, maybe it's the reason.

Please check it. Thanks!

Could you upload the pre-trained models and instructions for loading them, so we can verify the results reported in the paper? Thanks!