The fooling rate from using your code are much higher than those mentioned in the paper

psandovalsegura / pytorch-gd-uap

Generalized Data-free Universal Adversarial Perturbations in PyTorch

MIT License

16 stars 0 forks source link

The fooling rate from using your code are much higher than those mentioned in the paper #3

Open Aristozhang opened 3 years ago

Aristozhang commented 3 years ago

hello,I just reproduced your code. Using UAP existing under perturbations folder and ILSVRC2012 for verification, I found that the fooling rate was much higher than that in the paper. For example, UAP generated by VGG16-no- data was used.When the ILSVRC data set and VGG-16 model are used to verify, the fooling rate is 0.909, while the one mentioned in this paper is 0.63 .What is the reason for this?

psandovalsegura commented 3 years ago

It's likely a normalization issue, but I'll need to check. This was also mentioned here https://github.com/psandovalsegura/pytorch-gd-uap/issues/1#issue-739900244