glin2022
commented
2 weeks ago - The image above shows the output from cifar-miss-gau-adv.ipynb. All accuracies have reached over 80%, and the situation you described did not occur.
- In previous experiments, we observed similar cases but not red dots. In addition, for the generator-based AP methods, the purifier model $g$ is not required to satisfy $g(x+\delta)=x$. In general, we only focus the classification accuracy.
So, if the quality of the generated images is poor, it is normal. However, because your accuracy is low, I cannot determine whether these red dots are normal.
P.S. Recently, I have been preparing a submission for CVPR and preparing a rebuttal for another conference in the next several weeks. For some general questions, I will reply as soon as possible. For others, I may check later.
erhul
Rokishii
Elysia200207
As shown in the training screenshot of cifar-miss-gau-adv.py (I converted cifar-miss-gau-adv.ipynb into a .py file), after 69 epochs of training, the robust accuracy under clean samples and various attacks has only increased from the initial 13% to about 46%, and the increase is unstable. Therefore, there is a considerable gap with the 90% in the original paper. Why is this?
In addition, the red dots in the red box appear in the purified image generated in the right picture. Is this normal?