attack success rate seems different from paper

Question

attack success rate seems different from paper

yuxu915 opened this issue 3 years ago · 1 comments

hi, when I try to reproduce your results in Table 1, results seems different from that in paper.
For example, using AMT-GAN/test.py, I try to calculate ASR of clean image in CelebA-HQ, attacking ir152, the results are 0.02,0.019,0.017,0.049, while in paper it's 3.80. The corresponding test image IDs are: 008793, 047073, 055622, 166607.
I wonder if I miss something.
Thanks!

Answer 1 · 2022-08-09T08:14:45.000Z

Hi,
In the current repository of AMT-GAN, we have added an additional loss in the cycle reconstruction path (See backbone/solver.py Line312-313, Line393-394). This will make the AMT-GAN generate more stable and natural face images (FID degrades from 34 (in our official paper) to 27). As a trade-off, it will slightly affect the attack strength. If you want to have a better ASR or just want to reproduce the results, delete self.LPIPS(fake_A, image_s), self.LPIPS(fake_B, image_r) in Line312-313, self.LPIPS(purified_A, image_s), self.LPIPS(purified_B, image_r) in Line393-394.