Dear authors,
thank you for your open source work.
I am currently reproducing your impersonation attack experiments on the CelebA-HQ dataset. The PSR results for both clean and protected images using ir152, irse50, and facenet align perfectly with the results reported in your paper.
However, I encountered an anomaly with the mobile_face model. While the PSR for protected images reaches around 90.16% (which is great), the PSR for clean images abnormally surges to 46.46% (compared to 12.68% in the paper).
Could you please advise on what might be causing this discrepancy and how to resolve it? Are there any specific preprocessing steps or model weights I might have missed?
Thank you in advance for your time and help!
Dear authors,
thank you for your open source work.
I am currently reproducing your impersonation attack experiments on the CelebA-HQ dataset. The PSR results for both clean and protected images using ir152, irse50, and facenet align perfectly with the results reported in your paper.
However, I encountered an anomaly with the mobile_face model. While the PSR for protected images reaches around 90.16% (which is great), the PSR for clean images abnormally surges to 46.46% (compared to 12.68% in the paper).
Could you please advise on what might be causing this discrepancy and how to resolve it? Are there any specific preprocessing steps or model weights I might have missed?
Thank you in advance for your time and help!