larger kernel_size, smaller distance

[ybsu]

Hi Yingpeng, it is my honor to read your paper:Translation-Invariant attacks. I have tried your idea in my own task: face recognition, to generate some adversarial examples in a black-box manner, but I find the mean value of distance between the original pictures and the corresponding perturbed picture become smaller with the bigger kernel_size which is contrary to your result in the paper, I have not solve this problem yet. Would you mind giving me some suggestions to analysis this phenomenon? Thanks.
Best regards,
Looking forward to you reply.

[dongyp13]

Hi, Sorry for the late reply. What is the threat model used in your experiments? If you choose the L-infty norm threat model, after you apply the sign gradient, the distance should be similar to that without translation-invariance. Or you may try using smaller kernel size. Best, Yinpeng

…

在 2020年12月15日，下午7:53，Suqingyong ***@***.***> 写道： Hi Yingpeng, it is my honor to read your paper:Translation-Invariant attacks. I have tried your idea in my own task: face recognition, to generate some adversarial examples in a black-box manner, but I find the mean value of distance between the original pictures and the corresponding perturbed picture become smaller with the bigger kernel_size which is contrary to your result in the paper, I have not solve this problem yet. Would you mind giving me some suggestions to analysis this phenomenon? Thanks. Best regards, Looking forward to you reply. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#13>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AC2ZQAM3XT4NI6OMRV4RQYTSU5E45ANCNFSM4U4HD2FA>.