inspire-group / DetectorGuard

Code for "DetectorGuard: Provably Securing Object Detectors against Localized Patch Hiding Attacks"
MIT License
14 stars 0 forks source link

求助 #4

Closed XinDong10 closed 1 year ago

XinDong10 commented 1 year ago

大佬,你的对抗样本(文中提到的500张都是怎么生成的呀),是使用的哪个攻击算法呀,是自己手动生成的嘛?

xiangchong1 commented 1 year ago

没有用具体的攻击算法,robustness是用certification procedure (algorithm 2)来完成的,得到的结果是对于同一个threat model下的任何攻击算法都适用的。可以再仔细去看一下Section 5.2对certified recall的定义 另外还可以看一下这两个notes https://github.com/inspire-group/PatchCleanser/blob/main/misc/notes_on_robustness_evaluation.md#are-you-looking-for-attack-code-in-this-repository https://github.com/inspire-group/adv-patch-paper-list#empirically-robust-defenses-vs-provablycertifiably-robust-defenses

XinDong10 commented 1 year ago

作者大佬,您好,我看了评估鲁棒性的算法以及您的另一篇objectseeker的文章,在您的这篇文章中,算法二有一个输入x,请问在实际测试certified recall时,您的这个x都是干净图像嘛。既然不需要具体的攻击算法,那您的这个x输入是干净图像嘛,因为没有具体的攻击算法,那么x也就没有办法是一个patch攻击后的对抗样本了。

------------------ 原始邮件 ------------------ 发件人: "inspire-group/DetectorGuard" @.>; 发送时间: 2023年9月13日(星期三) 凌晨2:16 @.>; @.**@.>; 主题: Re: [inspire-group/DetectorGuard] 求助 (Issue #4)

没有用具体的攻击算法,robustness是用certification procedure (algorithm 2)来完成的,得到的结果是对于同一个threat model下的任何攻击算法都适用的。可以再仔细去看一下Section 5.2对certified recall的定义 另外还可以看一下这两个notes https://github.com/inspire-group/PatchCleanser/blob/main/misc/notes_on_robustness_evaluation.md#are-you-looking-for-attack-code-in-this-repository https://github.com/inspire-group/adv-patch-paper-list#empirically-robust-defenses-vs-provablycertifiably-robust-defenses

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

xiangchong1 commented 1 year ago

对的,certification procedure的input一定是一张clean image。 certification结果的意义是,在同一个threat model下的attacker在这张input image上无论怎么加patch ,我们defense是不是能一直都做出正确的prediction。如果input image是一张clean image,这个certification结果很好理解。但是如果input是一张有patch的图片,得到的certification结果的意义是,无论怎么在这张已经有一个patch的图片上加另一个任意的patch,defense能不能做出正确预测,这个结果的意义就不明确了。

XinDong10 commented 1 year ago

嗯嗯 感谢大佬作者的耐心回复 我看了下object seeker的代码也是先求出脆弱的区域 然后再和patch会存在的区域求逻辑与的结果 然后再统计为1的个数 也就是看看这种位置的patch有多大可能落在脆弱的区域。这个脆弱的区域是用防御算法在干净图像上预测的bm是否满足下界求出的,也就是算法2。

------------------ 原始邮件 ------------------ 发件人: Chong Xiang @.> 发送时间: 2023年9月13日 19:26 收件人: inspire-group/DetectorGuard @.> 抄送: XinDong10 @.>, Author @.> 主题: Re: [inspire-group/DetectorGuard] 求助 (Issue #4)

对的,certification procedure的input一定是一张clean image。 certification结果的意义是,在同一个threat model下的attacker在这张input image上无论怎么加patch ,我们defense是不是能一直都做出正确的prediction。如果input image是一张clean image,这个certification结果很好理解。但是如果input是一张有patch的图片,得到的certification结果的意义是,无论怎么在这张已经有一个patch的图片上加另一个任意的patch,defense能不能做出正确预测,这个结果的意义就不明确了。

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>