CW attack doesn't work.maybe i don't know how to run correctly. help me

[xieshisheng]

No description provided.

[Harry24k]

Please let me know the error situation.

[xuxiangsun]

Sir, I also encounter this issue, I found when I use the parameters default in your code, the success rate of adversarial attack is close to 0. Cause I do not check the original paper about C&W, so I am not sure how to do. Can you help me ? @Harry24k

[Harry24k]

Sir, I also encounter this issue, I found when I use the parameters default in your code, the success rate of adversarial attack is close to 0. Cause I do not check the original paper about C&W, so I am not sure how to do. Can you help me ? @Harry24k

With default c, you can't easily get adversarial images. Set higher c like 1.

I also used c=1 in the demo.

[xuxiangsun]

Ok,  I will try it now 1137582789 Email:1137582789@qq.com Signature is customized by Netease Mail Master On 07/30/2020 22:04, Harry Kim wrote: Sir, I also encounter this issue, I found when I use the parameters default in your code, the success rate of adversarial attack is close to 0. Cause I do not check the original paper about C&W, so I am not sure how to do. Can you help me ? @Harry24k With default c, you can't easily get adversarial images. Set higher c like 1. — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

[xieshisheng]

set c like 1, I tried, but the code doesn't work too.I also check the souce code, i found the source code are not wrong.So i dont know where is wrong.

[xieshisheng]

I found, set c to 10 or up(1 is not enough), CW attack can changed image to adversarial image with high success rate

[xuxiangsun]

Let me try, thanks for your precious experience! @xieshisheng

[xuxiangsun]

By the way, sir, how do you set iters and lr? @xieshisheng

[xieshisheng]

I dont change iters and lr.
source code
"loss1 = nn.MSELoss(reduction='sum')(a, images)
loss2 = torch.sum(self.c*f(a))"
i think MSEELoss force new image as same as old image, if c is small ,gradient of loss2 is small . So the "adversarial" image is just as same as image, but not truly adversarial image.

[xuxiangsun]

I cannot agree with you any more. But I think the iteration maybe to large...... 1137582789 Email:1137582789@qq.com Signature is customized by Netease Mail Master On 07/31/2020 10:05, xieshisheng wrote: I dont change iters and lr. source code "loss1 = nn.MSELoss(reduction='sum')(a, images) loss2 = torch.sum(self.c*f(a))" i think MSEELoss force new image as same as old image, if c is small ,gradient of loss2 is small . So the "adversarial" image is just as same as image, but not truly adversarial image. — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

[xieshisheng]

hhhhh, maybe large iteration is more effective.I‘ll try it later.thank you

[xuxiangsun]

thanks too! 1137582789 Email:1137582789@qq.com Signature is customized by Netease Mail Master On 07/31/2020 10:31, xieshisheng wrote: hhhhh, maybe large iteration is more effective.I‘ll try it later.thank you — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.