smartcameras / colorfool Goto Github PK

Hi, it is an interesting work. I test ColorFool on one GPU, attacking ResNet50 with 1000 images randomly chosen from ImageNet. I need nearly 20 hours to finish the attack (attacking 1000 images). Is this normal? Thanks!

Hi! Nice to read such interesting work!

I have tested the code on 1000 ImageNet images with the default three networks (AlexNet, ResNet15, and ResNet50), and got similar results as reported in your paper.
However, when I tried attacking an Inception-v3 model, the success rate dropped to only ~65%. I repeated with two image resolutions: 224x224 and 299x299, and max trials =1000 and 1500, the results remain almost the same.
Question: Have you ever tried other networks? Could you please help me figure out what may be happening here?

Thanks~!

There is only Colorfool one way to attack, It seems that there is no model attack after adding filter. Do you have more complete code?

Hi!
I am glad to read such a very interesting paper.
I recently experimented with your code on nips-1k and was confused about the results. (ResNet50)
The attack success rate of the output is 91.6%. But when I tested the saved adversarial samples, the success rate of the attack was only 87.2%. What is the reason for this? Is it reasonable?

Thanks!

I cannot find decoder.pth in Google.drive