search

qilong-zhang / Pytorch_Universal-adversarial-perturbation

A PyTorch implementation of universal adversarial perturbation (UAP) which is more easy to understand and implement.
54 stars 5 forks source link

Does it a deepfool or universal adversarial perturbations? #1

Closed DAJINZI01 closed 3 years ago

DAJINZI01 commented 3 years ago

I completely look your code in universal_pert.py, and I find that there is nothing about with the UAP.

qilong-zhang commented 3 years ago

Hi @DAJINZI01, Universal Adversarial Perturbation (UAP) is based on the Deepfool algorithm, i.e., It iterative craft single perturbation on 10,000 images. So the resultant perturbation is a universal adversarial perturbation. You can check the detail in their paper.

DAJINZI01 commented 3 years ago

Based on what you said, I can user another algorithm, i.e., I-FGSM, MI-FGSM, generate "single" perturbation(that is accumulate the noise based each iteration to a total noise). Is it right?

qilong-zhang commented 3 years ago

The official code is based on Deepfool, but you can also use any other algorithm to craft your UAP.