L2dist can‘t convergence

IBM / Autozoom-Attack

Codes for reproducing query-efficient black-box attacks in “AutoZOOM: Autoencoder-based Zeroth Order Optimization Method for Attacking Black-box Neural Networks” , published at AAAI 2019

https://arxiv.org/abs/1805.11770

Apache License 2.0

57 stars 22 forks source link

L2dist can‘t convergence #2

Closed wzp123456 closed 5 years ago

wzp123456 commented 5 years ago

Hi, I run your code on my dataset. After finding the adversarial example, the L2 distance is larger as the number of iterations increases. It looks like that the estimated gradient can't guide the direction of the drop in L2 distance. Do you have any advice? Thanks!

chunchentu commented 5 years ago

Hi,

It seems like AutoZOOM cannot find a good estimation for the gradient vector. I would suggest to increase the number of random vector ( specified by the option num_rand_vec) to improve the estimation of the gradient vector.

Bests, Chun-Chen

wzp123456 commented 5 years ago

Thanks for your reply! I will try again according to your suggestion.