Implement Zeroth Order Optimization based Attack

yifangyin commented 3 years ago

https://arxiv.org/pdf/1708.03999.pdf

gchhablani commented 3 years ago

The original paper talks about a few datasets/models/optimizers/tricks, listed below. I am checking the ones I'm planning to implement based on paper results and ease of coding. Please let me know in case of any issues.

Datasets

[x] MNIST
[x] CIFAR10
[ ] ImageNet

Models

[x] Custom CNN (for MNIST/CIFAR10)
[ ] InceptionV3 (for ImageNet)

Optimizers

[x] ZOO-Adam
[ ] ZOO-Newton

Tricks

[ ] Importance Sampling
[ ] Hierarchical Attacks
[ ] Resetting Adam

EDIT:

I will be implementing it based on the PyTorch version here but will try to make it framework-agnostic. The models will be in PyTorch, however.
Original Implementation: https://github.com/huanzhang12/ZOO-Attack/blob/master/l2_attack_black.py

someshsingh22 commented 3 years ago

Looks great! @gchhablani can you also add the code references for crediting and future references

someshsingh22 commented 3 years ago

@gchhablani MNIST is already present Imagenet is in #35, Opening InceptionV3 as a first issue

Adversarial-Deep-Learning / code-soup