Framartin / lgv-geometric-transferability

Source of the ECCV22 paper "LGV: Boosting Adversarial Example Transferability from Large Geometric Vicinity"

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

cannot find attack_csgld_pgd_torch.py

zhengmeixi opened this issue · comments

Thanks for your amazing work and detailed repo. But I cannot find the attack file attack_csgld_pgd_torch.py. I'm not sure whether I missed something. Could you please tell me where I can find it? Thanks!

Hello @zhengmeixi ,
Thank you for your interest in my work and for spotting this. I forgot to add this file! This should be fixed with eb024bd
If you want to simply want to use LGV, without all the experiments of the paper, I recommend to use the refactored implementation merged in the torchattacks library: https://adversarial-attacks-pytorch.readthedocs.io/en/latest/wrappers.html#module-torchattacks.wrappers.lgv
Anyway, feel free to reach out if you have any doubt on this code base.
Best.

Hi,
Thanks for your answer :) I do hava another question. In func ExtendedProjectedGradientDescentPyTorch in utils/attacks.py, self._compute_torch is used to generate adversarial images, but it seems that this func is not degfined.

The _compute_torch() method of ExtendedProjectedGradientDescentPyTorch should be inherited from its parent class ProjectedGradientDescentPyTorch of the ART library. Did you install adversarial-robustness-toolbox? I used it in its 1.6.0 version. A later version of ART may have refactored ProjectedGradientDescentPyTorch._compute_torch(). Please let me know if it does not work.