kira0096 / cgattack

Official repository for CVPR 2022 paper 'Boosting Black-Box Attack with Partially Transferred Conditional Adversarial Distribution'

Official repository for CVPR 2022 paper Boosting Black-Box Attack with Partially Transferred Conditional Adversarial Distribution.

This project is developed based on Python 3.6.

Install prerequisites

pip install -r requirements.txt

Download pre-trained model

Download the pretrained models and dataset [download link] and unzip it with

unzip pretrained.zip

Then you can conduct the untargeted attack for CIFAR-10 evaluation without training.

Robustness evaluation

• Evaluate our CG-ES against TARGET_MODEL [resnet.sh|densenet.sh|vgg.sh|pyramidnet.sh] by running

sh scripts/cifar_unt/TARGET_MODEL

Citation

Please cite our paper in your publications if it helps your research:

@inproceedings{Feng_CGATTACK_2022,
  title={Boosting Black-Box Attack with Partially Transferred Conditional Adversarial Distribution},
  author={Feng, Yan and Wu, Baoyuan and Fan, Yanbo and Liu, Li and Li, Zhifeng and Xia, Shutao},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2022}
}

cgattack's Issues

My experimental result for attacking resnet with target class 0 in Cifar is worse than Table 1.

I test the target attack with target class 0 against resNet by using their pretrained model on their provided testset, and find that the avg/median. query i 1238/751, which are larger than 696/421 in Table 1 of the paper.

My 1238/751 result is obtained by setting the targetn as resnet and loading the CGlow with your provided resnet_model.pth.tar on your provided 1000 images of test_cifar.npy. The attack parameters is the default value of your resnet.sh except for setting the target class as 0 as the same of Table 1. Of course, I filter images with the ground truth class 0.

Is there any parameter setting inconsistent with yours? Thank you very much for your response.

Could you please share us with the code about ImageNet and the dimension reduction by DCT?

requirements.txt havet package version

There is no specific version in the requirements.txt,How to determine the version of a dependent package.

About the train_cifar.npy and test_cifar.npy

Supplementary Material Results

Hello! Excuse me, where are the results of the supplementary material mentioned in the paper?