Requirements:

Python 3.7.16
torch 1.8.0+cu111
torchvision 0.9.0+cu111
tqdm 4.65.0
numpy 1.21.6
pillow 9.5.0

Experiments:

The code consists of three Python scripts. Before running the code, you need to complete the following two steps:

Download Data: Download the data from the provided link (https://pan.baidu.com/s/1NlenXev0cN1l55ZSVQ-_nw; password: d6tn) and place it in the benign_image/ directory.

Calculate Class Activation Maps (CAM): Compute the class activation maps and place them in the CAM/ directory.

Running the code

untaregt_attack_example.py：Non-targeted attack
taregt_attack_example.py：targeted attack
victim_one.py: test

Acknowledgments:

Code refer to
https://github.com/Harry24k/adversarial-attacks-pytorch
https://github.com/frgfm/torch-cam
https://github.com/erbloo/dr_cvpr20
https://github.com/RobustBench/robustbench