• Python >= 3.11, <3.12
• CUDA >= 11.8
• GCC >= 4.5
• poetry >= 1.4.2

curl -sSL https://install.python-poetry.org | python3 -

poetry env use <python-version>

If you use pyenv, you have to run pyenv local <python-version> in advance.

poetry install

• imagenet
  1. cd data/imagenet
  2. Download ILSVRC2012_img_val.tar and ILSVRC2012_devkit_t12.tar.gz from ImageNet official site
  3. mkdir val && tar -xf ILSVRC2012_img_val.tar -C ./val
  4. tar -xzf ILSVRC2012_devkit_t12.tar.gz
  5. python build_dataset.py
  6. mv val val_original && mv ILSVRC2012_img_val_for_ImageFolder val

poetry run python -B run_evaluation.py -p ../configs/config_apgd.yaml -g 0 -o ../result --log_level 20 --cmd_param attacker_name:str:APGD 

poetry run python -B run_evaluation.py -p ../configs/config_acg.yaml -g 0 -o ../result --log_level 20 --cmd_param attacker_name:str:ACG 

poetry run python -B run_evaluation.py -p ../configs/config_acg_t.yaml -g 0 -o ../result --log_level 20 --cmd_param attacker_name:str:ACG 

poetry run python -B run_evaluation.py -p ../configs/config_acg_r.yaml -g 0 -o ../result --log_level 20 --cmd_param attacker_name:str:ReACG 

poetry run python -B run_evaluation.py -p ../configs/config_reacg.yaml -g 0 -o ../result --log_level 20 --cmd_param attacker_name:str:ReACG 

selected_models.yaml shows the model names used in the experiments. You can generate adversarial examples for <model_name> by specifying model_name:str:<model_name> after --cmd_param. Similarly, you can specify the dataset and batch size as dataset:str:<dataset>, batch_size:int:<batch_size>.

config/config_<attacker_name>.yaml specifies the hyperparameters of the attack <attacker_name>. You can set different parameters by editing this file.