This is the python implementation of our paper, "Diversified Adversarial Attacks based on Conjugate Gradient Method" , accepted to ICML2022. paper(arxiv)

• Install python libraries.

pip install -r requirements.txt

• Complie .cpp and .c codes.

cd src/utils/cluster_coef_c
python setup.py build_ext -i

• Set ImageNet dataset.
  The directory name is the same as auto-attack

After your download, ls outputs the follow.

storage/ILSVRC2012/ILSVRC2012_img_val_for_ImageFolder/val

• Downloads the robsust models from RobustBench.

cd src
python get_models.py

export CUBLAS_WORKSPACE_CONFIG=:4096:8

• Fix PYTHONHASHSEED to 0.

export PYTHONHASHSEED=0

• ImageNet

  1. cd ../storage/ILSVRC2012
  2. Download ILSVRC2012_img_val.tar and ILSVRC2012_devkit_t12.tar.gz from ImageNet official site

  $ ls
  ILSVRC2012_img_val.tar

  3. mkdir val && tar -xf ILSVRC2012_img_val.tar -C ./val

  4. tar -xzf ILSVRC2012_devkit_t12.tar.gz

  5. python build_dataset.py

Attack on CIFAR-10

python -B run_cifar10_attack.py -o ../debug -g 0 --log_level 20 --param ./params/robustbench/cifar10/autoconjugate.yaml ./params/robustbench/cifar10/di.yaml  --experiment -bs 10

Attack on ImageNet

python -B run_imagenet_attack.py -o ../debug -g 0 --log_level 20 --param ./params/robustbench/imagenet/autoconjugate.yaml ./params/robustbench/cifar10/di.yaml  --experiment -bs 10

Attack on CIFAR-100

python -B run_cifar100_attack.py  -o ../debug -g 0 --log_level 20 --param ./params/robustbench/cifar100/autoconjugate.yaml ./params/robustbench/cifar10/di.yaml  --experiment -bs 10

(find ../result/ -maxdepth 7 |grep -e AUTOP -e AUTOC | xargs -L1 -P1 python run_evaluator_from_csv.py -ns 1  -r && find ../result/ -maxdepth 7 |grep AUTOC | xargs -L1 -P1 python run_evaluator_from_csv.py -ns 5 -r ) > cifar10_cw_result.csv

and open cifar10_cw_result.csv

Ding2020MMA,AUTOConjugaterestart-1,WideResNet-28-4,\cite{Ding2019},53.40,Wed Jan 26 10:59:21 2022
Ding2020MMA,AUTOConjugate,WideResNet-28-4,\cite{Ding2019},55.77,Wed Jan 26 10:59:21 2022

• 1st column: model name listed in RobustBench.
• 2nd column: the algorithm name.
  AUTOConjugaterestart-1 mean ACG with one restart.
• 3rd column: the architecture of model
• 4th column: the citation of the adversarial training method
• 5th column: the attack success rates
• 6th column: the execution start time

• Docker
  https://www.docker.com/
• NVIDIA Container Toolkit
  https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/overview.html

Build Docker Command

docker build --rm -t autocg:latest .

Create docker instance

docker run -it --gpus all -v $PWD/src:/AutoCG/src autocg /bin/bash

Start created container instance

docker start -ai [ContainerID]

Connect started container

docker attach [ContainerID]

Detach from Container

[control-P] [control-Q]

@inproceedings{yamamura2022,
    title={Diversified Adversarial Attacks based on Conjugate Gradient Method}, 
    author={Keiichiro Yamamura and Haruki Sato and Nariaki Tateiwa and Nozomi Hata and Toru Mitsutake and Issa Oe and Hiroki Ishikura and Katsuki Fujisawa},
    booktitle={ICML},
    year={2022}
}