Towards Transferable Adversarial Attacks on Vision Transformers

Recover the environment by

conda env create -f environment_transformer.yml

The used datasets are sampled from ImageNet. Unzip clean_resized_images.zip to ROOT_PATH of utils.py.

ViTs models from timm:

• vit_base_patch16_224
• deit_base_distilled_patch16_224
• levit_256
• pit_b_224
• cait_s24_224
• convit_base
• tnt_s_patch16_224
• visformer_small

CNNs and robustly trained CNNs from TI and here.

Change ROOT_PATH of utils.py.

python our_attack.py --attack OurAlgorithm --gpu 0 --batch_size 1 --model_name vit_large_patch16_224 --filename_prefix yours 

• attack: the attack method, OurAlgorithm, OurAlgorithm_MI or OurAlgorithm_SGM
• model_name: white-box name, vit_large_patch16_224, levit_256, cait_s24_224, tnt_s_patch16_224
• filename_prefix: additional names for the output file

sh run_evaluate.sh gpu model_{model_name}=method_{attack}-{filename_prefix}

If you use our method for attacks in your research, please consider citing

@article{Wei2021TowardsTA,
 title={Towards Transferable Adversarial Attacks on Vision Transformers},
 author={Zhipeng Wei and Jingjing Chen and Micah Goldblum and Zuxuan Wu and Tom Goldstein and Yu-Gang Jiang},
 journal={ArXiv},
 year={2021},
 volume={abs/2109.04176}
}