When-Adversarial-Training-Meets-Vision-Transformers

Adv Training of ViTs @ NeurIPS 2022 "When Adversarial Training Meets Vision Transformers: Recipes from Training to Architecture" by Yichuan Mo, Dongxian Wu, Yifei Wang, Yiwen Guo, Yisen Wang.

Requirements

Run pip install -r requirement.txt to install all requrements!

CIFAR-10

Vanilla adversarial defense methods：

# AT
CUDA_VISIBLE_DEVICES=0 python train_cifar.py --model "deit_tiny_patch16_224" --out-dir "./pgd_vanilla" --method 'AT' --seed 0
# TRADES
CUDA_VISIBLE_DEVICES=1 python train_cifar.py --model "deit_tiny_patch16_224" --out-dir "./trades_vanilla" --method 'TRADES' --seed 0
# MART
CUDA_VISIBLE_DEVICES=2 python train_cifar.py --model "deit_tiny_patch16_224" --out-dir "./mart_vanilla" --method 'MART' --seed 0

You can use --model to select other ViT variants to train.

Example for AT after combining ARD and PRM：

CUDA_VISIBLE_DEVICES=0 python train_cifar.py --model "deit_tiny_patch16_224" --n_w 10 --out-dir "./pgd_architecture" --method 'AT' --seed 0 --ARD --PRM

You can use --method to select other defense methods!

Imagenette

First, you need to download the ImageNette-v1 dataset (the old version of ImageNette) to the local path ./data from here.

Vanilla adversarial defense methods：

# AT
CUDA_VISIBLE_DEVICES=0 python train_imagenette.py --model "deit_tiny_patch16_224" --out-dir "./pgd_vanilla" --method 'AT' --seed 0
# TRADES
CUDA_VISIBLE_DEVICES=1 python train_imagenette.py --model "deit_tiny_patch16_224" --out-dir "./trades_vanilla" --method 'TRADES' --seed 0
# MART
CUDA_VISIBLE_DEVICES=2 python train_imagenette.py --model "deit_tiny_patch16_224" --out-dir "./mart_vanilla" --method 'MART' --seed 0

You can use --model to select other ViT variants to train.

Example for AT after combining ARD and PRM：

CUDA_VISIBLE_DEVICES=0 python train_imagenette.py --model "deit_tiny_patch16_224" --n_w 10 --out-dir "./pgd_architecture" --method 'AT' --seed 0 --ARD --PRM

You can use --method to select other defense methods!

ImageNet-1k

Vanilla AT：

python train_imagenet.py --model "swin_base_patch4_window7_224_in22k" --out-dir "./pgd_vanilla" --seed 0

You can use --model to select other ViT variants to train.

Example for AT after combining ARD and PRM：

python train_imagenet.py --model "swin_base_patch4_window7_224_in22k" --n_w 2 --out-dir "./pgd_architecture" --seed 0 --ARD --PRM

Checkpoint

Using our open-source code, we re-ran the experiment on ImageNet-1k on A100 GPUs. You can download the checkpoint directly from the following link: ViT-B (checkpoint, AA), Swin-B (checkpoint, AA). The pretrained models are also available on Robustbench.

Acknowlegements

This repository is built upon the following four repositories:
https://github.com/yaodongyu/TRADES
https://github.com/YisenWang/MART
https://github.com/rwightman/pytorch-image-models
https://github.com/RulinShao/on-the-adversarial-robustness-of-visual-transformer.

Cite this work

If you find our code is useful, please cite our accompanying paper!

@inproceedings{moadversarial,
  title={When Adversarial Training Meets Vision Transformers: Recipes from Training to Architecture},
  author={Mo, Yichuan and Wu, Dongxian and Wang, Yifei and Guo, Yiwen and Wang, Yisen},
  booktitle={NeurIPS}
  year={2022}
}