AttackRider

Code for our IJCAI-25 paper Accelerating Adversarial Training on Under-Utilized GPU

Image Datasets

Requirements

Python 3.10, torch==1.13, numpy==1.26, torchvision==0.14

Datasets

Download CIFAR-10, CIFAR-100 and TinyImageNet, and change the dataset_path in the code accordingly.

Running Examples

Training

cd image_datasets/
python train_attackrider_bullet.py --dataset CIFAR-10 --base-at Bullet_TRADES --e 2
python train_attackrider_bullet.py --dataset CIFAR-10 --base-at Bullet_PGDAT --e 2
python train_attackrider_dbac.py --dataset CIFAR-10 --e 6
python train_attackrider_baseat.py --dataset CIFAR-10 --base-at TRADES --e 6
python train_attackrider_baseat.py --dataset CIFAR-10 --base-at PGDAT --e 6

where --dataset can also be CIFAR-100 or TinyImageNet.

Evaluation with AutoAttack

python test_aa.py --model-path ./CIFAR-10_results/AR_Bullet_TRADES_beta10.0_e2/ep_best.pt --dataset CIFAR-10

Tabular Datasets

Requirements

Python 3.8, torch==1.13, numpy==1.19.2, torchvision==0.14

Setup PyTorch environment in the same way as https://github.com/yandex-research/rtdl-revisiting-models

Datasets

Download Jannis and CoverType from https://github.com/yandex-research/rtdl-revisiting-models and put them under ./tabular_datasets/data/

Running Examples

Training

cd tabular_datasets/
python train_attackrider_bullet.py --dataset jannis --epsilon 0.1 --e 2 
python train_attackrider_dbac.py --dataset jannis --epsilon 0.1 --e 6
python train_attackrider_baseat.py --dataset jannis --epsilon 0.1 --e 6
python train_attackrider_bullet.py --dataset covtype --epsilon 0.05 --e 2
python train_attackrider_dbac.py --dataset covtype --epsilon 0.05 --e 3
python train_attackrider_baseat.py --dataset covtype --epsilon 0.05 --e 3

Evaluation with PGD-100 Attack

python test_pgd100.py --checkpoint-path ./output/covtype/FT-Transformer_AT_e3/checkpoint.pt --dataset covtype --epsilon 0.05

Reference

Part of the code is based on the following repo:

• https://github.com/yaodongyu/TRADES
• https://github.com/qizhangli/ST-AT
• https://github.com/yandex-research/rtdl-revisiting-models
• https://openreview.net/forum?id=eAPrmf2g8f2