Revisiting Adversarial Training under Long-Tailed Distributions

Code for CVPR 2024 "Revisiting Adversarial Training under Long-Tailed Distributions".

Environment

• Python (3.9.12)
• Pytorch (2.1.0)
• torchvision (0.16.0)
• CUDA
• AutoAttack
• advertorch

Content

• ./datasets: Generate long-tailed datasets.
• ./models: Models used for training.
• train_at_bsl_cifar10.py: AT-BSL on CIFAR-10-LT.
• train_at_bsl_cifar100.py: AT-BSL on CIFAR-100-LT.
• train_at_bsl_tiny_imagenet.py: AT-BSL on Tiny-ImageNet-LT.
• at_bsl_loss.py: Loss function for AT-BSL.
• pgd_attack.py: Use PGD to select the best epoch during training.
• eval.py: Evaluate the robustness under various attacks.

Run

• AT-BSL using ResNet18 on CIFAR-10-LT

CUDA_VISIBLE_DEVICES='0' python train_at_bsl_cifar10.py --arch res --aug none

• AT-BSL-RA using ResNet18 on CIFAR-10-LT

CUDA_VISIBLE_DEVICES='0' python train_at_bsl_cifar10.py --arch res --aug ra

• AT-BSL using WideResNet-34-10 on CIFAR-100-LT

CUDA_VISIBLE_DEVICES='0' python train_at_bsl_cifar100.py --arch wrn --aug none

• AT-BSL-AuA using WideResNet-34-10 on CIFAR-100-LT

CUDA_VISIBLE_DEVICES='0' python train_at_bsl_cifar100.py --arch wrn --aug aua

• AT-BSL-RA using PreActResNet-18 on Tiny-ImageNet-LT

CUDA_VISIBLE_DEVICES='0' python train_at_bsl_tiny_imagenet.py --aug ra

• Evaluation

CUDA_VISIBLE_DEVICES='0' python eval.py --model_path INSERT-YOUR-MODEL-PATH

Pre-trained Models

• The pre-trained models can be downloaded from the Google Drive.

Reference Code

[1] RoBal: https://github.com/wutong16/Adversarial_Long-Tail

[2] REAT: https://github.com/GuanlinLee/REAT

[3] AT: https://github.com/MadryLab/cifar10_challenge

[4] TRADES: https://github.com/yaodongyu/TRADES

Citation

@inproceedings{yue2024revisiting,
  title={Revisiting Adversarial Training under Long-Tailed Distributions},
  author={Yue, Xinli and Mou, Ningping and Wang, Qian and Zhao, Lingchen},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={24492--24501},
  year={2024}
}