Alleviating the Effect of Data Imbalance on Adversarial Training [pdf]

Introduction: We explore the challenges and limitations of adversarial training on a long-tailed dataset. It is not to address the long-tail problem itself. Instead, we study how to improve adversarial training when the training data is imbalanced.

Requirements

1. pytorch >= 1.9.0
2. torchvision
3. numpy
4. tqdm
5. mmcv

Adversarial Training

python train.py --arch [resnet, wrn] --dataset [cifar10, cifar100] --imb [imbalanced ratio] --ext [existing ratio] --save [the name you want to save your model] --exp [experiment name]