Revisiting Adversarial Training Under Long-Tailed Distributions

Deep neural networks are vulnerable to adversarial attacks, leading to erroneous outputs. Adversarial training has been recognized as one of the most effective methods to counter such attacks. However, existing adversarial training techniques have predominantly been evaluated on balanced datasets, whereas real-world data often exhibit a long-tailed distribution, casting doubt on the efficacy of these methods in practical scenarios. In this paper, we delve into the performance of adversarial training under long-tailed distributions. Through an analysis of the prior method "RoBal" (Wu et al., CVPR'21), we discover that utilizing Balanced Softmax Loss (BSL) alone can obtain comparable performance to the complete RoBal approach while significantly reducing the training overhead. Then, we reveal that adversarial training under long-tailed distributions also suffers from robust overfitting similar to uniform distributions. We explore utilizing data augmentation to mitigate this issue and unexpectedly discover that, unlike results obtained with balanced data, data augmentation not only effectively alleviates robust overfitting but also significantly improves robustness. We further identify that the improvement is attributed to the increased diversity of training data. Extensive experiments further corroborate that data augmentation alone can significantly improve robustness. Finally, building on these findings, we demonstrate that compared to RoBal, the combination of BSL and data augmentation leads to a +6.66% improvement in model robustness under AutoAttack on CIFAR-10-LT. Our code is available at: https://github.com/NISPLab/AT-BSL.