Taming the Long Tail: Rebalancing Adversarial Training via Adaptive Perturbation

2026-05-13

AI Summary

This work addresses the degradation of adversarial robustness and skewed optimization objectives in long-tailed learning, which arise from class imbalance during adversarial training. To tackle these challenges, the authors propose RobustLT, a novel framework that introduces an adaptive perturbation mechanism to dynamically modulate the strength of adversarial perturbations across different classes. This approach reveals the dual role of perturbations in shaping the training distribution and enables joint optimization of long-tailed recognition and adversarial robustness. By integrating a rebalancing strategy, RobustLT significantly enhances both overall model robustness and class-wise performance equilibrium, achieving state-of-the-art results on multiple long-tailed benchmark datasets.

Abstract

Deep neural networks are highly vulnerable to adversarial examples, i.e.,small perturbations that can significantly degrade model performance. While adversarial training has become the primary defense strategy, most studies focus on balanced datasets, overlooking the challenges posed by real-world long-tail data. Motivated by the fact that perturbations in adversarial examples inherently alter the training distribution, we theoretically investigate their impact. We first revisit adversarial training for long-tail data and identify two key limitations: (i) a skewed training objective caused by class imbalance, and (ii) unstable evolution of adversarial distributions. Furthermore, we show that perturbations can simultaneously address both adversarial vulnerability and class imbalance. Based on these insights, we propose RobustLT, a plug-and-play framework that adaptively adjusts perturbations during adversarial training. Extensive experiments demonstrate that RobustLT consistently enhances adversarial robustness and class-balance on long-tailed datasets. The code is available at \href{https://github.com/zhang-lilin/RobustLT}{https://github.com/zhang-lilin/RobustLT}.

Problem

• long-tail
• adversarial training
• class imbalance
• adversarial robustness
• perturbation

Innovation

• adversarial training
• long-tail learning
• adaptive perturbation
• class imbalance
• robustness