DCARL: Decoupled-Curriculum for Adversarial Robustness Learning under Long-Tailed Distributions

Deep neural networks are highly susceptible to adversarial attacks. Although adversarial training is an effective defense, its efficacy in long-tailed settings remains limited. Current methods are constrained in long-tailed scenarios by early representation bias, imbalanced optimization, and insufficient class-aware adaptivity. To address these challenges, we propose Decoupled-Curriculum Adversarial Robustness Learning, a two-stage framework. The Initial Representation and Balancing stage employs inter-class margin adjustment via LDAM with a Deferred Re-weighting schedule to build balanced, transferable representations. The Adaptive Correctness-Aware Robustness Learning stage adapts the per-class PGD perturbation budget using class-wise correctness signals and optimizes a balanced loss that combines mean, medium-class protection, and tail-class reinforcement components, aligning robust learning across head, medium, and tail classes. Experiments on standard long-tailed benchmarks validate the effectiveness of our approach, yielding consistent improvements in both natural and robust performance.