Resurrecting the Tail: AGenerative Curriculum Framework for Self-Supervised Adversarial Robustness on Long-Tailed Distributions

Adversarial Training(AT) on long-tailed distributions triggers catastrophic vulnerability in tail classes. Prior work has predominantly relied on passive loss reweighting to compensate for data imbalance, leaving the underlying scarcity of tail-class samples unaddressed at the data level. In this work, we propose an active manifold-completion strategy that integrates conditional diffusion generation with self-supervised adversarial contrastive learning. Our frame-work introduces three core innovations. First, we exploit diffusion-based generative priors to repair the fragmented feature manifolds of minority classes by synthesizing on-manifold samples conditioned on tail-class labels. Second, we propose LT SSL AT, a dual-path self-supervised adversarial framework that decouples semantic invariance learning (clean-augmented views) from robustness learning (clean vs. adversarial pairs). Third, we introduce a curriculum scheduling weight s(t) that progressively shifts optimization from an Self-Supervised Learning(SSL)-only warm-up phase to joint adversarial alignment, preventing representation collapse. On CIFAR-10-LT (IR=10), CD-SSL achieves 35.12% Balanced Robustness under AutoAttack, surpassing the state-of-the-art TAET baseline by 4.58 percentage points. Extensive experiments across imbalance ratios 10, 20, 50, and 100 on three benchmarks (CIFAR-10-LT, CIFAR-100-LT, and MedMNIST) confirm consistent gains over both long-tailed and adversarial baselines.