Saved in:
Bibliographic Details
Main Authors: Xian, Guanmeng, Yang, Ning, Yu, Philip S.
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2605.02183
Tags: Add Tag
No Tags, Be the first to tag this record!
Table of Contents:
  • Adversarial training is effective on balanced datasets, but its robustness degrades under longtailed class distributions, where tail classes suffer high robust error and unstable decision boundaries. We propose Manifold-Constrained Adversarial Training (MCAT), a unified framework that enforces the semantic validity of adversarial examples by penalizing deviations from class-conditional manifolds in feature space, while promoting balanced geometric separation across classes via an ETF-inspired regularization. We provide theoretical results that link geometric separation to lower bounds on adversarially robust margins, and show that manifold-constrained adversarial risk upperbounds robust risk on high-density semantic regions. Extensive experiments on standard longtailed benchmarks demonstrate consistent improvements in overall, balanced, and tail-class adversarial robustness.