Robust Adversarial Training for Industrial Defect Classification with Long-Tailed Data

Abstract

Deep neural networks are vulnerable to adversarial examples which fool model predictions by adding imperceptible perturbations to natural examples. Adversarial training is effective in defending against adversarial attacks but faces a challenge with long-tailed data, where the over-compression of tail feature space undermines the reliability of defect classification models. To address this problem, we propose APRCB-AT, a novel adversarial training framework that integrates an adaptive perturbation radius and class-balanced loss. We find a positive correlation between the aggressiveness of adversarial examples and the perturbation radius within a certain range. Based on this, APRCB-AT assigns a larger perturbation radius in a certain range to the tail classes, while incorporating class-balanced loss with regularization to penalize the head classes. The experimental results show that APRCB-AT achieves 70.28% robust accuracy against adversarial attacks, surpassing existing methods, such as LSRG-DRW(67.22%) and REAT(68.06%).