Weakly Supervised Contrastive Adversarial Training for Learning Robust Features from Semi-supervised Data

The existing adversarial training (AT) methods often suffer from incomplete perturbation, i.e., not all non-robust features are perturbed during the generation of AEs, which causes remaining correlations of non-robust features with labels captured by the target model, i.e., suboptimal learning of robust features. However, fulfilling complete perturbation, i.e., perturbing as many non-robust features as possible, is not easy due to the challenges of unidentifiability of robust/non-robust features and the sparsity of labeled data. To overcome these challenges, we propose a novel solution called Weakly Supervised Contrastive Adversarial Training (WSCAT). WSCAT fulfills complete perturbation for better learning of robust features by blocking the correlations between non-robust features and labels, via complete AE generation over partially labeled data grounded in information theory. The solid theoretical analysis and the extensive experiments conducted on widely adopted benchmarks verify the superiority of WSCAT.