3SAT: A Simple Self-Supervised Adversarial Training Framework

Abstract

The combination of self-supervised learning and adversarial training (AT) can significantly improve the adversarial robustness of self-supervised models. However, the robustness of self-supervised adversarial training (self-AT) still lags behind that of state-of-the-art (SOTA) supervised AT (sup-AT), even though the performance of current self-supervised learning models has already matched or even surpassed that of SOTA supervised learning models. This issue raises concerns about the secure application of self-supervised learning models. The inclusion of adversarial training turns self-AT into a challenging joint optimization problem, and recent studies have shown that the data augmentation methods necessary for constructing positive pairs in self-supervised learning negatively impact the robustness improvement in self-AT. Inspired by this, we propose 3SAT, a simple self-supervised adversarial training framework. 3SAT conducts adversarial training on original, unaugmented samples, reducing the difficulty of optimizing the adversarial training subproblem and fundamentally eliminating the negative impact of data augmentation on robustness improvement. Additionally, 3SAT introduces a dynamic training objective scheduling strategy to address the issue of model training collapse during the joint optimization process when using original samples directly. 3SAT is not only structurally simple and computationally efficient, reducing self-AT training time by half, but it also improves the SOTA self-AT robustness accuracy by 16.19\% and standard accuracy by 11.41\% under Auto-Attack on the CIFAR-10 dataset. Even more impressively, 3SAT surpasses the SOTA sup-AT method in robust accuracy by a significant margin of 11.25\%. This marks the first time that self-AT has outperformed SOTA sup-AT in robustness, indicating that self-AT is a superior method for improving model robustness.