A Survey of Harnessing Self-Supervision Against Adversarial Attacks

Nowadays, there has been a strikingly remarkable success in computer vision, owing to the boom of deep learning. Nevertheless, the revelation of adversarial examples has highlighted the reality that even highly potent deep learning models possess fatal vulnerabilities. On the other hand, the abundance of unlabeled data has fueled the surge in self-supervised learning, which has emerged as a prominent avenue for leveraging unlabeled data to enhance model performance. To address adversarial examples, this paper reviews the research results related to adversarial examples from three main aspects, namely adversarial attack, adversarial defense, and the fusion of adversarial defense and self-supervised learning framework, and analyzes the challenges that may be faced by the current research on utilizing self-supervised learning to fortify the adversarial robustness of deep learning model and discusses the prospect of remarkable methodologies on counteracting adversarial attacks and acquiring informative representation from self-supervised learning in the future.