Soft Actor-Critic Based Anti-Attack XSS Detection

Fan Zhang, Haoran Huang, Hongwu Lv, Meijuan Jia, Juntao Liu

Abstract: Cross-site scripting (XSS) attacks have become increasingly prevalent and sophisticated with the rapid development of Internet technology, posing a severe threat to network security. Deep learning-based XSS detection models autonomously learn complex features and pat-terns, boosting accuracy and adaptability. However, they often fail to detect novel and sophisticated adversarial tactics, leading to significant security breaches. To coun-teract these limitations, this study propose a novel Soft Actor-Critic (SAC) based anti-attack XSS detection. Our method enhances decision-making robustness by leveraging maximum entropy principles, thereby increasing the randomness and unpredictability of strategy selection. Additionally, we propose a unique Agent-Environment interaction framework specifically designed for XSS de-tection, which significantly improves the model's learning efficiency and adaptability. Comprehensive experimental validations demonstrate that our approach significantly outperforms existing methods in detecting a variety of XSS attack scenarios, particularly against unknown and complex attack patterns, thereby exhibiting enhanced adaptability and robustness.