Enhancing Cross-Site Scripting (XSS) Attacks Detection through Modern Transformer Architecture Optimizations

Abstract

Cross-Site Scripting (XSS) remains a widespread and damaging threats to web applications, as highlighted by the OWASP Top 10. While various detection methods exist, they often struggle to keep pace with the sophistication of attack vectors and obfuscation techniques. This paper implements an approach for enhancing XSS detection by leveraging modern optimizations within the transformer architecture. Our methodology uses a custom transformer encoder model trained on an aggregated dataset of nearly 100,000 samples, including newly collected XSS payloads. To enhance performance and efficiency, we integrate two key architectural improvements: Rotary Positional Embeddings (RoPE) to achieve a superior contextual understanding of HTTP payloads, and Flash Attention to significantly accelerate training and inference speeds while reducing memory consumption. Experimental results show good performance of our model that achieves an accuracy of 99.38 with high recall and precision. An ablation study demonstrates that the integrated optimizations improve detection accuracy by 0.11 percentage points, while reducing training time by approximately 32% and peak GPU memory usage by approximately 30% relative to standard Transformer configurations. Comparative evaluation against a Random Forest baseline further reveals a clear contextual understanding advantage over traditional frequency-based approaches, justifying the architectural complexity. The proposed model effectively captures structural patterns in sophisticated payloads that typically evade classical methods. This work presents an efficient, and accurate solution in real-time XSS threat detection.