Privacy-Preserving Federated Deep Learning for Real-Time XSS Attack Detection

Web applications continue to face increasing securityrisks due to the rapid evolution of Cross-Site Scripting (XSS) payloadsand the growing sophistication of attackers. Traditional centralizeddetection systems struggle to maintain high accuracy in real-timeenvironments and often require access to sensitive user data,creating additional privacy concerns. Addressing these challengesrequires a detection mechanism that can adapt to diverse datasources, withstand adversarial manipulation, and operate withoutexposing client-side information. The proposed work introduces aprivacy-preserving federated deep learning framework thatcombines LSTM–CNN hybrid feature extraction with secureaggregation to detect XSS attacks across distributed clients in realtime. The system enables collaborative model training withoutsharing raw data, thereby improving generalization whilemaintaining confidentiality. Robustness is further enhanced througha token-aware preprocessing pipeline and resilient gradientaggregation, allowing the model to handle highly obfuscated andpolymorphic XSS payloads. Experimental evaluation demonstratesthat the proposed architecture achieves superior performancecompared to existing centralized and federated detectionapproaches, yielding 96–98% accuracy, 95– 97% precision, and afalse-positive rate as low as 2–4%. The system also shows strongresistance to adversarial crafting, poisoning attempts, and inference-based threats. These results confirm the suitability of the proposedframework as a secure, scalable, and high-performance solution formodern web security environments.Index Terms—Federated learning, deep learning, XSS attackdetection, privacy-preserving security, intrusion detection, real-timeweb security.