Privacy-Preserving Federated Deep Learning for 6G Network Security Monitoring

Abstract

Deep federated learning (DFL) has emerged as an effective paradigm for privacy-preserving decentralized intelligence in sixth-generation mobile networks. Increasing deployment of intelligent network services introduces challenges associated with high communication overhead, susceptibility to model poisoning attacks, limited interpretability, and inefficient hyperparameter optimization, which reduce the reliability of federated intrusion detection. To address these limitations, an integrated framework combining Explainable Fast Topic-aware Temporal Dual-path Convolutional Attention (EFTTDCA) and the Artificial Circulation System Algorithm (ACSA) is presented. The dual-path attention architecture captures both short-term malicious traffic patterns and long-term temporal dependencies to improve intrusion detection in virtual network functions, whereas ACSA performs adaptive hyperparameter optimization for stable convergence and reduced computational complexity. A lightweight federated communication mechanism with secure model-update transmission and an explainability module further enhances privacy, transparency, and deployment efficiency. Experimental evaluation using the InSDN, CICIDS2017, Kitsune, 6G mIoT Network Traffic, and 6G-AI-EdSecure datasets achieves an average detection accuracy of 98.98%, precision of 98.92%, recall of 98.94%, and F1-score of 98.95%, together with lower communication overhead, faster convergence, and improved resource efficiency compared with existing federated intrusion detection approaches. These findings demonstrate that the proposed framework provides accurate, secure, interpretable, and computationally efficient intrusion detection for privacy-preserving 6G virtual network environments.