FedDTKG: Federated Temporal Graph Learning with Adaptive Loss for Robust 5G Attack Detection under Extreme Class Imbalance

作者：S. Sheikhi, Lauri Lovén, Susanna Pirttikangas, Panos Kostakos · 发表于：International Conference on Modeling, Analysis and Simulation of Wireless and Mobile Systems · 年份：2025 · DOI：10.1109/mswim67937.2025.11308990 · 研究领域：Computer Science

The distributed architecture and massive connectivity of 5G networks create significant security vulnerabilities that are challenging to address with centralized monitoring due to data privacy restrictions. Furthermore, the traffic data in such environments is characterized by extreme class imbalance, where critical but rare attacks are vastly outnumbered by benign traffic (with observed ratios exceeding 1:200). This paper introduces FedDTKG, a novel federated learning framework designed to provide robust, privacy-preserving intrusion detection under these challenging conditions. The framework features two key innovations: (1) a Temporal-Aware Self-Adaptive Graph Attention Network (TASA-GAT) that explicitly models the temporal dynamics and relational structure of network flows and (2) an Adaptive Synthetic Focal Loss (ASFL) that counters class imbalance by dynamically tuning its focus and incorporating a feature-level variance regularization term to improve minority class representation. We conduct a comprehensive evaluation on a non-IID distribution of 5G traffic data. In a centralized setting, FedDTKG achieves a state-of-the-art F1-macro score of 0.8756, significantly outperforming traditional ML and standard GNN baselines that fail to detect minority classes. In the federated setting, FedDTKG maintains a high F1-macro of 0.7546, whereas conventional GNNs fail completely, demonstrating our model’s resilience to statistical heterogeneity. The findings validate FedDTKG as an ...