Federated Learning-based Explainable Intrusion Detection System: A Hybrid Transformer–GNN Approach for Privacy-Preserving and Robust Cybersecurity
Main Article Content
Abstract
We propose a novel federated learning-based explainable intrusion detection system (IDS) that integrates a hybrid Transformer–graph neural network (GNN) architecture to address the challenges of privacy-preserving and robust cybersecurity. Modern IDSs face limitations in handling distributed network environments where data privacy is critical, while existing methods often lack interpretability and fail to capture both sequential and relational patterns in network traffic. The proposed framework combines the strengths of federated learning, Transformer attention mechanisms, and GNNs to enable collaborative training across multiple clients without sharing raw data, thereby preserving privacy. The Transformer module processes sequential network traffic data to extract contextual features, while the GNN analyzes topological relationships within the network graph, enabling a comprehensive understanding of both temporal and structural anomalies. These features are then fused to enhance detection accuracy. Furthermore, an explainability module based on feature attribution techniques provides transparency into the model’s decisions, aiding cybersecurity analysts in identifying root causes of threats. The system also incorporates an adaptive aggregation strategy to dynamically weigh client contributions during federated updates, improving robustness against heterogeneous data distributions. Experimental validation on benchmark datasets demonstrates superior performance in detecting intrusions while maintaining interpretability. The proposed method not only advances the state-of-the-art in IDS by unifying privacy, accuracy, and explainability but also offers practical insights for deploying scalable and trustworthy cybersecurity solutions in real-world distributed systems.