Privacy-Preserving Graph-Based Intrusion Detection System: A Federated Learning Approach with GNN and XAI
Maninti Venkateswarlu, Ch. Karthik, M. Shivamani, Anusha Sathyanarayanan Rao · 2025
Today's digital networks are defended by the use of Intrusion Detection Systems (IDS), but they continue to face problems like data privacy, scalability, and transparency. To solve these problems, this paper proposes a new framework for intrusion detection using both Federated Learning (FL) and Explanaible Artificial Intelligence (XAI). Unlike other traditional models where there is a need to collect sensitive data in one central location which raises privacy issues, our approach is different. It allows several devices (nodes) to work collaboratively in training a model while never sharing their data. This system is built from the Graph Neural Network (GNN) framework, the best at capturing complex patterns as well as relationships within the traffic data of different networks. With GNN, detection of both familiar and new types of cyber attacking is enhanced. In order to make the system more tranparent and trustworthy, we add tools for explanation such as SHAP and LIME that aid in forming model predictions. Preliminary tests with our framework on popular datasets like UNSW-NB15 and CICIDS 2017 demonstrated high accuracy with a reduction in false positives and overall, robust performance, even with unbalanced data. In summary, this model addresses both the concerns of privacy and providing adequate explanations as to why certain decisions have been made. This fulfills increasing mandates on trustworthy AI, especially within the domain of cybersecurity