AI-DRIVEN INTRUSION DETECTION SYSTEM FOR DIGITAL TWINS USING FEDERATED LEARNING

Abstract

Digital Twin (DT) technology provides close-to-real-time replicas of physical systems for monitoring and optimization but also widens the cyber-attack surface. Traditional intrusion detection systems (IDS) rely on centralized training and data aggregation, which raises privacy and scalability concerns. This paper proposes an AI-driven intrusion detection framework based on Federated Learning (FL) for heterogeneous Digital Twin environments. A hybrid CNN–LSTM model is trained locally on DT nodes, while a secure aggregation server aggregates encrypted model updates. Differential privacy and Byzantine-robust 8 aggregation improve privacy and resilience. The experimental results on CICIDS-2018 and ToN-IoT (simulated) show promising accuracy (96.4% and 94.8%), low communication overhead, and robustness against poisoning.