FedCPS: A Federated Deep Learning Framework for Privacy-Preserving Security in Cyber-Physical Systems

Abstract

: The increasing integration of computation, communication, and control within Cyber-Physical Systems (CPS) has enabled intelligent automation across critical infrastructures such as smart grids, industrial plants, and healthcare systems. However, the massive interconnectivity and continuous data exchange among CPS components expose these systems to severe cybersecurity threats, including data tampering, false data injection, and distributed denial-of-service (DDoS) attacks. Traditional deep learning (DL)-based intrusion detection models, while powerful, rely on centralized data aggregation, leading to privacy risks, scalability limitations, and communication overhead. To overcome these challenges, this paper proposes FedCPS, a novel Federated Deep Learning framework for privacy-preserving and scalable CPS security. In FedCPS, each CPS node trains a local deep learning model on its private data and shares only encrypted model updates with a global aggregator, which performs federated averaging to generate a unified global model. The framework integrates differential privacy, secure aggregation, and Byzantine-resilient mechanisms to protect data confidentiality and ensure model robustness against malicious updates. Experimental evaluations using SWaT, WADI, and NSL-KDD datasets demonstrate that FedCPS achieves 96.8% detection accuracy, 30% lower latency, and 50% reduced communication cost compared to centralized DL models. These results confirm that FedCPS effectively balances privacy, efficiency, and scalability, establishing it as a practical solution for next-generation secure and intelligent CPS environments.