Privacy-Preserving Deep Reinforcement Learning for Secure Resource Orchestration in Cyber-Physical Systems

Abstract

This research addresses the critical challenge of secure and efficient resource allocation in Cyber-Physical Systems (CPS) by introducing a Deep Reinforcement Learning (DRL) framework integrated with privacy-preserving federated learning. Unlike traditional methods, our approach ensures that raw data remains localized, thereby mitigating privacy risks and enhancing trust within the CPS ecosystem. A custom-designed reward function is proposed to optimize both resource utilization and privacy assurance, balancing performance and security goals. To strengthen data confidentiality, we incorporate a variant of Differential Privacy, which increases the privacy budget without significantly compromising data utility—achieving a privacy guarantee of 0.8 while maintaining over 92% model accuracy. Experimental validation on a smart grid test bed demonstrates the efficacy of the proposed model, achieving a 17.6% improvement in resource allocation efficiency, a 23% reduction in communication overhead, and a 12% increase in system throughput compared to baseline DRL models without privacy constraints. Overall, the framework demonstrates state-of-the-art performance in optimizing resources in complex, distributed CPS environments while upholding stringent privacy requirements. The proposed method offers a scalable and secure solution for next-generation CPS applications in smart infrastructure.