Privacy-Preserving AI: Leveraging Federated Reinforcement Learning in Distributed Systems

Organizing computing distributed resources in conjunction with machine learning algorithms in the age of big data is critical. This paper presents the FRL approach that combines Federated Learning and Reinforcement Learning to address problems in scale, speed, and data privacy. The overall framework of the Federated Reinforcement Learning (FRL) model bestows the use of distributed computing at different agents to train a standard RL model. This framework guarantees data privacy utilizing differential privacy and secure multi-party computation (SMPC). Quantitative analysis proves the proposed FRL framework retrieves a better result than a standard RL strategy implemented in the central controller. In the 100th episode of the study, the authors’ FRL framework had better results than the centralized RL approach in autonomous navigation, resource management and game-playing. Thus, the FRL framework gained a total benefit of 95 cumulative. 6, 98. 4, and 98. 3, and the rewards obtained for the centralized RL approach were 90. 5, 92. 0, and 93. 8, respectively. Also, there was lower communication overhead about the generic FRL framework, where the assessment was recorded at 33. At the same time, 100 parameter exchanges were performed; the decentralized RL method was measured at 6 units, while the centralized RL method was measured at 41. 3 units. The joint plot analysis established that the traditional FRL framework training strategy offers quicker convergence and heightened model robustness, as the collaborative training approach provides. The heatmap visualization confirms the enhancement of the capability to control the communication overhead.