Securing Next-Generation Networks against Eavesdroppers: FL-Enabled DRL Approach

Anticipated advancements in 5G wireless networks and beyond would necessitate an increased emphasis on security measures to accommodate the projected rise in demand for connections and services. Therefore, this paper aims to investigate the physical layer security (PLS) to evaluate the privacy of authorized users in multi-cellular networks, which represent a fundamental architecture in next-generation networks. Each cell is assumed to include a base station (BS) that serves multiple users. This scenario also takes into account the presence of several eavesdroppers. Every BS functions as a reinforcement learning (RL) agent that must undergo training in order to optimize security. To enhance the safety and speed of training, a federated learning (FL) technique is utilized. In this approach, a central unit regularly receives the neural network (NN) weights from the agents, updates them, and then transfers the result back to the agents to update their model. We examine and compare two deep RL methodologies, specifically deep Q-network, and Reinforce deep policy gradient. The findings of our research demonstrate the influence of the number of eavesdroppers on security, as well as the impact of the number of cells and the aggregation frequency of neural network parameters.