The paper studies the secrecy communication threatened by a single eavesdropper in Energy Harvesting (EH)-based cognitive radio networks, where both the Secure User (SU) and the jammer harvest, store, and utilize RF energy from the Primary Transmitter (PT). Our main goal is to optimize the time slots for energy harvesting and wireless communication for both the secure user as well as the jammer to maximize the long-term performance of secrecy communication. A multi-agent Deep Reinforcement Learning (DRL) method is proposed for solving the optimization of resource allocation and performance. Specifically, each sub-channel from the Secure Transmitter (ST) to the Secure Receiver (SR) link, along with the jammer to the eavesdropper link, is regarded as an agent, which is responsible for exploring optimal power allocation strategy while a time allocation network is established to obtain optimal EH time allocation strategy. Every agent dynamically interacts with the wireless communication environment. Simulation results demonstrate that the proposed DRL-based resource allocation method outperforms the existing schemes in terms of secrecy rate, convergence speed, and the average number of transition steps.
Keywords: cognitive radio network; deep reinforcement learning; energy harvesting; physical layer security.