Secure Resource Allocation for NOMA PLC Networks with Multi-agent DRL

This paper investigates the intelligent downlink secure transmission in a non-orthogonal multiple access (NOMA) network over power line channels, where each NOMA pair consists of one distant and nearby terminal user (UE) and the distant UE could wiretap the nearby UE. In order to maximize the security sum rate of all nearby UEs while guaranteeing the targeted data rate requirements of all UEs, a joint secure subchannel assignment and power allocation problem is first established. To deal with it, a deep reinforcement learning (DRL) scheme is then proposed. In specific, this scheme uses a deep Q-network (DQN) to learn the optimal decision policy for each NOMA pair by combining the compressed local observation and an aggregation information from other NOMA pairs as the input. All the input dimensions of the DQNs, the local compression networks and the central aggregation networks are independent of the PLC network size. In another word, this proposed scheme is scalable, which ca be easily applied in the practical system where the network size is dynamically changing. Simulation results verify that the effectiveness of this proposed scheme compared to benchmark scheme.