A SAC-based DRL approach for secure RIS-assisted NOMA transmission against internal and external eavesdropping

Abstract

This paper investigates a reconfigurable intelligent surface (RIS)-assisted secure non-orthogonal multiple access (NOMA) network in the presence of both internal eavesdroppers (IEs) and external eavesdroppers (EEs). To mitigate threats of IEs, a reversed successive interference cancellation (SIC) decoding order is adopted. Maximizing the sum secrecy rate is framed as a non-convex optimization challenge, where the primary decision variables are the RIS phase shifts and the transmit precoding matrix. A soft actor-critic (SAC)-based deep reinforcement learning (DRL) scheme is developed to solve this problem. Numerical results verify that the proposed scheme surpasses deep deterministic policy gradient (DDPG) and exceeds conventional successive convex approximation (SCA) at low transmit power. It also converges faster and operates more stably than the DDPG-based approach.