Secrecy Outage Probability Fairness in Intelligent Reflecting Surface Assisted Uplink Channels—Alternating Optimization Versus Deep Learning

This paper explores the fairness issues on physical layer security (PLS) in intelligent reflecting surface (IRS) assisted multiple-user uplink systems. Due to unknown instantaneous eavesdropper channel state information (CSI), it is not possible to acquire exact secrecy rate of a PLS system. In this paper, we introduce secrecy outage probability (SOP) as a security metric, instead of secrecy rate as used in most existing works, and formulate a minimization problem of maximum (min-max) SOP among multiple users. To solve this problem, we propose two independent approaches: one is alternating optimization (AO) and the other is a deep learning based (DL) scheme. The AO scheme decouples the problem into two sub-problems to alternately optimize phase shift matrix and receiver beamforming vectors, which gives a near-optimal performance but with a high complexity. The DL scheme, on the other hand, works based on neural networks through offline training, which is used for online generation of phase shift matrix and receiver beamforming vectors with a lower complexity. As traditional self-supervised neural networks cannot achieve a good solution to the max-min problem, we design a multiple-stage booster (MSB) framework to solve this problem. Simulations demonstrated that SOP is improved significantly with the proposed schemes compared to benchmark schemes. In particular, the AO scheme outperforms the DL-based approach slightly at the cost of a relatively high computational complexity.