Secure beamforming via machine learning-driven joint optimization in IRS-assisted wireless systems

This paper proposes a Machine Learning-enhanced Gradient-Method Based Linear Optimization (ML-GMBLO) framework for secure beamforming in intelligent reflecting surface (IRS)-assisted 6G wireless networks. The method jointly optimizes beamforming vectors, artificial noise, IRS phase shifts, and user associations through an iterative process. A hybrid offline–online architecture is employed, where a deep neural network (DNN) trained on representative channel features provides near-optimal initializations. This reduces convergence iterations by approximately 40% compared to Alternating Optimization with Semi-Definite Programming (AO-SDP) under identical 4-bit phase quantization. Under the same channel estimation errors and Doppler conditions, ML-GMBLO achieves around 30% higher secrecy rate (5.2 vs. 4.0 bps/Hz), 35% lower latency, and near-linear scalability with respect to IRS size. This integrated approach enhances physical layer security while maintaining compatibility with hardware constraints.