Pilot-driven deep learning based RIS-assisted beamforming for secrecy rate maximization

Secure beamforming in reconfigurable intelligent surface (RIS)-assisted multiuser downlink systems is challenging due to high computational complexity and complex channel state information (CSI) estimation. This work proposes a pilot-driven beamforming network (PilotBeamNet) that jointly designs base-station (BS) transmit beamforming and quantized RIS phases directly from uplink pilot received signals with legitimate-user location cues to capture geometry. The framework avoids explicit channel estimation and slow iterative algorithms. A convolutional module reads each pilot frame, a long short term memory (LSTM) block with lightweight temporal attention aggregates them, and two simple heads output the beamformers and the discrete RIS phases. The location cues are embedded and fused with the features extracted from the pilot frames by the convolutional, LSTM, and temporal attention modules. Training maximizes ergodic secrecy rate (ESR) through Monte Carlo sampling of unknown eavesdropper channels, enabling robustness without requiring eavesdropper CSI. Once trained, PilotBeamNet performs single-pass inference with latency determined only by network depth. Across all tested conditions, PilotBeamNet achieves 10%–30% ESR improvement depending on the signal-to-noise ratio (SNR), pilot length, and RIS size, while reducing inference latency by more than an order of magnitude compared to alternating optimization (AO) and outperforming multilayer perceptron (MLP) baselines. It also maintains consistent performance under phase quantization and delivers higher secrecy rates across all evaluated configurations. • PilotBeamNet enables end-to-end secure RIS beamforming from uplink pilots without explicit channel estimation. • CNN-LSTM-attention predicts BS beamforming and quantized RIS phases with low latency. • Outperforms AO/MLP across SNRs and RIS sizes and remains robust to quantization.