Vision-aided beamforming and eavesdropper detection in UAV-borne intelligent reflecting surface assisted wireless systems

Abstract

The increasing demand for high-capacity, secure, and intelligent wireless communication in 6G and beyond has motivated the integration of emerging technologies such as Aerial Intelligent Reflecting Surfaces (AIRS), machine learning, and computer vision. This thesis proposes a novel framework that leverages visual sensing and RF signal characteristics for beam selection and eavesdropper detection in UAV-borne IRS-assisted multi-user wireless networks. The proposed system employs visual sensing information, extracted from images captured by UAV-mounted cameras using YOLOv10, to identify the location and spatial features of legitimate users. These visual features, combined with a sequence of previous beam information, are used as input to a Gated Recurrent Unit (GRU) based deep learning model. The model predicts the top-K beam indices with high confidence, significantly reducing the beam search space and enhancing received signal-to-noise ratio (SNR). Evaluation on the DeepSense6G dataset confirms that the proposed model achieves over 99% top-5 beam prediction accuracy and minimal power loss compared to exhaustive search-based beam selection.