Adversarial Training Framework for Multi-Hop Semantic Transmission Against Malicious Attacks

With the rapid integration and development of critical technologies， such as wireless communications and artificial intelligence， various intelligent unmanned devices have been widely deployed in wireless ad-hoc network scenarios for exploration and data collection tasks. This proliferation has generated substantial volumes of image data， spawning a significant demand for massive data transmission. In dynamic wireless environments， traditional coding and transmission schemes that focus on accurate symbol delivery are gradually encountering bottlenecks in information compression， thus proving inadequate to address the core challenge of massive data transmission under constrained bandwidth. As a new paradigm integrating intelligence and communication， semantic communication delves into the novel dimension of information semantics itself. By adopting the joint source-channel coding （JSCC） approach， it significantly enhances transmission efficiency and spectrum utilization through semantic-level extraction， compression， and reconstruction of data， thereby providing innovative ideas and methodologies for future data communication. However， owing to the node mobility and multi-hop relay characteristics of wireless ad-hoc networks， relay nodes are vulnerable to malicious attacks， such as jamming， tampering， and eavesdropping. These attacks can severely compromise the quality of image data transmission. Therefore， designing an efficient and robust multi-hop transmission mechanism for image transmission is crucial. Existing research on image semantic communication primarily focuses on communication-friendly environments， with little consideration given to the impact of malicious attacks on the image semantic transmission process in wireless ad-hoc networks. Aiming at the poor anti-attack capability of image semantic transmission in wireless ad-hoc network scenarios， this study conducts research on image semantic multi-hop transmission algorithms and proposes a transformer sliding window-based semantic multi-hop transmission framework. By utilizing self-attention mechanisms to achieve efficient extraction of multi-scale semantic features and combining joint source-channel coding methods， it effectively enhances end-to-end transmission efficiency. Furthermore， to address the issue of malicious node attacks in wireless ad-hoc networks， this study proposes an adversarial training algorithm designed for semantic multi-hop transmission models. The method formulates a loss function for multi-hop image transmission and enhances the anti-attack capability of the proposed algorithm by dynamically adjusting and optimizing the semantic encoding and decoding strategies at each node within the wireless multi-hop transmission system. This approach effectively improves the algorithm’s resistance to attacks， while reducing semantic distortion in the received images. Moreover， considering the issue of semantic distortion accumulation caused by noise interference in the semantic multi-hop transmission system， a transformer-based denoising module is designed to mitigate semantic distortion resulting from channel noise interference and malicious attacks during the multi-hop transmission process， thereby further improving the robustness of the semantic multi-hop transmission. Simulation results indicate that in the additive white Gaussian noise （AWGN） channel， the proposed semantic multi-hop transmission strategy outperforms baseline methods across metrics， such as peak signal-to-noise ratio （PSNR） and multi-scale structural similarity index measure （MS-SSIM）. In both AWGN and Rayleigh channels containing malicious attacks， the proposed semantic multi-hop transmission strategy demonstrates significantly superior image reconstruction quality compared with existing schemes. Experimental results confirm that introducing the image denoising module and adversarial training effectively enhances the robustness of the image semantic multi-hop transmission system.