CSBA: Covert Semantic Backdoor Attack Against Intelligent Connected Vehicles

Semantic communication (SemCom) can reduce data traffic for intelligent connected vehicles (ICVs), given the limited wireless spectrum available. However, it is important to recognize that deep learning-based SemCom is vulnerable to backdoor attacks, which pose significant security risks to ICVs. Therefore, it is crucial to investigate these security risks before integrating SemCom into ICVs. To this end, this study introduces a novel backdoor attack known as Covert Semantic Backdoor Attack (CSBA), specifically designed for SemCom-enabled ICVs. Unlike existing backdoor attack techniques that rely on noticeable triggers, CSBA analyzes the self-contained semantics in transmitted images to determine if they contain the target semantic required for initiating a backdoor attack. Moreover, in the event of an attack by CSBA, the target semantics disappear from the recovered image while the rest of the image remains unchanged, ensuring that the attack remains invisible. Experimental results confirm the effectiveness and the stealthiness of the proposed CSBA schemes across various wireless channel conditions and attack ratios.

I. Introduction

The intelligent connected vehicles (ICVs) are poised to revolutionize transportation efficiency and safety by leveraging data from various terminals to optimize AI-enabled services such as autonomous driving and urban road monitoring. However, the increasing demand for mobile data traffic is putting pressure on the limited wireless spectrum [1]. To address this challenge, researchers are turning to semantic communication (SemCom) for solutions [2], [3]. As shown in Fig. 1, vehicles collect traffic condition information, such as images of the leading vehicle. By incorporating SemCom into ICVs, the semantic features from these images are extracted and transmitted wirelessly to nearby cars [3], [4]. Upon receiving these characteristics, neighboring nodes reconstruct the information or perform downstream tasks, thereby reducing the need for wireless resources [5], [6].