Defending Backdoor Attacks in Task-Oriented Multi-User Semantic Communication: A Contrastive Training Approach

Multi-user semantic communication enables multiple users (such as mobile devices) to collaborate on intelligent tasks, reducing communication load while enhancing efficiency. However, it is vulnerable to backdoor attacks due to the absence of auditing mechanisms for mobile users. In such attacks, malicious attackers can inject carefully crafted samples to implant backdoor neurons into the target model, which can be triggered to mispredict the target class based on the attacker's intent. Most existing defense methods for backdoor attacks are inadaptable to semantic communication systems, post-analysis, and are unable to respond timely. To address these challenges, we propose Reject-BD, a contrastive training-based defense mechanism tailored for multi-user semantic communication systems. Firstly, Reject-BD can actively apply test gradients to local semantic encoders. Then, contrastive training is used to amplify the differences in the updated test gradients of different clients to predict the credence degree of each client. Finally, the global semantic decoder is dynamically updated according to each client's credence degree. Experimental results demonstrate that Reject-BD effectively reduces the success rate of backdoor attacks while maintaining relatively high classification accuracy of the semantic decoder compared to five baseline methods.