Semantic Viruses: How to Destroy a Task Oriented Semantic Communication Network

Semantic communication has gained significant attention for its potential to address the limitations of traditional communication systems. However, most current research focuses on architectural design, neglecting the critical issue of communication security. In this paper, we reveal the security vulnerabilities in existing semantic communication systems. First, we introduce a knowledge base-assisted task-oriented image semantic communication network. Next, we design two types of semantic viruses that target the knowledge base: a static adversarial semantic virus and a dynamic diffusion semantic virus. The adversarial semantic virus generates a fixed-strength attack, while the diffusion semantic virus attack follows a Gaussian distribution. Finally, for multi-task scenarios, we propose a unified multi-task diffusion semantic virus capable of switching between task-specific viruses using different task embeddings. Experiments on semantic segmentation and image recognition tasks demonstrate that the proposed method effectively degrades task performance at the receiver.