Mixture of Experts-Based Semantic Communication for Secure Artificial Intelligence of Things

Deep learning-based semantic communication systems have become prevalent in efficiently transmitting information in Artificial Intelligence of Things (AIoT), e.g., Intelligent Transportation Systems (ITS). However, their deployment over wireless networks introduces significant security vulnerabilities, including semantic noise attacks, physical layer adversarial attacks, and eavesdropping attacks. These threats may occur concurrently in practical scenarios, while existing defense mechanisms are typically designed to counter a single type of attack and struggle to address multiple heterogeneous threats. To tackle this, this paper introduces a Mixture of Experts-based Semantic Communication (MESC) system, which adaptively selects suitable security experts based on the security requirements of AIoT nodes. Numerical results indicate that the MESC system is capable of integrating specialized defense solutions to effectively mitigate multiple attacks with little loss of communication accuracy. Furthermore, the scalability of MESC substantially reduces ongoing costs by allowing the addition of new security experts with minimal training expenses.