Secure Semantic Communications With Information-Theoretic Honeypoints

Abstract

The broadcast nature of wireless channels renders semantic communications vulnerable to eavesdropping. Existing defense mechanisms generally suffer from severe performance degradation for legitimate users. To address this limitation, this letter proposes a secure semantic communication framework that exploits a semantic honeypoint mechanism. By embedding a honeypoint into the model during training, the proposed framework deliberately perturbs the reconstruction process of the eavesdropper. Moreover, from an information-theoretic perspective, a rigorous theoretical framework for security is established. Experimental results over AWGN and Rayleigh fading channels demonstrate that the proposed method effectively prevents privacy leakage with only a marginal degradation in the transmission utility for legitimate users.