Semantic Covert Communication: Concealing Sensitive Image Information via Covertness-Oriented Artificial Noise

This paper investigates covert transmission in semantic communication systems, with a particular focus on protecting sensitive information embedded in images. While current research on semantic secure communication emphasizes semantic confidentiality, it often overlooks the covertness of semantic content. To bridge this gap, we propose a semantic covert communication scheme comprising a semantic covert artificial noise module and a deep learning module. The semantic covert artificial noise module ensures that the covertness-oriented artificial noise (CoAN) lies within the null space of the legitimate user’s channel, so that it perturbs only the eavesdropper while preserving the legitimate user’s ability to recover semantic information. The deep learning module is designed to distinguish sensitive content from the background in an image. Furthermore, the semantic covert communication neural network is trained using a customized information-bottleneck–inspired loss function that simultaneously enhances the semantic expressiveness of the representation observed by the legitimate user and improves the semantic representational capability of the CoAN at the eavesdropper, thereby inducing the eavesdropper to misinterpret the transmitted information as background. Simulation results demonstrate that the proposed scheme conceals sensitive information from the eavesdropper’s perspective while maintaining high-fidelity semantic recovery for the legitimate user.