A Secure Digital Signature Scheme for Deep Learning-Based Semantic Communication Systems

Deep learning-based semantic communication (DLSC) systems have shown significant potential by transmitting only the semantic information of data, thereby greatly enhancing the efficiency of wireless networks. However, the openness of wireless channels and the inherent vulnerabilities of neural network models make DLSC systems susceptible to various types of attacks. While traditional physical layer keys (PLKs), which leverage reciprocal channel characteristics and randomness between legitimate users, offer a degree of security, they face challenges in ensuring sufficient key randomness and reliable distribution, particularly in static or multi-device environments. To address these issues, we propose a novel digital signature scheme that integrates reconfigurable intelligent surfaces (RIS), multimodal semantic features, and the receiver’s radio frequency fingerprint. Our scheme generates semantic keys (SKey) from multimodal data using deep learning, and combines them with RF fingerprint keys (RFKey), derived from the receiver’s unique radio frequency fingerprint. A key derivation function (KDF) then produces a dynamic digital signature key (DSK) and its corresponding public key (DSK_pub), which is used for signature verification. The DSK is employed to sign a payload consisting of the message and a dynamic PLK, ensuring unique, secure signatures, while the receiver uses the public key to verify their authenticity. This scheme has been implemented in a real-world DLSC platform using SDR and OAI, with experimental results demonstrating high security and minimal performance impact, making it a promising solution for securing DLSC systems.