Secure Federated Semantic Communication With Model Division Multiple Access Against Gradient Inversion Attacks

This letter proposes a federated semantic communication (FedSC) framework based on split learning (SL) and model division multiple access (MDMA), named FedSC-SL-M. To preserve model privacy, the model is partitioned across the participating entities into three layers, namely the tunable layer (P1) on the user side, the middle layer (P2) at the base station, and the task execution layer (P3) on the receiver side. At the P1, a MDMA-based semantic transmission scheme is designed to exploit both shared and personalized semantic information through multi-user semantic communication, thereby enabling the reuse of shared semantic features. The system achieves maximal semantic transmission rate under constraints of semantic accuracy, transmit power, bandwidth, etc. Experimental results demonstrate that FedSC-SL-M outperforms baseline approaches in image classification tasks across varying signal-to-noise ratio (SNR) conditions and two datasets, not only significantly reducing communication rounds but also effectively resisting gradient inversion attacks (GIAs).