FedDistillSC: A Distributed Semantic Communication System Based on Federated Distillation

In this paper, we propose a federated-distillation-based distributed semantic communication system for image classification tasks, called FedDistillSC, which considers multiple transmitters with varying channel conditions. After local training with public and private datasets, the local models of each transmitter are knowledge distilled by weighted aggregation based on the classification probabilities generated at the receiver to achieve knowledge sharing without exposure of private data, improving the semantic extraction capability of each transmitter. Moreover, our method does not upload/download model parameters, thus saving bandwidth resources and avoiding data security problems caused by model parameter leakage. Simulation results show that the proposed FedDistillSC exhibits excellent Top-1 accuracy performance under three channels. Especially for those transmitters with low signal-to-noise ratio, knowledge distillation could signiflcantly improve their semantic extraction capability.