Federated KD-Assisted Image Semantic Communication in IoT Edge Learning

The evolution from fifth-generation mobile communications (5G) to beyond 5G (B5G) will lead to more ubiquitous and smarter paradigms in the Internet of Things (IoT). Communication will shift from classical information theory to a semantic communication (SemCom) paradigm driven by artificial intelligence (AI) to enhance capacity and optimize resources. Image semantic communication (ISC) will empower IoT applications such as drone image acquisition. However, ISC requires suitable devices and sufficient computing resources to support complex neural network models, posing a significant challenge. To address this, we propose a federated semantic feature distillation (FedSFD) architecture to improve the global performance of ISC by combining federated learning (FL) and feature distillation (FD) for feature knowledge transfer. First, lightweight IoT device models in the edge group and the powerful server model alternately minimize losses to update parameters. After learning middle-layer features of all edge models, the server can guide individual device models. Secondly, we incorporate the information bottleneck (IB) concept into the design of loss functions to balance compression and reconstruction. Thirdly, focusing on the trade-off between local training and knowledge interaction, FedSFD achieves image semantic reconstruction without sharing private data, ensuring personalization and privacy protection in the FL framework. Finally, compared to the baseline, simulation experiments show that the proposed approach achieves better ISC reconstruction and noise robustness during group ISC.