An Efficient Federated Learning Framework for Training Semantic Communication System

TL;DR

This work addresses private training of semantic communication systems by integrating federated learning to exploit client data without leakage. It introduces FedLol, a loss-based aggregation scheme, and a partial-update strategy that reduces communication overhead by updating the heavy semantic encoder/decoder less frequently than other components. Empirical results on ImageNet10 non-IID clients and DIV2K show that FedLol outperforms FedAvg, FedProx, and MOON in PSNR and MS-SSIM, while achieving about 25% savings in transmitted data. The approach demonstrates that selective parameter transmission and a robust aggregation rule can yield state-of-the-art performance for privacy-preserving, efficient FL in semantic communication tasks.

Abstract

Semantic communication has emerged as a pillar for the next generation of communication systems due to its capabilities in alleviating data redundancy. Most semantic communication systems are built upon advanced deep learning models whose training performance heavily relies on data availability. Existing studies often make unrealistic assumptions of a readily accessible data source, where in practice, data is mainly created on the client side. Due to privacy and security concerns, the transmission of data is restricted, which is necessary for conventional centralized training schemes. To address this challenge, we explore semantic communication in a federated learning (FL) setting that utilizes client data without leaking privacy. Additionally, we design our system to tackle the communication overhead by reducing the quantity of information delivered in each global round. In this way, we can save significant bandwidth for resource-limited devices and reduce overall network traffic. Finally, we introduce a mechanism to aggregate the global model from clients, called FedLol. Extensive simulation results demonstrate the effectiveness of our proposed technique compared to baseline methods.

Figures (3)

• Figure 1: The left-hand side figure demonstrates the proposed Federated Learning framework in a semantic communication system. The right-hand side figure shows the architecture of the Semantic Encoder and the detailed components of the Swin Transformer blocks.
• Figure 2: (a) The data distribution of each client using non-IID data partition. The color bar indicates the number of data samples, while each rectangle points out the number of data samples of a specific class in a client. (b) The convergence of the FedAvg in terms of full-update and partial-update. (c) Communication cost of the full-update and partial-update scenarios.
• Figure 3: The PSNR (a) and MS-SSIM (b) values of the proposed algorithm compared to other benchmarks.