Cross-Domain Federated Semantic Communication with Global Representation Alignment and Domain-Aware Aggregation

TL;DR

The paper tackles the challenge of cross-domain data in federated learning for DL-based semantic communication aimed at image reconstruction. It introduces a cross-domain FL framework that builds a global representation from local features and enforces alignment with client-domain features, coupled with a domain-aware aggregation strategy to prevent domination by any single domain. The approach, implemented with a Swin Transformer-based DeepJSCC and evaluated on the PACS dataset, achieves superior cross-domain generalization and reconstruction quality (PSNR/MS-SSIM) across most domains and channel conditions, compared to FedAvg, FedProx, and MOON. The work demonstrates meaningful improvements in robustness to domain shift and channel variability, with a quantified trade-off between cross-domain generality and per-domain specialization, and provides convergence analysis and practical guidelines for lambda tuning. This framework advances practical semantic communication by enabling scalable, privacy-preserving learning across diverse data distributions in wireless networks.

Abstract

Semantic communication can significantly improve bandwidth utilization in wireless systems by exploiting the meaning behind raw data. However, the advancements achieved through semantic communication are closely dependent on the development of deep learning (DL) models for joint source-channel coding (JSCC) encoder/decoder techniques, which require a large amount of data for training. To address this data-intensive nature of DL models, federated learning (FL) has been proposed to train a model in a distributed manner, where the server broadcasts the DL model to clients in the network for training with their local data. However, the conventional FL approaches suffer from catastrophic degradation when client data are from different domains. In contrast, in this paper, a novel FL framework is proposed to address this domain shift by constructing the global representation, which aligns with the local features of the clients to preserve the semantics of different data domains. In addition, the dominance problem of client domains with a large number of samples is identified and, then, addressed with a domain-aware aggregation approach. This work is the first to consider the domain shift in training the semantic communication system for the image reconstruction task. Finally, simulation results demonstrate that the proposed approach outperforms the model-contrastive FL (MOON) framework by 0.5 for PSNR values under three domains at an SNR of 1 dB, and this gap continues to widen as the channel quality improves.

Figures (7)

• Figure 1: The example of domain shift for the DL-based Semantic Communication: the model performs well under the trained domain but fails to obtain good performance for other domains.
• Figure 2: A) Provide the general components of the semantic communication system that follow the deep learning-based joint source-channel coding direction. B) Present the implementation of a federated learning framework for training the DL-based semantic communication system, which requires training on a large amount of data.
• Figure 3: Visualization of data from the same class but different domains, showing that their latent features and distributions drift significantly. Each domain can be regarded as a distinct task; therefore, simply aggregating models trained on different tasks may degrade task-specific performance and hinder generalization.
• Figure 4: (LHS) The local feature representation extractions and the construction of global representations; (RHS) The domain-aware aggregation approach to overcome the dominance of a subset of clients.
• Figure 5: The comprehensive performance of the proposed frameworks against other FL approaches for four kinds of domain data. The comparison performance is measured in terms of PSNR and MS-SSIM metrics.