Personalized Federated Learning for GAI-Assisted Semantic Communications

Abstract

Semantic Communication (SC), which focuses on transmitting meaning rather than raw data, has emerged as the next-generation communication paradigm. However, the performance of SC is heavily influenced by network design and training methodologies. To address these challenges and enhance SC performance at the edge, we first introduce a Generative Artificial Intelligence (GAI)-assisted SC (GSC) model, which improves SC capabilities by optimizing the network architecture. Then, to achieve the efficient learning of GSC models deployed on each user, a Personalized Semantic Federated Learning (PSFL) framework is proposed. Specifically, in the local training phase, a Personalized Local Distillation (PLD) approach is employed, where each user selects a tailored GSC model as a mentor based on local resources. This mentor subsequently distills knowledge to a unified student model, ensuring compliance with the model isomorphism requirements of FL. In the global aggregation phase, an Adaptive Global Pruning (AGP) scheme is applied, dynamically pruning or expanding the aggregated global model based on real-time channel conditions. This mechanism effectively balances accuracy and communication energy efficiency. Finally, numerical results validate the feasibility and efficacy of the proposed PSFL framework, demonstrating its potential to enhance SC performance in edge environments significantly.