Enhancing Deep Joint Source-Channel Coding in Varying Wireless Channels via MoE Approach

Joint Source-Channel Coding (JSCC) holds substantial promise for revolutionizing wireless networks, especially in semantic communication systems for 6 G and beyond. However, classical JSCC models trained for specific channel conditions often struggle to adapt to dynamically varying channel environments and various real-world scenarios, resulting in poor generalization. To address these challenges, this paper proposes a novel end-to-end JSCC framework incorporating a Mixture of Experts (MoE) architecture. Specifically, our model adopts a three-stage training process and designs a gating mechanism to dynamically select the appropriate decoder(s) based on the estimated channel state. Experimental results varying practical wireless channel conditions demonstrate that the MoE-JSCC architecture clearly outperforms the standard deep JSCC model trained across all conditions. It also maintains performance comparable with specialized decoders for specialized channels. The proposed framework enhances adaptability and effectiveness across various scenarios, achieving the effect of transfer learning, making it promising for next-generation wireless networks. Index Terms-Joint source-channel coding, mixture of experts, channel adaptation, gate selection, image communications.