Semantics Guided Diffusion for Deep Joint Source and Channel Coding Design

Semantic communication (SemCom) offers a promising avenue for enhancing transmission efficiency, especially as traditional bit-level communication approaches its theoretical limits. A key enabling technology in SemCom is deep learning-based joint source and channel coding (DeepJSCC). However, existing DeepJSCC methods lack interpretability because they transmit semantic information implicitly by mapping data to latent features. Additionally, the black-box nature of neural networks makes it challenging to ensure the reliable transmission of critical semantics. To address these challenges, we propose advancing DeepJSCC towards a more “semantic” approach. Specifically, we suggest transmitting interpretable and lightweight semantics as side information alongside JSCC latent features. At the receiver end, we introduce a novel latent diffusion model designed for wireless communication, trained from scratch to integrate seamlessly with DeepJSCC. Using the semantic side information, the receiver employs the proposed semantics-guided latent diffusion for denoising. Furthermore, since accurate channel state information (CSI) is essential in practice, we propose estimating CSI directly from the channel output. The estimated CSI is then used for step matching in the denoising diffusion process, enabling CSI-free transmission. Finally, the denoised feature is fed into the JSCC decoder to reconstruct the image. Numerical results demonstrate that our proposed scheme can achieve comparable performance without accurate CSI. Additionally, guided by semantic information and leveraging the powerful diffusion model, our method surpasses current DeepJSCC schemes, delivering satisfactory reconstruction performance even at SNR = −5dB. This proposed scheme highlights the potential of incorporating diffusion models in future SemCom systems and suggests several promising applications.