Channel-Adaptive Denoising Diffusion Models for Reliable Semantic Communications

Semantic communication (SC) aims to deliver the profound meaning of information, in which deep-learning-based joint source-channel coding (DeepJSCC) is usually utilized to enable end-to-end communications. DeepJSCC-based SC frame-works perform well in wireless image transmission tasks, particularly under channel conditions with low signal-to-noise ratios (SNRs). However, existing methods still cannot adaptively eliminate different levels of channel noises, limiting the adaptability and effectiveness of DeepJSCC-based SC frameworks. To this end, we design a Channel-Adaptive Denoising Diffusion Model in the SC framework (SC-CAD2M), which can adaptively learn the transmitted semantic features under uneven noise distributions. Specifically, we design the Noise-Adaptive Mask (NAM) module to adjustably reconstruct semantic features across various SNR conditions. Furthermore, we incorporate NAM and CAD2M into the DeepJSCC-based SC framework, which are capable of adaptively predicting and eliminating a diverse range of noises from received semantic features across different channel conditions. We conduct extensive simulations to evaluate the algorithm's performance. Results demonstrate the SC-CAD2M framework's superiorities and adaptabilities for image transmission tasks in dynamic and complex communication scenarios.