SC-Diffusion: Parameter Generation for Task-Oriented Semantic Communication Systems via Conditional Diffusion Model

Task-oriented semantic communications (ToSC) has received significant attention as a promising paradigm for realizing more efficient and intelligent data services. However, ToSC systems often suffer from limited generalization capabilities, requiring retraining to meet performance demands under varying channel conditions. In recent years, artificial intelligence generated content (AIGC) has shone in computer vision (CV) and natural language processing (NLP), and its potential in wireless communications is also emerging. Motivated by these advances, we propose semantic communications (SC)-diffusion in this paper, which generates high-performance parameters for ToSC systems to address the inherent challenges of semantic communications. Specifically, SC-diffusion begins by using an autoencoder to extract latent representations from trained system parameters. A diffusion model is then trained to generate these latent representations from random noise. In particular, to ensure that the generated parameters are adapted to the real-time communication environment, we incorporate channel information as conditional information into the diffusion model. Finally, the latent representations are decoded by the autoencoder’s decoder to yield the final system parameters. In experiments across various ToSC architectures and real-world datasets, SC-diffusion consistently generates models that perform comparable to or better than the original trained models, with minimal additional computational overhead.