Model-enabled task-oriented semantic communications through knowledge synchronization

Existing semantic communications (SemCom) rely on a shared dataset for training deep learning based encoders/decoders. However, this prerequisite deviates from reality, as the cost of establishing shared datasets is enormous. In this paper, different from the previous prerequisite, we focus on establishing SemCom in the absence of a shared dataset. Specifically, we assume that each communication participant possesses a non-IID private dataset. We propose a Model-Enabled Task-Oriented SemCom (MESC), which establishes SemCom through sharing trained models between participants. First of all, the receiver trains a cGAN and shares it with the sender for information providing. Secondly, the sender utilizes her private dataset and samples generated by the receiver’s cGAN to train her encoder. The encoder is trained through a contrastive learning paradigm with a knowledge synchronization scheme. The designed knowledge synchronization scheme associates samples from different non-IID datasets based on their categories, ensuring that the semantic symbols extracted from different datasets are interoperable and reasonable. The trained encoder is subsequently transmitted back to the receiver, who then utilizes it to train the decoder using the receiver’s private dataset. Extensive experimental results demonstrate that MESC outperforms the SemCom benchmark and traditional coding methods, achieving state-of-the-art communication performance.