Federated Learning for Semantic Communication Based on CNNs and Transformer

This study focuses on the latest research advancements in the field of semantic communication. Traditional communication systems prioritize the transmission of raw data, whilst semantic communication emphasizes conveying the meaning represented by the data. However, the extracted semantic information is often ambiguous and subject to subjective evaluation. To address this problem, this study proposes a model that combines a convolutional neural network (CNN) with a Transformer, called DeepSC‐CT. The model utilizes a CNN to extract semantic information from the data, followed by a Transformer model to capture spatial relationships and contextual information within the semantic content. We utilize federated learning to train the model and propose an adaptive aggregation algorithm to accelerate the convergence process. Moreover, we expand the single‐modality semantic communication model to encompass multiple modalities, such as texts, audio, and images. Furthermore, this study introduces a learnable position‐encoding method for the Transformer. The experimental results and visual effects of audio and image restoration demonstrate that the proposed method exhibits impressive performance and that the proposed model shows robust data restoration capabilities under various signal‐to‐noise ratio conditions.