Two-Stage Attention Based Collaborative Inference in Semantic Communication

This paper presents a heterogeneous federated semantic communication system for time series forecasting tasks (HSC-TSF) within an edge network. To extract forecasting-oriented semantic information, we devise a two-stage attention mechanism comprising in-window attention and cross-window attention. The in-window attention efficiently extracts local temporal features with linear complexity, while the cross-window attention achieves global awareness by aggregating information from the same positions across all windows. In addition, an unsupervised heterogeneous federated learning architecture is adopted to facilitate knowledge sharing among all mobile devices. This approach summarizes semantic features generated from a public dataset and utilizes cross-correlation matrices to enhance the performance of each heterogeneous device individually. Simulation results demonstrate that the proposed HSC-TSF system achieves outstanding performance compared with both SC-based and traditional technique-based baseline systems under additive white Gaussian noise, Rayleigh, and Rician channels. HSC-TSF attains the lowest mean square error across three time series datasets.