HFL: Heterogeneous Federated Learning Framework Based on Transformer for Wireless Channel Prediction

In frequency-division duplex (FDD) large-scale multiple-input multiple-output (MIMO) systems, acquiring accurate downlink channel state information (CSI) is crucial for enhancing system performance but faces challenges such as feedback overhead and channel aging. Federated learning (FL) offers a privacy-preserving collaborative CSI prediction approach; however, the inherent heterogeneity of base station environments and antenna configurations in real-world networks impedes the application of FL algorithms. To address this issue, we design a Heterogeneous Federated Learning (HFL) Framework based on Transformer. This framework employs heterogeneous neural network architectures where each client shares a powerful backbone network that learns universal channel dynamics through FL. Additionally, we propose a hierarchical attentionbased dynamic weighted federated aggregation mechanism to tackle convergence challenges caused by heterogeneity. We also investigate the impact of the shared backbone network scale on prediction performance and computational cost, validating its scalability. Simulation results demonstrate that the proposed HFL framework significantly outperforms both local-only training and standard federated averaging baseline methods in terms of CSI prediction accuracy under varying user speeds and signal-to-noise ratios (SNRs), proving the method's effectiveness in handling heterogeneity.