Federated Generative-Adversarial-Network-Enabled Channel Estimation

Abstract

Accurately estimating channel state information is essential for meeting the quality-of-service requirements of modern applications and scenarios. Deep learning techniques have proven effective in acquiring channel conditions with low pilot overhead in massive connectivity scenarios. However, accessing channel data brings new challenges related to transmission overhead, privacy concerns, scalability, heterogeneous network support, and adaptability to dynamic environments. We propose a federated generative-adversarial-network-enabled channel estimator to address these challenges. We refine the coarse least-squares estimation results for their low complexity and fast convergence. To ensure accuracy, we designed a double U-shaped network. The Lipschitz continuous function is applied to discriminators for spectral normalization. We then propose a federated learning framework to utilize the training process. The local generator parameters are updated at the center, reducing communication overhead and privacy concerns. To deal with nonindependent and identically distributed datasets, the discriminators dynamically push away the predictions by dynamic regularization to obtain a more robust aggregated generative model at the center. Furthermore, we propose a motivation scheme that benefits users participating in the training process, encouraging them to join and take advantage of edge/cloud computing capabilities. Numerical results demonstrate that the proposed federated generative adversarial network-enabled channel estimator provides high estimation accuracy and reduces the burden on pilots. The proposed dynamic regularization terms and motivation scheme boost performance efficiently with low communication cost and high participation.