Federated Learning Enabled Channel Estimation for RIS-Aided Multi-User Wireless Systems

Channel estimation is one of the essential tasks of realizing reconfigurable intelligent surface (RIS)-aided communication systems. However, the RIS introduces a high-dimension cascaded channel with complicated distribution. In this case, deep learning (DL) enabled channel estimation has been proposed to tackle this problem. In most previous works, model training is conducted via centralized model learning, in which the base station (BS) collects training data from all users and lead to excessive transmission overhead. To address this challenge, this paper proposes a federated deep residual learning neural network (FDReLNet)-base channel estimation framework in an RIS-aided multi-user OFDM system. For each user, we design a deep residual neural network updated by the local dataset and only send model weights to the BS so as to train a global model. To verify the effectiveness and robustness of the FDReLNet, we update the well-trained global model to the newly joint user and test its performance. The simulation results demonstrate that our proposed FDReLNet can significantly reduce transmission overhead while maintain satisfactory channel estimation accuracy.

Automated summary

This paper proposes a federated deep residual learning neural network (FDReLNet)-based channel estimation framework in an RIS-aided multi-user OFDM system. By updating a deep residual neural network on a local dataset and sending only model weights to the base station (BS), a global model is trained. Simulation results demonstrate that the proposed FDReLNet can significantly reduce transmission overhead while maintaining satisfactory channel estimation accuracy.