Learn Efficiently Without a Server: RIS-Aided Federated Learning

Learning in a fully decentralized environment without the assistance of a parameter server (PS) may not always be feasible, particularly for energy-constrained Internet of Things (IoT) devices facing challenges such as lack of line-of-sight (LOS) links or poor channel conditions between clients. On the other hand, the presence of a PS introduces privacy concerns, especially for sensitive applications. The inversion attack, also known as input recovery from gradient, poses a burgeoning threat to the security and privacy of federated learning (FL). This vulnerability allows a 'curious' PS to partially recover clients' private data. To enable energy-efficient learning in a fully decentralized topology, we propose a novel FL approach where clients learn a global model exclusively relying on reconfigurable intelligent surfaces (RISs). The RISs are continuously configured to enable each client to communicate with only one 'other' client at each learning iteration, thus, focusing the beam and utilizing energy efficiently. Additionally, we demonstrate that by altering the communication links (i.e., dynamically changing which client communicates with which) while iterating, via adjusting the RIS configuration, we can maintain the same convergence speed of standard tree topology based FL (PS-based FL). Hence, the proposed algorithm significantly reduces energy consumption compared to fully decentralized FL approach which suffers from slow convergence rate due to sparsity of the network connectivity graph while preserving privacy of clients' data from a curious PS.