Abstract
Federated learning (FL) is a distributed machine learning method where models on edge devices are trained without sharing personal data and model aggregation occurs solely on a central server. Typically, FL involves clients with limited computing capability and a central server which is much more powerful, therefore it is critical to reduce the client-side overhead. In this study, we propose binary to real-valued federated learning (BRFL), a novel algorithm deploying binary weight networks (BWNs) on clients and real-valued weight network (RWN) on the central server. Gradients are computed locally on clients and sent to the server for aggregation and global model updates. The updated global weights are binarized and then distributed to all clients. To address clients’ privacy concerns, we propose an activation function satisfying differential privacy (DP) for the local BWNs. Experimental results show that BRFL offers lower communication overhead, and after applying DP, the global model performs even better than existing solutions.