Title:Evaluating Multi-Global Server Architecture for Federated Learning

Abstract:Federated learning (FL) with a single global server framework is currently a popular approach for training machine learning models on decentralized environment, such as mobile devices and edge devices. However, the centralized server architecture poses a risk as any challenge on the central/global server would result in the failure of the entire system. To minimize this risk, we propose a novel federated learning framework that leverages the deployment of multiple global servers. We posit that implementing multiple global servers in federated learning can enhance efficiency by capitalizing on local collaborations and aggregating knowledge, and the error tolerance in regard to communication failure in the single server framework would be handled. We therefore propose a novel framework that leverages the deployment of multiple global servers. We conducted a series of experiments using a dataset containing the event history of electric vehicle (EV) charging at numerous stations. We deployed a federated learning setup with multiple global servers and client servers, where each client-server strategically represented a different region and a global server was responsible for aggregating local updates from those devices. Our preliminary results of the global models demonstrate that the difference in performance attributed to multiple servers is less than 1%. While the hypothesis of enhanced model efficiency was not as expected, the rule for handling communication challenges added to the algorithm could resolve the error tolerance issue. Future research can focus on identifying specific uses for the deployment of multiple global servers.