Uncertainty Measured Active Client Selection for Federated Learning in Smart Grid

Abstract

Federated learning is a hot machine learning research direction, its goal is to train a high quality central model while protecting the privacy of all parties, and it has a broad application prospect in smart grid and other fields. However, in federated learning with massive client participation, it is impossible to have all clients participate in training and model aggregation every time due to the limitation of communication and computing resources. Usually the method of selecting clients for federated learning is random, some studies have studied this problem from aspects of client data quality, model training effect, communication and computing resources, etc. In this paper, we propose an active client selection algorithm from the perspective of model uncertainty, this algorithm is called uncertainty measured active client selection in FL (UCS-FL). The server actively selects a subset of clients to participate in the FL training, and the unselected clients do not need to train in this round, saving computing and communication resources. Perform a thorough empirical analysis of the image classification task to demonstrate the excellent performance of UCS-FL against baseline in the context of monitored FL settings. Finally, we describes the real-world application of the proposed architecture, especially in smart grid scenarios.