Pre-Training Model and Client Selection Optimization for Improving Federated Learning Efficiency

Federated learning(FL), a privacy preserving distributed paradigm of machine learning, enables different clients to learn a global model jointly without sharing their own data. The way of the global model initialization and the selection scheme of the client have a crucial impact on the efficiency of FL. However, most existing studies randomly initialize the global model, which cannot improve the efficiency of FL. In addition, due to noteworthy differences in data characteristics and device functionality between different clients, selecting clients randomly cannot balance the differences in these aspects, which is not conducive to improving the convergence speed and model accuracy of the global model. As a result, we exploit the pre-training model for federated learning and model the client selection as a multi-objective problem and apply a multi-objective optimization algorithm (NSGA-II) to conduct it. We define this client selection scheme as multi-objective optimization selection (MOOS). Before each round of local models aggregation, MOOS provides multiple client combinations for the server. We assess the performance of pre-training model and MOOS in the classification tasks of EMNIS T and CIFAR-l 0 using the deep neural network (DNN). Experimental results demonstrate that the pre-training model can improve the efficiency of FL and compared with the prevalent client selection schemes, MOOS has faster convergence speed, and better accuracy.