MMPL: Multi-Objective Multi-Party Learning via Diverse Steps

Abstract

Multi-party learning is a specific framework of distributed learning which is widely exploited in the medical system and mobile data analysis. By installing a central server, individual devices update the model parameters instead of sharing sensitive data. This method can protect data privacy with the highest measures. But during the communication round, it is difficult to balance the model performance and the computational costs. In this article, we propose a novel framework for multiparty learning, named Multi-objective Multi-Party Learning via Diverse Steps (MMPL). We regard multi-party learning as a multi-objective problem and employ evolutionary optimization for analysis. Within the design of our framework, we try to use the neural network as a bridge to connect evolutionary optimization with multi-party learning. During the study, we propose a novel space-searching strategy for complex encoding problems. An individual wised value is installed on each private device for a differentiated treatment. Experimental results between our method and the comparative algorithms show that the proposed algorithm can achieve better performance while partly ameliorating the time-consuming problem.