Towards Method of Horizontal Federated Learning: A Survey
Dianqi Liu, Liang Bai, Tianyuan Yu, Aimin Zhang
Abstract
With the rapid evolution of artificial intelligence and deep learning, people pay increasing attention to the value of data privacy, which inspires the development of federated learning technologies. As an innovative machine learning application for decentralized data, federated learning unites different clients to cooperatively train on the premise of strictly protecting the data privacy. Among all categories of federated learning frameworks, horizontal federated learning is the most common and popular type. Under the setting of horizontal federated learning, a series of optimization methods based on Fedavg greatly improve the availability of federated learning, but they still face the Non-IID problem seriously. Therefore federated personalization technology is proposed to adapt federated model to local data, which has achieved impressive success in dealing with data heterogeneity and further expanding the application scenarios of federated learning. In this paper, characteristics of different categories of federated learning are described in detail, which enable horizontal federated learning to be distinguished clearly. In addition, we summarize the classical optimization methods of horizontal federated model and the problems it faces. To comprehensively sum up the research for these problems and challenges, we also introduce personalized horizontal federated learning and strategies to implement it.