联邦学习可视化:挑战与框架

联邦学习是一种保证数据隐私安全的分布式机器学习方案.与传统的机器学习的可解释性问题类似,如何对联邦学习进行解释是一个新的挑战.文中面向联邦学习方法的分布式与隐私安全性的特性,探讨联邦学习的可视化框架设计.传统的可视化任务需要使用大量的数据,而联邦学习的隐私性决定了其无法获取用户数据.因此,可用的数据主要来自服务器端的训练过程,包括服务器端模型参数和用户训练状态.基于对联邦学习可解释性的挑战的分析,文中综合考虑用户、服务器端和联邦学习模型3个方面设计可视化框架,其包括经典联邦学习模型、数据中心、数据处理和可视分析4个模块.最后,介绍并分析了2个已有的可视化案例,对未来通用的联邦学习可视分析方法提出了展望.

Visualization for Federated Learning:Challenges and Framework

Federated learning(FL)is a distributed machine learning solution,which guarantees data privacy and security.Similar to the problem of the interpretability of traditional machine learning,how to explain FL is a new challenge.Based on the characteristics of the distribution and privacy security of FL methods,this paper explores how to design the visualization framework of FL.Traditional visualization tasks often require a large amount of data.However,the privacy feature of FL determines that it cannot retrieve clients’data.Therefore,the available data mainly comes from the training process of the server-side,including server-side model parameters and client training status.Based on the analysis of the challenge of the interpretability of FL,this paper designs a visualization framework that takes into account the clients,the server-side,and FL models.The framework consists of the classical FL model,data storage center,data processing module and visual analysis interface.Finally,this article introduces two existing visualization cases and discusses a more general visual analysis method in the future.