基于分布式存储框架的无中心联邦学习  

Distributed-Based Decentralized Federated Machine Learning

作　　者：王丽华[1] 程翔 杨宁彬 宫碧瑶 黄泽宇 陈美燕 WANG Lihua;CHENG Xiang;YANG Ningbin;GONG Biyao;HUANG Zeyu;CHEN Meiyan(The 20th Research Institute,China Electronics Technology Group Corporation,Xi′an 710068,China;Xi′an Branch,China Academy of Space Technology,Xi′an 710100,China;School of Information Science and Technology,Northwest University,Xi′an 710127,China)

机构地区：[1]中国电子科技集团公司第二十研究所,陕西西安710068 [2]中国空间技术研究院西安分院,陕西西安710100 [3]西北大学信息科学与技术学院,陕西西安710127

出　　处：《电子科技》2023年第8期29-34,共6页Electronic Science and Technology

基　　金：国家自然科学基金(61602381);陕西省大学生创新创业训练计划项目(S202110697318,S202110697486,S202110697529)。

摘　　要：联邦学习是一种新的机器学习范式,其允许多个参与者在不共享原始数据的情况下以隐私安全的方式协作地训练一个共享的机器学习模型。由于联邦学习可以解决数据孤岛问题,因此其具有广泛的应用价值。然而在传统联邦学习中,使用单一的中央服务器聚合模型可能会导致单点故障问题。为了克服传统联邦学习中的可能存在的单点故障问题,文中提出一种基于区块链的分布式联邦学习(Distributed Federated Learning,DFL),利用区块链的特点,将存储模型的任务委托给区块链网络中的节点。文中提出了一种异步聚合策略,能够让参与者在任意时间加入联邦学习,从而减少参与者的等待时间。为了克服区块链存储限制,文中还设计了一种模型分块策略。该策略将大规模模型分块以满足区块链的存储要求。通过在多个数据集上训练多种机器学习模型来评估DFL,实验结果表明DFL在克服单点故障的同时实现了优于传统方法的性能。Federated learning is a new machine learning paradigm that allows multiple participants to collaboratively train a shared machine learning model in a private and secure way without sharing raw data.Federated learning has wide application value because it can solve the problem of data island.However,in traditional federated learning,using a single central server aggregation model can lead to a single point of failure problem.In order to overcome the possible single point of failure problem in traditional federated learning,this study proposes a blockchain-based distributed federation learning(DFL),which takes advantage of the characteristics of blockchain to delegate the task of storing the model to nodes in the blockchain network.An asynchronous aggregation strategy is proposed,which enables participants to join federated learning at any time reducing the waiting time of participants.To overcome the blockchain storage limitation,a model chunking strategy is designed to chunk the large-scale model to fit the blockchain storage requirements.The proposed DFL is evaluated by training multiple machine learning models on multiple data sets,and the experimental results show that DFL achieves better performance than traditional methods while overcoming single point of failure.

关 键 词：联邦学习 隐私安全 数据孤岛 区块链 单点故障 分布式 异步聚合 模型分块 

分 类 号：TP309[自动化与计算机技术—计算机系统结构] TP181[自动化与计算机技术—计算机科学与技术]