一种基于区块链的隐私保护异步联邦学习  

A blockchain-based privacy-preserving asynchronous federated learning

作　　者：高胜[1] 袁丽萍 朱建明[1] 马鑫迪 章睿[3] 马建峰 Sheng GAO;Liping YUAN;Jianming ZHU;Xindi MA;Rui ZHANG;Jianfeng MA(School of Information,Central University of Finance and Economics,Beijing 100081,China;School of Cyber Engineering,Xidian University,Xi7an 710071,China;Institute of Information Engineering,Chinese Academy of Science,Beijing 100093,China)

机构地区：[1]中央财经大学信息学院,北京100081 [2]西安电子科技大学网络与信息安全学院,西安710071 [3]中国科学院信息工程研究所,北京100093

出　　处：《中国科学：信息科学》2021年第10期1755-1774,共20页Scientia Sinica(Informationis)

基　　金：国家自然科学基金(批准号:62072487,61902290);北京市自然科学基金(批准号:M21036);全国统计科学研究(批准号:2020LD01);陕西省重点研发计划(批准号:2020ZDLGY09-06,2019ZDLGY12-04)资助项目。

摘　　要：联邦学习能够在保障本地数据隐私前提下利用分布式数据和计算资源实现机器学习模型联合训练.现有异步联邦学习有效解决了同步联邦学习所存在的计算资源浪费、训练效率低等问题.然而,现有异步联邦学习通过聚合不同节点训练得到局部模型,并通过中心服务器完成全局模型更新,内生性地受制于中心化信用模式,存在单点失效、隐私泄露等问题.为此,提出了一种基于区块链的隐私保护异步联邦学习,通过上链局部模型并通过共识算法生成全局模型,保证异步联邦学习的可信性.为了保证联邦学习的隐私性,同时提高模型效用,提出利用差分隐私中的指数机制以高概率选择贡献度高的模型梯度,并分配较低的隐私预算以保证局部模型的隐私性.另一方面,针对异步联邦学习时钟不同步问题,提出了双因子调整机制进一步提高全局模型效用.最后,理论分析与实验结果表明所提出的方案能有效保证异步联邦学习的可信性和隐私性,同时提高了模型效用.Federated learning enables the joint training of machine learning models by utilizing distributed data and computing resources while protecting local data privacy. The existing asynchronous federated learning can effectively solve the problems such as waste of computing resources and low training efficiency caused by synchronous learning. However, it aggregates local models from different nodes and updates the global model through the central server, which makes it endogenously subject to the centralized trust mode and suffers from some issues such as single point of failure and privacy leakage. In this paper, we propose a blockchain-based privacy-preserving asynchronous federated learning, which ensures the trustability by storing local models into the blockchain and generating the global model through the consensus algorithm. In order to guarantee the privacy of federated learning and improve the model utility, the exponential mechanism of differential privacy is used to select model gradients with high contribution at high probability, and a lower privacy budget is allocated to ensure the model privacy. In addition, in order to solve the problem of clock desynchronization in asynchronous federated learning, we propose a two-factor adjustment mechanism to further improve the global model utility. Finally,theoretical analysis and experimental results demonstrate that our proposed scheme can effectively guarantee the trustability and privacy of the asynchronous federated learning while improving the model utility.

关 键 词：联邦学习 区块链 差分隐私 模型效用 异步训练 

分 类 号：TP309[自动化与计算机技术—计算机系统结构] TP311.13[自动化与计算机技术—计算机科学与技术] TP181