电子学报2025，Vol.53Issue(7)：2482-2499,18.DOI:10.12263/DZXB.20241078
基于区块链的分层联邦学习系统
Blockchain-Based Hierarchical Federated Learning System
摘要
Abstract
Federated learning can build a distributed and secure computing environment at the cloud-edge-terminal for application scenarios with high data privacy protection and real-time requirements.As a cross-device distributed learn-ing,client heterogeneity and privacy security are two critical issues.Firstly,under the conditions of client data heterogeneity and device heterogeneity,there are large differences in response speed and data distribution,which can lead to lag between clients and greatly affect the performance of federated learning.Secondly,in terms of privacy security,federated learning still has security problems such as single-point attack on the central server,untrustworthy clients,and inference attacks.In this paper,we designed a hierarchical federated learning system FATChain to solve the above problems.Firstly,for the problem of client heterogeneity,an efficient client selection mechanism is proposed to group the selected clients according to their response speeds,and cluster sampling based on the representative gradient is used for each group of clients to ensure that clients with unique data distributions are selected,and synchronous and asynchronous training are combined through hi-erarchical bridging,which reduces the pressure caused by global synchronization while solving the problem of data and de-vice heterogeneity.At the same time,a weighted aggregation algorithm based on the influence function is designed to im-prove the aggregation weight of high-quality local models,to solve the problem that global accuracy is affected by the high weight of low-quality local models due to data heterogeneity,to accelerate the convergence of the global model,and to im-prove the accuracy of model training.Secondly,to address the privacy and security issues,the federated learning algorithm is combined with the blockchain to achieve decentralization and solve the problem of single-point attack.A poisoning attack detection module is set up in the system to filter out the unqualified local updates before aggregation,solving the problem of poisoning attack.And the approach that participant nodes in the blockchain grouping only upload the updates without gener-ating the blocks is utilized,which effectively prevents the inference attack caused by the malicious participant.The analysis shows that the proposed federated learning system well achieves privacy security protection for all parties,while the perfor-mance is greatly improved compared to similar schemes with good scalability.And it is suitable for large-scale application scenarios with high requirements for privacy protection.关键词
区块链/联邦学习/隐私保护/聚类采样/模型聚合/云边端Key words
blockchain/federated learning/privacy protection/cluster sampling/model aggregation/cloud-edge-ter-minal分类
信息技术与安全科学引用本文复制引用
胡荣磊,刘思惠,段晓毅,左珮良,张艳硕..基于区块链的分层联邦学习系统[J].电子学报,2025,53(7):2482-2499,18.基金项目
中央高校基本科研业务费资金资助(No.3282023017,No.3282024052,No.3282024058) Fundamental Research Funds for the Central Universities(No.3282023017,No.3282024052,No.3282024058) （No.3282023017,No.3282024052,No.3282024058）
