通信学报 ›› 2024, Vol. 45 ›› Issue (4): 201-215.doi: 10.11959/j.issn.1000-436x.2024072

• 学术通信 • 上一篇    

基于多级代理许可区块链的联邦边缘学习模型

葛丽娜1,2,3, 栗海澳1,3, 王捷1,2,3   

  1. 1.广西民族大学人工智能学院, 广西 南宁 530006
    2.广西混杂计算与集成电路设计分析重点实验室, 广西 南宁 530006
    3.广西民族大学网络通信工程重点实验室, 广西 南宁 530006
  • 收稿日期:2023-09-08 修回日期:2023-12-01 出版日期:2024-04-30 发布日期:2024-05-27
  • 作者简介:葛丽娜(1969- ),女,广西环江人,博士,广西民族大学教授、硕士生导师,主要研究方向为信息安全、物联网、区块链和智能计算等。
    栗海澳(1999- ),男,河南信阳人,广西民族大学硕士生,主要研究方向为网络与信息安全、联邦学习和区块链。
    王捷 (1995- ),女,广西柳州人,广西民族大学讲师,主要研究方向为网络与信息安全和区块链。
  • 基金资助:
    国家自然科学基金资助项目(61862007);广西自然科学基金资助项目(2020GXNSFBA297103)

Federated edge learning model based on multi-level proxy permissioned blockchain

Li’na GE1,2,3, Haiao LI1,3, Jie WANG1,2,3   

  1. 1.School of Artificial Intelligence, Guangxi Minzu University, Nanning 530006, China
    2.Guangxi Key Laboratory of Hybrid Computation and IC Design Analysis, Nanning 530006, China
    3.Key Laboratory of Network Communication Engineering, Guangxi Minzu University, Nanning 530006, China
  • Received:2023-09-08 Revised:2023-12-01 Online:2024-04-30 Published:2024-05-27
  • Supported by:
    The National Natural Science Foundation of China(61862007);Guangxi Natural Science Foundation(2020GXNSFBA297103)

摘要:

针对零信任边缘计算环境下联邦学习面临的隐私安全及学习效率低等问题,提出了一种边缘计算中基于多级代理许可区块链的联邦学习模型,设计多级代理许可区块链构建联邦边缘学习可信底层环境,实现分层模型聚合方案缓解模型训练压力,利用秘密共享和差分隐私设计混合策略增强模型隐私。针对边缘客户端可信度为零或极差的问题,设计了基于信誉验证的联邦任务节点选择算法,将正向训练样本及本地模型作为信誉奖励,完善安全验证方案,进一步保证模型抵御恶意敌手攻击的有效性。实验结果表明,在40%恶意敌手的攻击下,相较于现有的先进方案,所提方案准确率提升了10%,以较高的模型准确率实现了较高的隐私安全。

关键词: 联邦学习, 区块链, 数据安全, 隐私保护, 边缘计算

Abstract:

Aiming at the problems of privacy security and low learning efficiency faced by federated learning in zero trust edge computing environment, a federated learning model based on multi-level proxy permission blockchain for edge computing was proposed. The multi-level proxy permission blockchain was designed to establish a trusted underlying environment for federated edge learning, and the hierarchical model aggregation scheme was implemented to alleviate the pressure of model training. A hybrid strategy was devised to enhance model privacy using secret sharing and differential privacy. A federated task node selection algorithm based on reputation verification was devised to address the problem of zero or extremely poor credibility of edge clients. Positive training samples and the local model were utilized as reputation rewards to refine the security verification scheme, and further ensure the effectiveness of the model against malicious adversaries. Experimental results show that under the attack of 40% malicious adversaries, compared with the existing advanced schemes, the accuracy of the proposed scheme is improved by 10%, and high privacy security is achieved with high model accuracy.

Key words: federated learning, blockchain, data security, privacy-preserving, edge computing

中图分类号: 

No Suggested Reading articles found!