基于多级信任度量的群体学习后门威胁防护  

Swarm Learning Backdoor Threat Protection Based on Multilevel Trust Measurement

作　　者：陈贝 李高磊[1] CHEN Bei;LI Gaolei(School of Electronic Information and Electrical Engineering,Shanghai Jiao Tong University,Shanghai 200240,China)

机构地区：[1]上海交通大学电子信息与电气工程学院,上海200240

出　　处：《现代信息科技》2023年第18期119-124,128,共7页Modern Information Technology

基　　金：国防基础科研项目(JCKY2020604B004);上海市科委“科技创新行动计划”(22511101200,22511101202)。

摘　　要：群体学习是一种基于区块链的分布式模型协同训练框架。由于接入设备类型和用户信任关系多变,群体学习中可能存在由恶意节点发起的投毒行为和后门传播效应。文章从多级信任度量的角度展开研究,首先,通过组合针对区块链的日蚀攻击和针对联邦学习的分布式后门攻击构造一种具有强传染性的攻击方法;其次,结合基于数字签名的用户身份认证和基于模型逆向的后门异常检测建立一个多级信任度量模型;最后,利用群体学习的交叉验证机制进行后门模型剔除以及异常节点注销。实验结果表明,该文提出的防护方案能够有效增强群体学习框架下机器学习模型的安全性。Swarm Learning is a blockchain-based distributed model collaborative training framework.Due to the variable types of access devices and user trust relationships,there may be poisoning behaviors and backdoor propagation effects initiated by malicious nodes in Swarm Learning.This paper studies from the perspective of multi-level trust measurement.First,constructs a highly contagious attack method by combining eclipse attack against blockchain and distributed backdoor attack against Federated Learning.Secondly,a multi-level trust measurement model is established by combining user identity authentication based on digital signature and backdoor anomaly detection based on model reversal.Finally,the cross-validation mechanism of Swarm Learning is used to remove the backdoor model and write off abnormal nodes.Experimental results show that the protection scheme proposed in this paper can effectively enhance the security of machine learning models under the framework of Swarm Learning.

关 键 词：群体学习 区块链 联邦学习 多级信任度量 后门防御 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP309[自动化与计算机技术—控制科学与工程]