首页|基于对比训练的联邦学习后门防御方法

基于对比训练的联邦学习后门防御方法

扫码查看
原文链接
  • 万方数据
  • 维普
针对现有联邦学习后门防御方法不能实现对模型已嵌入后门特征的有效清除同时会降低主任务准确率的问题,提出了一种基于对比训练的联邦学习后门防御方法ContraFL.利用对比训练来破坏后门样本在特征空间中的聚类过程,使联邦学习全局模型分类结果与后门触发器特征无关.具体而言,服务器通过执行触发器生成算法构造生成器池,以还原全局模型训练样本中可能存在的后门触发器;进而,服务器将触发器生成器池下发给各参与方,各参与方将生成的后门触发器添加至本地样本,以实现后门数据增强,最终通过对比训练有效消除后门攻击的负面影响.实验结果表明,ContraFL能够有效防御联邦学习中的多种后门攻击,且效果优于现有防御方法.
Backdoor defense method in federated learning based on contrastive training
In response to the inadequacy of existing defense methods for backdoor attacks in federated learning to effec-tively remove embedded backdoor features from models,while simultaneously reducing the accuracy of the primary task,a federated learning backdoor defense method called ContraFL was proposed,which utilized contrastive training to dis-rupt the clustering process of backdoor samples in the feature space,thereby rendering the global model classifications in federated learning independent of the backdoor trigger features.Specifically,on the server side,a trigger generation algo-rithm was developed to construct a generator pool to restore potential backdoor triggers in the training samples of the global model.Consequently,the trigger generator pool was distributed to the participants by the server,where each par-ticipant added the generated backdoor triggers to their local samples to achieve backdoor data augmentation.Experi-mental results demonstrate that ContraFL effectively defends against various backdoor attacks in federated learning,out-performing existing defense methods.

federated learningbackdoor attackcontrastive trainingtriggerbackdoor defense

张佳乐、朱诚诚、成翔、孙小兵、陈兵

展开 >

扬州大学信息工程学院,江苏 扬州 225127

中国民航大学民航飞联网重点实验室,天津 300300

南京航空航天大学计算机科学与技术学院,江苏 南京 211106

联邦学习 后门攻击 对比训练 触发器 后门防御

国家自然科学基金江苏省基础研究计划自然科学研究项目江苏省高等学校基础科学自然科学研究项目中国博士后科学基金中国民航大学民航飞联网重点实验室开放基金江苏省研究生科研创新计划

62206238BK2022056222KJB5200102023M732985MHFLW202304KYCX23_3534

2024

10.11959/j.issn.1000-436x.2024063

通信学报
中国通信学会

通信学报

CSTPCD北大核心
影响因子：1.265
ISSN：1000-436X
年,卷(期)：2024.45(3)
  • 40