动态信誉驱动的联邦学习恶意攻击检测与隐私保护协同优化方法

打开文本图片集

引用格式：，，，等.动态信誉驱动的联邦学习恶意攻击检测与隐私保护协同优化方法[J].现代电子技术，2026，49（7） ：69-73.

关键词：联邦学习；动态信誉；自适应差分隐私；数据投毒攻击；模型投毒攻击；协同优化中图分类号：TN919-34；TP309；TP181 文献标识码：A 文章编号：1004-373X（2026）07-0069-05

A dynamic reputation-driven federated learning co-optimization approach for malicious attack detection and privacyprotection

GUHaoran1，CHEN Yuanyuan'，YANG Yicheng²，WU Chenyang' （1.School of Informationand Communication Enginering，North UniversityofChina，TaiyuanO3oo51，China; 2.School ofComputer Science and Technology，North University of China，Taiyuan O3oo51，China）

Abstract：Open deployment of federated learning faces multiplechalenges，including maliciousattacksand privacy leakage.Thetraditionalmethods，however，oftenoptimizeasinglemetricseparately，whichleadstoimbalancebetweenmalicious attack detectionandprivacyprotection，makingthemdificulttocopewiththethreatofcompositeaacks.Inviewofthisthe authorproposesadynamicreputation-drivenfederatedlearningco-optimizationmethodformaliciousattackdetectionandprivacy protection.Onthebasis ofthedesignofadynamicreputationassessment module，themethodbreaksthrough the performance botleneckofpartitioneddefense.Areputationevaluationmechanismincorporatingmulti-dimensionaldynamicmetricsisdesigned tocomprehensivelycalculatethereal-timetraininglossdeviationoftheclient，trainingdelayanomaly，andhistoricalbehavioral trustworthinesssoastodynamicallyanditerativelyupdateitsreputationscore.Thesystemwillimplementaquarantine mechanismfortheclientwhosereputationvalueisconsistentlybelowapresetthresholdduetomultiplemaliciousbehavior penalties.Onthebasisoftheclient'sreal-timereputationrating，thehierarchicaldiferentialprivacymechanismimplements diferentialprivacyudgetalocationndachievesadaptiveadjustmentofprivacyprotectionstrength，whichefectivelysegards themodelcontributionandprivacysecurityofthenormaldata whilesuppresing theimpactof maliciousgradients.The experimentsshowthattheproposed methoddemonstratessignificantadvantagesinthemaliciousdetectionaccuracyand privacy protection incomparison with the traditional detectionalgorithms.Ina word，the proposed method is reliable.

Keywords：federatedlearning;dynamicreputation；adaptivedifferential privacy;datapoisoningatack；model poisoning attack;co-optimization

0 引言

随着联邦学习在物联网、边缘计算等分布式场景中的广泛应用，其安全问题逐渐成为制约技术落地的核心挑战[。（剩余5921字）