异质性感知的自适应安全联邦聚合方法

打开文本图片集

关键词：异质性感知；度量隐私；四分位距；梯度泄露攻击；梯度敏感性；隐私安全 中图分类号：TP309 文献标志码：A 文章编号：1001-3695（2025）12-029-3752-07 doi：10.19734/j. issn.1001-3695.2025.04.0142

Heterogeneity aware adaptive federated security aggregation method

Li Xiaohuia†，Wang Chaoa，Wang Shiyuana，Zhang Xingʰ，Lan Jiee （a.Schoolofcs&fEeCuoscCtefceogUs nology， Jinzhou Liaoning 121001，China）

Abstract：This paperproposed theHAFSA toaddress thechallngesof modelperformancedegradationinnon-independent andidenticallydistributed（Non-ID）dataenvironments，insuicientdynamicsensitivityprivacyprotection，androbustnes imbalancecausedbymalicious atacksandcommunicationdelaysinfederatedlearning.Themethodcontainedthree keycomponents.Firstly，the serverdynamicalladjustedaggregation weightsbasedonlocal model deviation degrees andperformedlatencycompensationfordelayed clients.Secondly，clientsadded local noisetosensitivelayersacording togradientsigificancewhiletheserverdynamicallyalocatedcentralprivacybudgetsbymeasuringprivacylevelsthroughlocalupdatedeviations.Thirdly，theserverseletedvalidclientsusing interquartilerange methodbycombininglossvaluesandmodelupdate similarities，thenimplementedtruncated medianaggregation fortheseselectedclients.ExperimentalresultsshowthatHAFSA outperformed comparative algorithms on the CIFAR-10 and MNIST datasets，with accuracy improved by 12%～20% ，convergencespeedmorethandoubled，andefectiveresistancetogradientleakageatacks.Ablationexperimentsfurtherverifiedthe significantcontributionofeachcomponent to systemperformance.The experiments confirmedthat HAFSAhasachieved anefficientbalanceamong privacyprotection，communication eficiency，and modelrobustness inasynchronous federatedlearning.

Keywords：heterogeneityawareness；metricprivacy；interquartlerange；gradientleakageatack；gradientsensitivity；privacy and security

0 引言

联邦学习（federatedleaming，FL）[1]作为分布式机器学习领域的革命性范式，凭借“数据不动，模型动”的核心机制成功破解了数据孤岛与隐私保护之间的博弈困局[2]。（剩余19227字）