网络安全中的联邦学习：性能、鲁棒性与对抗性威胁

Nkuako,KojoA. Monterey,CA;NavalPostgraduateSchool https://hdl.handle.net/10945/74155 ThispublicationisaworkoftheU.S.GovernmentasdefinedinTitle17,UnitedStatesCode,Section101.CopyrightprotectionisnotavailableforthisworkintheUnitedStates. DownloadedfromNPSArchive:Calhoun NAVALPOSTGRADUATESCHOOL MONTEREY,CALIFORNIA THESIS THISPAGEINTENTIONALLYLEFTBLANK THISPAGEINTENTIONALLYLEFTBLANK DistributionStatementA.Approvedforpublicrelease:Distributionisunlimited. FEDERATEDLEARNINGINCYBERSECURITY:PERFORMANCE,ROBUSTNESS,ANDADVERSARIALTHREATS KojoA.NkuakoLieutenant,UnitedStatesNavyBS,KwameNkrumahUniversityofScienceandTechnology,2009MS,SouthernIllinoisUniversity,Carbondale,2013 Submittedinpartialfulfillmentoftherequirementsforthedegreeof MASTEROFSCIENCEINCOMPUTERSCIENCE fromthe NAVALPOSTGRADUATESCHOOLJune2025 Approvedby:ArmonC.BartonAdvisor MarkoOrescaninSecondReader GeoffreyG.XieChair,DepartmentofComputerScience THISPAGEINTENTIONALLYLEFTBLANK ABSTRACT Modernmilitaryoperationsrequireresilient,privacy-preservingdataprocessingacrossdistributedsystems,wherecentralizedlearningoftenfailstomeetsecurityandoperationaldemands.Federated learning(FL)provides a decentralized solution byenablingmodel training without transferring raw data.This thesis evaluates FL’sperformance,robustness,and vulnerability to adversarial attacks in network trafficclassificationatthetacticaledge.Usingsupervisedmodels—randomforest,ExtremeGradientBoosting,andconvolutionalneuralnetworks(CNNs)—thisstudycomparesFLtocentralizedlearningusingInternetofThings(IoT)-23andNavalPostgraduateSchool(NPS)Enterprise Research Network(ERN)datasets.Results show FL matchescentralizedperformance in binary classification but struggles with multiclass tasks,particularlyunderclassimbalance.Labelflippingattackspreserveoverallaccuracywhiledegradingminority-class performance.A novel backdoor method is introduced,embeddingfeature-leveltriggerswithoutchanginglabels,reducingmacroandweightedF1scoresbyupto10%.Thisworkcontributesafederatedsimulationframework,modelcomparison,andthreatevaluation.Findingsrevealtheneedforimprovedmetrics,classbalancing,andintegrateddefensemechanismstoenhanceFLresilience.TheseinsightsinformfuturedeploymentofFLinmilitaryandcoalitiondomainswheredatasensitivity,resourceconstraints,andoperationalsecurityremainparamount. THISPAGEINTENTIONALLYLEFTBLANK TableofContents 1Introduction 123341.1Motivation...........................1.2ResearchQuestions.......................1.3Scope.............................1.4Contributions..........................1.5Organization.......................... 2Background5 572.1TrafficClassification.......................2.2FeatureSelection........................2.3SupervisedMachineLearningModels................2.4FederatedLearningArchitectures..................2.5FederatedLearningAggregatingAlgorithms.............2.6PublicandEnterpriseTrafficAnalysis................2.7AdversarialAttacksinFederatedLearning..............811141720 3Methodology25 25284042463.1EvaluationMetrics........................3.2DatasetandPreprocessing.....................3.3SystemArchitectureandExperimentalSetup.............3.4ModelImplementationandTrainingStrategy.............3.5PoisoningAttack........................ 4Results 5156874.1BinaryClassificationPerformance.................4.2MulticlassClassificationPerformance................4.3DataPoisoning......................... 5Discussion 97 vii 5.1KeyFindings..........................5.2InterpretationofResults.....................5.3ComparisonwithPriorWork....................5.4ImplicationsforFederatedLearninginCybersecurity..........9798101103 1071081106.1SummaryofFindings......................6.2Limitations...........................6.3FutureWork.......................... 121InitialDistributionList ListofFigures Figure2.1Figure2.2Figure2.3Trafficclassificationdecisiontree..................Centralizedlearningarchitecture..................Federatedlearningarchitecture...................91212Figure3.1Figure3.2Figure3.3Figure3.4Figure3.5Figure3.6Figure3.7Figure3.8Figure3.9Figure3.10Figure3.11Figure3.12Figure3.13Figure3.14Figure3.15Figure3.16Figure3.17Figure3.18Networkarchitecturewithdatacollectionpoints..........Flowsegmentationlogic......................Packet-levelfeaturebreakdown...................Burst-levelfeaturebreakdown....................Flowlabelhistogram(logscale)..................Flowlabeldistributionbyclient..................Scatterplotofunshuffledlabels...................Scatterplotofshuffledlabels....................Correlationmatrixofflowfeatures.................Featuredistribution:Benignversusmaliciousflows........Scattermatrixforflowfeatures...................Featureimportancedistribution(logscale).............Top20mostimportantfeatures...................Explainedvariancecurve......................F1-scoreheatmapforbinaryclassification.............F1-scoreheatmapformulticlassclassification...........Federatedlearningsetupwithoutpoisoning............Federatedlearn