2024利用生成式AI增強(qiáng)自主系統(tǒng)的安全性和韌性白皮書(shū)（英文版）-42正式版

EnhancingSecurity,Resilience,andSafetyofAutonomousSystemswithGenerativeAItii.aeRapidadvancesinautonomoussystemsandedgeroboticshaveunlockedunprecedentedopportunitiesinindustriesfrommanufacturingandtransportationtohealthcareandexploration.Increasingcomplexityandconnectivityhavealsointroducednewchallengesinensuringsecurity,resilience,andsafety.Asedgerobotsintegrateintoourdailylivesandcriticalinfrastructures,itisimperativethatwedevelopinnovativeapproachesthatcanraisethesesystems’trustworthinessandreliabilitytonewlevels.ThiswhitepaperexploresthetransformativepotentialofgenerativeAI(GenAI)toenhancethesecurity,resilience,andsafetyofautonomoussystemsandedgerobots.Wecanusethesecutting-edgetechnologiestomeettheuniquedistributedanddynamicchallengesofedgerobotics,tounlocknewlevelsofintelligence,adaptability,androbustness.GenAImodelsproducenewcontentbyanalyzingpatternsinadataset.Theyderivecharacteristicprobabilitydistributionsandapplythesetocreatenewdatapatternsthatareconsistentwiththeoriginal“real”dataset.EarliergenerationsofdiscriminativeAImodelsappliedconditionalprobabilitiestopredictoutcomesforpreviouslyunseendata.Theapproachisversatileandwell-suitedtoawiderangeofproblems,includingclassificationsandregressions.Theyexcelatdelineatingthedecisionboundariesthatdifferentiatebetweenvariousclassesorcategorieswithinthedataset.Thegrowingranksofgenerativetechniquesincludethosebasedontransformersandtheresultinglargelanguagemodels(LLMs),GenerativeAdversarialNetworks(GANs),VariationalAutoencoders(VAEs),GenerativeFlowModels(GFM),andGenerativeDiffusionModels(GDM).ThesehaveallopenedexcitingnewavenuesofAIresearch—withapplicationstodroneswarmsandintrusiondetection,physicalcommunicationsecurity,semanticcommunication,andmobilenetworks.1TheTechnologyInnovationInstitute’sSecureSystemsResearchCenter(TII-SSRC,AbuDhabi,UAE,https://www.tii.ae/secure-systems)isworkingtoapplyGenAItoitsworkextendingZeroTrustarchitectures(developedforinformationsecurity)toeveryaspectofinformationsecurityincyber-physicalsystems.Thus,SSRCconsidershowGenAIcanhelpguaranteesecurity,resilience,andsafetyfordrones,swarms,swarmsofswarms,autonomousterrestrialandmarinevehicles,commandsystems,andhuman/droneinteractions—andparticularlyinareaswhereGenAIoutperformstraditionalAI/MLapproaches.Examplesinclude:?

Individualapplications:healthmonitoring,stateestimation,predictivemaintenance,anomalydetection,self-healing,navigation,andemergencylandings.?

Fleetapplications:Swarmcoordination,swarmintelligence,collectivedecision-making.?

Human/Droneinteractions:communicationresilience,missionplanning,human-computerinteraction.?

Cybersecurityandresilience:Intrusiondetection,malwareclassification,threatsimulation.ThispaperwillfocusondronesbecauseSSRCisdoingsomuchworkinthisarea.Whatwelearnfromdronescanbeappliedmuchmoretoautonomousandcyber-physicalsystemsingeneral,includingcars,robots,embeddedsystems,andsmartcities.Bythesametoken,lessonslearnedintheseareascanbefoldedintoSSRC’sresearch.approachallowsorganizationstomoveawayfromphysicaldevicemanagementapproachesthatrequireemployeestocarrymultiplephones.1M.Xuetal.,“UnleashingthePowerofEdge-CloudGenerativeAIinMobileNetworks:ASurveyofAIGCServices.”arXiv,Oct.31,2023.doi:10.48550/arXiv.2303.16129Sidebar:GenAImodelsAsurveyofspecificgenerativeAImodelingtechniques,withtheirstrengthsandlimitationsasappliedtodronesafety,security,andresilience.PopularexcitementovergenerativeAImodelshasbeendrivenbyhighlypublicizednewserviceslikeChatGPT,whichcreateauthoritative-soundingresponsestotextpromptsusingspeciallytrainedLLMs.LargeLanguageModelsareAIsystemstrainedonvasttextdatasets.TheyuseDeepLearningtechniques,particularlyastructureknownasaTransformer,to“understand”andgeneratehuman-liketextbasedonthepatternsthey'velearned.LLMsanalyzetherelationshipsandcontextsinthetrainingdata.Theyuseavarietyoftechniquestobuildsimplifiedrepresentationsofthedata,allowingthemtomakeassociationsandcorrelationsbetweentheoriginaldataelements.Thesemodelsletthemcomposeresponsesthatcanmimichumanwritingstylesandcoverdiversetopics.VisualAIs,likeDALL-EandStableDiffusion,compilenewimagesfromtextandimageprompts.LLMsarewidelyused,andwidelyuseful,increatingcontent,code,translations,summaries,syntheticdata,andtostructureunstructureddatafromtexts,documents,images,audio,andotherpromptdata.Transformermodelsandtheservicesbuiltonthem—likeOpenAI’sChatGPT,Google’sGemini,andAnthropic’sClaude—haveattractedwidespreadattention,thankstotheirimpressiveabilitytocreatearticulate-seemingresponsestohumanprompts.These,andotherdomain-specificLLMsandSmallLanguageModels(SMLs),alsoshowpromiseforsupportinganalysis,research,anddevelopmenttoimprovedronesafety,securityandresilience.ParallelingtheseveryvisibleAIdevelopments,however,hasbeenalmostadecadeofprogressonnewclassesofGenerativeAImodelscouldautomateandacceleraterepresentation-building.Whilegenerativeapplicationsattractthemostattentions,thesenewmodelsalsoaredrivingadvancesinanalyzingdataandinteractingwiththeworldaroundus.OthergenerativeAImodels—GenerativeAdversarialNetworks,VariationalAutoencoders,GenerativeDiffusionModels,andNormalizingFlowModels—thoughrelativelyunknown,canmakesubstantialcontributionstodronesecurity,safety,andresilience.TransformerModels:Introducedin2017totranslatebetweenEnglishandFrenchtexts,TransformerModelsexcelatcapturinglong-rangedependenciesandcorrelationswithinunstructureddata.2Transformersleverageanovel“attentionmechanism”tolearntheconnectionsbetweenwordstohelpcreateembeddingsautomatically.Priortechniquesrequiredtranslatingrawtextintoavectorrepresentationusingaseparatemodel.Transformerscanbuildcomplexrepresentationsandlearnintricateconnectionsthroughtheirlayeredarchitecture,manner.,allowingresearcherstoprocesslargebodiesofunlabeledtextanddeveloplargelanguagemodelswithbillionsofparameters.Subsequentinnovationshavesupporteddocumentsummarization,composingquestion/answerassociationsacrosslargedatasets,codegeneration,in-depthanalysis,intrusiondetection,malwaredetection,andtranslatingcontrolsysteminstructionsacrossroboticarms.Theapproach’skeyadvantageisdistillingcontextfromcomplexdatasets.Challengesincludehallucination,longertrainingtime,slowerinference-building,heaviercomputationrequirements,andlargermodelsizescomparedtoothertechniques.2A.Vaswanietal.,“AttentionIsAllYouNeed.”arXiv,Aug.01,2023.doi:10.48550/arXiv.1706.03762.GenerativeAdversarialNetworks(GANs):Theseweredevelopedin2014tocreaterealisticsyntheticnumbers,faces,andanimalimages.3GANspittwoneuralnetworksagainsteachother:oneisrewardedforgeneratingmorerealisticcontent,andthesecondisrewardedfordetectingfakecontent.Inthiscompetition,thegeneratorimprovesitsabilitytocreaterealisticoutputsthatcanfoolthediscriminator.GANsarewidelyusedincontentgeneration.Sincethefirstversionsweredesignedtoworkwithimages,researchersarenowfindingcreativewaystotranslatedata,suchascodeornetworklogs,intoimagessuitableforGANprocessing.GANsaregoodforrealisticsyntheticdatasetsthatcanbeusedtoimproveautonomoussystemsandcybersecurityalgorithms.They,too,however,sufferfromfailureslikemodecollapseorcatastrophicforgetting.3I.J.Goodfellowetal.,“GenerativeAdversarialNetworks.”arXiv,Jun.10,2014.doi:10.48550/arXiv.1406.2661.VariationalAutoencoder(VAE):VAEswereintroducedin2014toimproveinferencesdrawnfromacontinuouslyvaryingdatastream.4Thetechniquehelpsfindefficientwaystorepresentdataandcanbeusedtocompressdataordetectanomaliesandthreats.TrainingVAEsprocessinvolvesteachingasetofencodersanddecoderstotranslaterawdataintoanintermediatelatentspacewithadifferentprobabilitydistribution.VAEscanbeusedindependentlyinapplicationslikeanomalydetection,designingbetterencodingschemes,dataaugmentation,andimagegeneration.Inaddition,theyareoftenusedtopre-structuredataforotheralgorithms,includingGANs,toimprovetheirresults.GenerativeDiffusionModel(GDF):GDFsemergedin2015toimprovelearning,sampling,inferences,andevaluationsthatwereinformedbynon-equilibriumthermodynamicsmodeling.5Thetechniqueaddsnoisetoasample(suchasanimage)andthenautomatesthedenoisingprocesstorevealthedata’sunderlyingstructure.Slightvariationscanleadtovalidnewtrainingdatasets.GDFsarewidelyusedinimagegenerationandcanimprovesignalclassificationinvariousdroneusecases.However,thetechniquerequireshighersamplingtimesanddemandsamorecomplexarchitecturethanGANsandVAEs.NormalizingFlowModels(NFMs):Thesewereintroducedbyresearcherstomakecomplexdatasimplertoworkwith.6Thesemodelstakeeasy-to-understanddistributions,likeanormalbellcurve,andtransformthemstepbystep.Eachstepisreversible,meaningwecanalwaysgobacktothestartifneeded.Thisprocess,calleda“flow,”movesfromasimplebeginningtoanendthat4D.P.KingmaandM.Welling,“Auto-EncodingVariationalBayes.”arXiv,Dec.10,2022.doi:10.48550/arXiv.1312.6114.5Yang,Ling,etal."Diffusionmodels:Acomprehensivesurveyofmethodsandapplications."ACMComputingSurveys56.4(2023):1-39./doi/10.1145/36262356Kobyzev,Ivan,SimonJDPrince,andMarcusA.Brubaker."Normalizingflows:Anintroductionandreviewofcurrentmethods."IEEEtransactionsonpatternanalysisandmachineintelligence43.11(2020):3964-3979./abstract/document/9089305/authorsresemblesthecomplicatedtargetdataset.Bydoingthis,itispossibletostudyandusethedatamoreeffectively.NFMshavebeenusedtogeneratehandwrittennumbers,images,etc.Newerusecasesincludeenhancedclassificationandencodingschemes.Thetrainingprocesscreatesamodelthattransformstheprobabilitydistributionofadatasetintoamorecomplex,fullyreversibledistribution.NFMscan,however,requirehighercomputationandtrainingtimesthantechniqueslikeGANsandVAEs.ThefollowingfiguresummarizesthemainGenAITechniquesandtheirapplicationsinthefieldofZeroTrustforautonomoussystems.GenAIUseCasesTheproliferationofdrone-technologyhasbroughtchallengesthatspancrossbetweendomains—individual,fleet,humancontrol,andcybersecurity.Theirgrowthandcomplexitydemandconstantinnovationtoredoubletheirtrustworthinessandreliability.Thefollowingapplications—whetherderivedfromUnmannedAerialVehicle(UAV)anddroneresearchorimportedfromotherdomains—haveimportantimplicationsforthefutureofUAVsandotherautonomoussystems.Note,too,thatmanyoftheseareearly-stageprojects,includedtogiveaflavorwhatGenAItoolsmightaccomplishastechniquesevolve.本報(bào)告來(lái)源于三個(gè)皮匠報(bào)告站（）,由用戶Id:673421下載,文檔Id:490694,下載日期:2025-01-23GenAIshowstremendouspotentialforimprovingZeroTrustframeworkstoenhancesecurity,resilience,andsafetyinindividualautonomoussystems,suchasdrones,self-drivingcars,robots,andembeddedsystems.Usecasesunderinvestigationincludeboostingself-awareness,anomalydetection,autonomousdriving,predictivemaintenance,faultmanagement,self-healing,andlandingsafely.Self-AwarenessOpportunity:Efficientlytranslatenoisy,blurry,andinconsistentdatatounderstandthedrone'scurrentstate—e.g.,compensatingformotionblurwhiletryingtodetectobstacles.Thefoundationofdronehealthisaccuratelycapturingandmakingsenseofitscurrentstate—includingtheconditionofitscurrenthardware,itsapplications,itsphysicallocation,anditssecurityposture.Intherealworld,thiscangetmessy,asvideofeedssuffermotionblur,GPSdatajitters,inertialguidancedatalosecalibration,andnoiseorgapsdegradeinternalmonitoringdata.Stateestimationiscrucialtoautonomousnavigationanddecision-making,andrawdatastreamsmustbeaccuratelycorrelatedwithposition,velocity,andorientation.7GenerativeAIcanhelpfillinmissingdataandfusedatafrommultiplesourcestoimprovestateestimation.8InnovationsinGenAIalgorithmslikeGANs,VAEs,andtraditionalMLalgorithmslikeLSTMcanfillinthesegaps,preservingvehiclesafetyviafaultdetection,predictivemaintenance,faultmanagementandsafe-landingprotocols.Forexample,innovativeGANapproachescanfillinmissingdataandmakeiteasiertofusedatastreamstocreateamoreaccuratestateassessment,9helpcorrelateinternallogdatawithacousticanalysis,10andidentifypotentialmechanicalissues.11Researchershavealsodevelopedtechniquesforgeneratingestimated-statevariablesusingConditionalGANs(CGANs)forindividualdronesanddroneswarms.127T.D.Barfoot,Stateestimationforrobotics.CambridgeUniversityPress,2017.8Liu,Guangyuan,NguyenVanHuynh,HongyangDu,DinhThaiHoang,DusitNiyato,KunZhu,JiawenKang,ZehuiXiong,AbbasJamalipour,andDongInKim.“GenerativeAIforUnmannedVehicleSwarms:Challenges,ApplicationsandOpportunities.”arXiv,February28,2024./10.48550/arXiv.2402.18062.9Y.He,S.Chai,andZ.Xu,"Anovelapproachforstateestimationusinggenerativeadversarialnetwork,"in2019IEEEInternationalConferenceonSystems,ManandCybernetics(SMC),2019,pp.2248-2253./document/891458510Y.Wang,A.Vinogradov,“ImprovingtheperformanceofconvolutionalGANusinghistory-stateensembleforunsupervisedearlyfaultdetectionwithacousticemissionsignalsAppl.Sci.,13(5)(2023),p.3136,doi:10.3390/APP1305313611S.Zheng,A.Farahat,andC.Gupta,“GenerativeAdversarialNetworksforFailurePrediction.”arXiv,Oct.04,2019.Accessed:Mar.15,2024.[Online].Available:/abs/1910.0203412.A.He,C.Luo,X.Tian,andW.Zeng,"AtwofoldSiamesenetworkforreal-timeobjecttracking,"inProceedingsoftheIEEEconferenceoncomputervisionandpatternrecognition,2018,pp.4834-4843./document/8578606AnomalydetectionOpportunity:Improveanalysisofdronesensordatatoidentifyabnormalconditions.Moreaccurate,multi-dimensionalsystem-staterecordscanalsohelpidentifyanomaliesrelevanttodronehealth.Forexample,VAEscanimprovefaultdetectionandisolation.Theycanalsoidentifythewarningsignsofstressinvarioussystems,toprioritizepredictivemaintenanceschedules.Typically,machine-learningclassificationalgorithmsaretrainedonmultipleclasses(suchasdatatagged“faulty”and“notfaulty”).Dataonrarerclasses,suchas“faulty”data,isoftenscarceinpublicdatasetsandtherealworld.Inthesecases,GenerativeAdversarialNetworks(GANs)canbevaluableinsynthesizingtheserareclasses—makingup“faulty”datathatlooksconditions.Inaddition,researchersareexploringhowLLMscouldhelpbettercontextualizetooperateinnewenvironments.13Forexample,DriveLLMcombinesLLMwithtraditionalautonomousnavigationalgorithmstosupportbetterreasoninganddecision-makingwhenrespondingtoedgecases.14Researchersfoundthemethodcouldimproveproactivedecision-makinginunexpectedcircumstances.Anotherapplication,TypeFly,enhancescommunicationbetweenhumansanddronesthroughanaturallanguageinterface15.Suchlargelanguagemodelsmay,however,manifestsurfacebias,inaccuracies,andhallucinationissuesthatrequireadditionalsafeguards.Similarly,MicrosoftResearchdiscussestheiradvancementsinintegratingChatGPTwithroboticstomakerobotcontrolmoreintuitivethroughnaturallanguage.They'veenabledChatGPTtounderstandandexecutetasksinphysicalenvironments,whichfacilitateseasierhuman-robotinteractionwithouttheneedforcomplexprogrammingknowledge.TheChatGPTteamhasdevelopeddesignprinciplesforlanguagemodelstosolveroboticstasks(involvingspecialpromptingstructuresandhigh-levelAPIs),andtheyhavedemonstratedhowChatGPTcanhandletaskslikeoperatingdronesandrobotarmsthroughuser-friendlycommandsandfeedback.Thedevelopersemphasizetheimportanceofsafetyandsimulationtestingbeforereal-worldapplication.16AdaptabilityOpportunity:Improvetranslationofautonomoussystemssoftwaretorunacrossdifferenthardwaremakes,models,andconfigurations.Autonomoussystemcontrollersmustbetrainedforaspecificmodelandconfiguration.Thiscancreatechallengeswhenupgradingindividualcomponentsoradoptingnewmodels.RTXisarobotcontrolLLMthatcantranslatecontrolpoliciestomanagedifferentroboticarmswithoutrefactoringthecontrolalgorithmsforthelatesthardware.Insometests,leveragingtheexperienceofothercontrollers,producedcontrolpoliciestheoutperformedthebestcontrolscustom-builtforanindividualarm.1713Wang,Lei,etal."Asurveyonlargelanguagemodelbasedautonomousagents."FrontiersofComputerScience18.6(2024):1-26./article/10.1007/s11704-024-40231-114Y.Cuietal.,“DriveLLM:ChartingthePathTowardFullAutonomousDrivingWithLargeLanguageModels,”IEEETransactionsonIntelligentVehicles,vol.9,no.1,pp.1450–1464,Jan.2024,doi:10.1109/TIV.2023.3327715.15Chen,Guojun,XiaojingYu,andLinZhong."TypeFly:FlyingDroneswithLargeLanguageModel."arXivpreprintarXiv:2312.14950(2023)./pdf/2312.1495016Vemprala,Sai,etal."Chatgptforrobotics:Designprinciplesandmodelabilities."arXivpreprintarXiv:2306.17582(2023)./abs/2306.1758217O.X.-E.Collaborationetal.,“OpenX-Embodiment:RoboticLearningDatasetsandRT-XModels.”arXiv,Dec.17,2023.doi:10.48550/arXiv.2310.08864.EarlyLLMs,likeGPT3.5,weretrainedonlargebodiesoftextscrapedfromtheInternet.Thesemodelslackedreal-worldexperiencethatcouldreflecthowvariousconfigurationsofrobotsandotherautonomoussystemsmakeandexecutedecisions.Researchintoroboticaffordancesexploreshowtoconstraineachrobotmodeltoactionsthatarefeasibleandappropriatefortheircapabilities.18ThisprovidesaframeworkforguidingLLMdevelopmentbasedonmorecompleteknowledgeofanoperationorprocedure.Atthesametime,thegroundingfunctiontranslatesthishigh-levelknowledgeintoexecutionbyaparticularrobotmodelinaspecifictargetenvironment.PredictivemaintenanceOpportunity:Predictthependingbreakdownofdronecomponentstooptimizemaintenance,repair,andpart-replacementschedules.Properlyrecordedandanalyzed,thedrone’ssensorandoperationaldatarevealimpendingmechanicalproblemsbeforebreakdownsoccur.Predictivealgorithmsletmaintenanceandrepaircrewsestablishregularschedules,prioritizemaintenance,andstayaheadofpartsinventories.Withadvancenoticeandplanning,evenmajorrepairsandreplacementscanbeperformedduringroutineservice.Theprobabilitiesofcostlybreakdownsand,worse,catastrophicfailuresdropsharply.Partscanbereplacedjust-before-needed,withservicelifecalculatedasafunctionofinstalled-partquality,servicetime,andoperationalprofile—slashingthecostsofreplacingperfectlysoundpartsonafixedschedule.TraditionalMLalgorithmsoftenlieattheheartofpredictivemaintenance.Forexample,metricslikeRemainingUsefulLife(RUL)andHealthIndicatorscanidentifymotoranomalies.ButsyntheticdatageneratedbyGANsandotherGenAIalgorithmscanimprovethealgorithms’performance.MultipleMLtechniques,includingGenAIalgorithms,canbecombinedtoimprovefaultdiagnosisandpredictivemaintenanceworkflows.19Forexample,GANtechniqueshavebeenappliedtoacousticsignalsfrommachinerytoidentifyandpredictfaultsnotidentifiedbyothermethods.20GANshavealsobeenusedtogeneratesyntheticmonitoring-datasetstohelptrainotherMLalgorithmstoimprovefailure-predictionandoptimizemaintenanceschedules.21GAN-FP,geneticadversarialnetworksforfailureprediction,specializeingenerating,balancing,andlabelingtrainingdatatoimproveperformanceofotherMLalgorithms22.FaultmanagementOpportunity:Identifyfaults,makedynamicadjustments,andeffectasafelandingwhenrequired.18M.Ahnetal.,“DoAsICan,NotAsISay:GroundingLanguageinRoboticAffordances.”arXiv,Aug.16,2022.Accessed:Mar.21,2024.[Online].Available:/abs/2204.0169119Z.Mianetal.,“Aliteraturereviewoffaultdiagnosisbasedonensemblelearning,”EngineeringApplicationsofArtificialIntelligence,vol.127,p.107357,Jan.2024,doi:10.1016/j.engappai.2023.107357.20Y.Wang,A.VinogradovImprovingtheperformanceofconvolutionalGANusinghistory-stateensembleforunsupervisedearlyfaultdetectionwithacousticemissionsignalsAppl.Sci.,13(5)(2023),p.3136,10.3390/APP1305313621Q.Fu,H.Wang,J.Zhao,andX.Yan,“AMaintenance-predictionMethodforAircraftEnginesusingGenerativeAdversarialNetworks,”in2019IEEE5thInternationalConferenceonComputerandCommunications(ICCC),Dec.2019,pp.225–229.doi:10.1109/ICCC47050.2019.9064184.22S.Zheng,A.Farahat,andC.Gupta,“GenerativeAdversarialNetworksforFailurePrediction.”arXiv,Oct.04,2019.Accessed:Mar.15,2024.[Online].Available:/abs/1910.02034GenAImodelscantransformdataforotherMLsystemstoimprovefaultdetectioninautonomoussystems.Forexample,VAEscanhelpcompressoperationaldataintomoreefficientrepresentationsforlongshort-termmemorynetworks(LSTN),atypeofrecurrentneuralnetwork.23Spatio-temporaltransformernetworkscancapturetrendsanddimensionsacrossdifferenttimescalestoimprovebatteryfaultdiagnosisandfailureprognosis,enhancingpredictivemaintenanceforUAVs.Forexample,theBERTerysystemcanspotsubtlechanges(changesinvisibletoearlierMLtechniques)thatsignalimpendingbatteryfailureasmuchas24hoursbeforethebatteriesfail.24GANshavebeenusedtogeneratetrainingsamplesandbuildinferencenetworksforaircraft-enginemonitoringdatatoimprovefailurepredictionsofotherMLalgorithms.25ResearchershavecombinedVAEsandLSTMtosupportcontinuouslearningfromvehiclesensordata,generatingsyntheticdataforwiderrangesoffaultscenarios.BytrainingotherMLalgorithmsonthiskindofsyntheticdata,Sadhuetal.achieved90%accuracyindetectingfaultsand99%accuracyinclassifyingthem.Demandsforcomputingpowerandrelativelyslowexecutionspeedaretopconcernswhenrunningthesekindsofalgorithmsonlow-costhardware.OnesolutionistoportthecomputationstoFPGAs,whicharemorepower-efficientthanGPUs.ThisishowSadhuetal.achieveda40xspeedup(athalfthepowerconsumption)fortheirVAE-LSTMfaultdetectionalgorithm.26VAEscanalsobeusedtotrainmodelsthatidentifynormaloperation.Usingthistechnique,Dhakletal.achieveda95.6%accuracyindetectingdeviationsindicativeoffaultsandanomaliesnotrepresentedinthetrainingdataset.27Whenaproblemarisesinadroneoritscommunicationsnetwork,thedronemustlandsafelytoavoidsecondarydamage.Tominimizethisrisk,MonteCarloalgorithmshavebeenusedtocalculate“targetlevelsofsafety”(levelsofacceptablerisk)forvariouslandingzones.28Techniqueslikethiscouldbecombinedwithtransformerstomakecontext-awaredecisionswhenafaultforcesaUAVsystemtoselectanappropriatelandingzone.Inthefuture,itmayalsobepossibletouseGenAItechniquesliketransformerstoletsystemsself-healinresponsetohardwarefailures,softwarebugs,ornetworkdisruption.Forexample,Khlaisamniangetal.haveproposedaframeworkforusingGenAItodetectanomalies,generatecode,debugit,andcreatereportsoncomputersystems.29Althoughstillinitsearlystages,thisworksuggestsdirectionsforfutureresearchonotherautonomoussystems.23P.Han,A.L.Ellefsen,G.Li,F.T.Holmeset,andH.Zhang,“FaultDetectionWithLSTM-BasedVariationalAutoencoderforMaritimeComponents,”IEEESensorsJ.,vol.21,no.19,pp.21903–21912,Oct.2021,doi:10.1109/JSEN.2021.3105226.24J.Zhao,X.Feng,J.Wang,Y.Lian,M.Ouyang,andA.F.Burke,"Batteryfaultdiagnosisandfailureprognosisforelectricvehiclesusingspatio-temporaltransformernetworks,"AppliedEnergy,vol.352,p.121949,2023.25Q.Fu,H.Wang,J.Zhao,andX.Yan,“AMaintenance-predictionMethodforAircraftEnginesusingGenerativeAdversarialNetworks,”in2019IEEE5thInternationalConferenceonComputerandCommunications(ICCC),Dec.2019,pp.225–229.doi:10.1109/ICCC47050.2019.9064184.26V.Sadhu,K.Anjum,andD.Pompili,“On-BoardDeep-Learning-BasedUnmannedAerialVehicleFaultCauseDetectionandClassificationviaFPGAs,”IEEETransactionsonRobotics,vol.39,no.4,pp.3319–3331,Aug.2023,doi:10.1109/TRO.2023.3269380.27R.Dhakal,C.Bosma,P.Chaudhary,andL.N.Kandel,"UAVfaultandanomalydetectionusingautoencoders,"inProceedingsofIEEE/AIAA42ndDigitalAvionicsSystemsConference.IEEE,2023,pp.1-8.28L.Tong,X.Gan,L.Yu,andH.Zhang,“EvaluationofSafetyTargetLevelofUnmannedAerialVehicleSysteminFusionAirspace,”in2022IEEEInternationalConferenceonArtificialIntelligenceandComputerApplications(ICAICA),Jun.2022,pp.375–379.doi:10.1109/ICAICA54878.2022.9844489.29P.Khlaisamniang,P.Khomduean,K.Saetan,andS.Wonglapsuwan,GenerativeAIforSelf-HealingSystems.2023,p.6.doi:10.1109/iSAI-NLP60301.2023.10354608.FleetGenAIcanhelpimproveswarmintelligence,swarmcoordination,andtherobustness,security,andefficiencyofunderlyingcommunicationsnetworksattheleveloffleetsorswarmsofautonomousthings.Inthiscontext,ZeroTrustsecurity,safety,andresiliencecomeintoplay—protectingdronefleets,improvingtheintegrityofsharedsensing,facilitatingbetterc