




版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
EnhancingSecurity,Resilience,and
SafetyofSystemswith
GenerativeAITableofSidebar:GenAImodels4GenAIUseCases9IndividualApplications9Self-Awareness10Anomalydetection11Autonomy12Adaptability12Predictivemaintenance13Faultmanagement14Fleet16Swarmintelligence17Swarmcoordination19Communicationresilience21Consensus22Human23Missionplanningandexecution24Human-machineteaming25Meta-learning26Datalabelingsynthesis27Cybersecurity28Malware29Intrusiondetection:31Integratingthreatintelligence33Policymanagement33ThreatSimulation:34Softwaresupplychainvisibility34Challenges34Cost35Compute36Adaptation36Ethical/Regulatory36Alignment36Privacy36Accuracy3737securityissues3838Glossary41Rapidadvancesautonomoussystemsedgeroboticsunlockedunprecedented
opportunitiesinindustriesfrommanufacturingtransportationhealthcareandexploration.
Increasingcomplexityandconnectivityhaveintroducednewchallengesinensuringsecurity,
resilience,andsafety.Asedgerobotsintegrateourdailylivesandcriticalinfrastructures,itis
imperativethatwedevelopinnovativethatcanthesetrustworthiness
andreliabilitynewlevels.ThiswhitepaperexploresthetransformativepotentialofgenerativeAI(GenAI)enhancethe
security,resilience,andsafetyofautonomoussystemsandedgerobots.canusethese
cutting-edgetechnologiesmeettheuniquedistributeddynamicchallengesofedge
robotics,unlocklevelsofintelligence,adaptability,androbustness.GenAImodelsproducenewcontentbyanalyzingpatternsinadataset.Theyderivecharacteristic
probabilitydistributionsandapplythesecreatenewdatapatternsthatareconsistentwiththe
original“real”dataset.EarliergenerationsofdiscriminativeAImodelsappliedconditionalprobabilitiesoutcomesforpreviouslyunseendata.Theapproachisversatileandwell-suitedawiderange
ofproblems,includingclassificationsandregressions.Theyexcelatdelineatingthedecision
boundariesthatdifferentiatebetweenvariousclassesorcategorieswithinthedataset.Thegrowingranksofgenerativetechniquesincludethosebasedontransformersandthe
resultinglargelanguagemodels(LLMs),GenerativeAdversarialNetworks(GANs),Autoencoders(VAEs),GenerativeFlowModelsandGenerativeDiffusionModels(GDM).
ThesehaveallexcitingavenuesofAIresearch—withapplicationsdroneswarms
andintrusiondetection,physicalsecurity,semanticcommunication,andnetwo1TheTechnologyInnovationInstitute’sSecureResearchCenter(TII-SSRC,AbuUAE,https://www.tii.ae/secure-systems)isapplyGenAIitsworkextendingZero
Trustarchitectures(developedforinformationsecurity)everyaspectofinformationsecurityin
cyber-physicalsystems.Thus,SSRCconsidershowGenAIcanhelpguaranteesecurity,resilience,andforswarms,swarmsofswarms,autonomousterrestrialmarinevehicles,commandsystems,human/droneinteractionsandparticularlyinareaswhereGenAIoutperformstraditionalapproaches.Examples•Individualapplications:healthmonitoring,stateestimation,predictivemaintenance,
anomalydetection,self-healing,navigation,andemergencylandings.
•Fleetapplications:Swarmcoordination,intelligence,collectivedecision-making.
•Human/Droneinteractions:communicationresilience,missionplanning,human-
computerinteraction.•Cybersecurityandresilience:Intrusiondetection,malwareclassification,threatsimulation.ThispaperwillfocusondronesSSRCisdoingsomuchworkinthisarea.Whatwe
learnfromdronescanbeappliedmoreautonomousandcyber-physicalsystemsin
general,includingcars,robots,embeddedandsmartcities.Bythesametoken,
lessonslearnedintheseareascanbefoldedSSRC’sresearch.approachallowsorganizationsmoveawayfromphysicaldevicemanagementapproaches
thatrequireemployeescarrymultiplephones.1M.“UnleashingthePowerEdge-CloudGenerativeinMobileNetworks:AAIGCServices.”arXiv,2023.doi:10.48550/arXiv.2303.16129Sidebar:GenAImodelsAsurveyofspecificgenerativeAImodelingtechniques,withtheirstrengthslimitationsas
applieddronesafety,security,andresilience.PopularexcitementovergenerativeAImodelsbeendrivenbyhighlypublicizedlikeChatGPT,whichcreateauthoritative-soundingresponsestextpromptsusingspecially
trainedLLMs.LargeLanguageareAIsystemstrainedonvasttextdatasets.TheyuseDeepLearning
techniques,particularlyastructureknownasaTransformer,“understand”andhuman-liketextbasedonthepatternslearned.LLMsanalyzetherelationshipscontextsinthedata.Theyuseavarietyoftechniquesbuildsimplifiedrepresentations
ofthedata,allowingthemmakeassociationsandcorrelationsbetweenoriginalelements.Thesemodelsletthemcomposeresponsesthatcanmimichumanwritingstylescoverdiversetopics.VisuallikeDALL-EandStableDiffusion,compileimagesfromtext
andimageprompts.LLMsarewidelyused,widelyuseful,increatingcontent,code,translations,summaries,
syntheticdata,structureunstructuredfromtexts,documents,images,andpromptdata.Transformermodelsandtheservicesbuiltonthem—likeOpenAI’sGoogle’sGemini,
andAnthropic’sClaude—attractedwidespreadattention,thankstheirimpressiveability
createarticulate-seemingresponseshumanprompts.These,andotherdomain-specific
LLMsandSmallLanguageModels(SMLs),showpromiseforsupportinganalysis,research,
anddevelopmentimprovedronesafety,securityandresilience.ParallelingtheseveryvisibleAIdevelopments,however,hasbeenalmostadecadeofprogress
onnewclassesofGenerativeAImodelscouldautomateandrepresentation-building.
Whilegenerativeapplicationsattractthemostattentions,thesenewmodelsaredriving
advancesinanalyzingdataandinteractingwiththeworldus.OthergenerativeAI
models—GenerativeAdversarialNetworks,VariationalAutoencoders,GenerativeDiffusion
Models,NormalizingFlowModels—thoughrelativelyunknown,makesubstantial
contributionsdronesecurity,safety,andresilience.TransformerModels:Introducedin2017translatebetweenandtexts,
TransformerModelsatcapturinglong-rangedependenciescorrelationswithin
unstructureddata2Transformersleverageanovel“attentionmechanism”learnthe
connectionsbetweenwordshelpcreateembeddingsautomatically.Priortechniquesrequired
translatingrawtextintoavectorrepresentationusingaseparatemodel.Transformerscanbuild
complexrepresentationsandlearnintricateconnectionsthroughlayeredarchitecture,
manner.,allowingresearchersprocesslargebodiesofunlabeledtextdeveloplarge
languagemodelsbillionsofparameters.Subsequentinnovationssupporteddocument
summarization,composingquestion/answerassociationsacrosslargedatasets,code
generation,in-depthanalysis,intrusiondetection,malwaredetection,translatingcontrol
systeminstructionsacrossroboticarms.Thekeyadvantageisdistillingcontextfrom
complexdatasets.Challengesincludehallucination,longertrainingtime,slowerinference-
building,heaviercomputationrequirements,andlargermodelcomparedother
techniques.2Vaswani“AttentionAllNeed.”arXiv,Aug.2023.doi:10.48550/arXiv.1706.03762.GenerativeAdversarialNetworks(GANs):Theseweredevelopedin2014createrealistic
syntheticnumbers,faces,andanimalimage3GANspittwonetworksagainstother:
oneisrewardedforgeneratingrealisticcontent,andthesecondisrewardedfordetecting
fakecontent.thiscompetition,thegeneratorimprovesitscreateoutputscanfoolthediscriminator.GANsarewidelyusedincontentgeneration.Sincethefirstversions
weredesignedworkwithimages,researchersarenowfindingcreativewaystranslatesuchascodeornetworklogs,intoimagessuitableforprocessing.GANsaregoodfor
realisticsyntheticdatasetsthatcanbeusedimproveautonomoussystemsandcybersecurity
algorithms.They,too,however,sufferfromfailureslikecollapseorcatastrophic3J.Goodfellow“GenerativeAdversarialNetworks.”arXiv,Jun.2014.doi:10.48550/arXiv.1406.2661.VariationalAutoencoder(VAE):VAEswereintroducedin2014improveinferencesdrawn
fromacontinuouslyvaryingdatastream.4Thetechniquehelpsfindefficientwaysrepresent
dataandbecompressdataordetectanomaliesthreats.TrainingVAEsprocess
involvesteachingasetofencodersanddecoderstranslaterawdataintoanintermediatelatent
spacewithadifferentprobabilitydistribution.canbeusedindependentlyapplicationsanomalydetection,designingencodingschemes,dataaugmentation,generation.addition,theyareoftenusedpre-structuredataforotheralgorithms,GANs,improvetheirresults.GenerativeDiffusionModel(GDF):emergedin2015improvelearning,sampling,
inferences,andevaluationsthatwereinformedbynon-equilibriumthermodynamicsmodeling5
Thetechniqueaddsnoiseasampleanimage)andautomatesthedenoising
processrevealthedata’sstructure.Slightvariationsleaddatasets.arewidelyusedinimagegenerationandcanimprovesignalclassificationvariousdroneusecases.However,thetechniquerequireshighersamplingdemands
amorecomplexarchitectureGANsandNormalizingFlow(NFMs):Thesewereintroducedbyresearchersmakecomplex
datasimplerworkwith6Thesemodelstakeeasy-to-understanddistributions,likeanormalbell
curve,andtransformthemstepbystep.Eachstepisreversible,meaningwealwaysgothestartifneeded.Thisprocess,calleda“”movesfromasimplebeginninganend4D.andM.Welling,“Auto-EncodingVariationalBayes.”arXiv,Dec.2022.10.48550/arXiv.1312.6114.5Ling,et"Diffusionmodels:Acomprehensivemethodsandapplications."ACMComputingSurveys56.4(2023):1-39./doi/10.1145/3626235
6Ivan,SimonJDPrince,MarcusA.Brubaker."Normalizingflows:introductionandreviewcurrentmethods."IEEEtransactionsonpatternanalysisand
intelligence43.11(2020):3964-3979./abstract/document/9089305/authorsresemblesthecomplicatedtargetdataset.Bydoingthis,itispossiblestudyandusethedata
moreeffectively.NFMshavebeenusedgeneratehandwrittennumbers,images,etc.Newer
usecasesincludeenhancedclassificationencodingschemes.Thetrainingprocessamodelthattransformstheprobabilitydistributionofadatasetacomplex,fully
reversibledistribution.NFMscan,however,requirehighercomputationandtrainingtimesthan
techniqueslikeGANsVAEs.ThefollowingfiguresummarizesthemainTechniquestheirapplicationsthefield
ofZeroTrustforautonomoussystems.GenAIUseTheproliferationofdrone-technologyhasbroughtchallengesspanbetweendomains
—individual,fleet,humancontrol,cybersecurity.Theirgrowthandcomplexitydemand
constantinnovationredoubletheirtrustworthinessandreliability.Thefollowingapplications—whetherderivedfromUnmannedAerialVehicle(UAV)anddrone
researchorimportedfromotherdomains—importantimplicationsforthefutureofUAVsand
otherautonomoussystems.too,thatmanyoftheseareearly-stageprojects,includedgiveaflavorwhatGenAItoolsmightaccomplishastechniquesevolve.本报告来源于三个皮匠报告站(),由用户Id:673421下载,文档Id:490694,下载日期:2025-01-23GenAIshowstremendouspotentialforimprovingZeroTrustframeworksenhancesecurity,
resilience,andsafetyinindividualautonomoussystems,suchasdrones,self-drivingcars,
robots,andembeddedsystems.Usecasesunderinvestigationincludeboostingself-
awareness,anomalydetection,autonomousdriving,predictivemaintenance,faultmanagement,
self-healing,andlandingsafely.Self-AwarenessOpportunity:Efficientlytranslatenoisy,blurry,andinconsistentdataunderstandthedrone's
currentstate—e.g.,compensatingformotionblurwhiletryingdetectobstacles.
Thefoundationofdronehealthisaccuratelycapturingandofitscurrentstate—
includingtheconditionofitscurrenthardware,itsapplications,itsphysicallocation,securityposture.theworld,thiscanmessy,asvideofeedsmotionblur,datajitters,inertialguidancelosecalibration,andnoiseorgapsdegradeinternaldata.Stateestimationiscrucialautonomousnavigationanddecision-making,andrawdatastreams
mustbeaccuratelycorrelatedwithposition,velocity,andorientation.7GenerativeAIcanhelpfill
inmissingdataandfusefrommultiplesourcesimprovestateestimation.8InnovationsGenAIalgorithmslikeGANs,VAEs,andtraditionalMLalgorithmslikeLSTMfillingaps,preservingvehiclefaultdetection,predictivemaintenance,managementandsafe-landingprotocols.Forexample,innovativeapproachescanmissingandmakeiteasierfusestreamscreateamoreaccuratestate
assessment9helpcorrelatedatawithacousticanalysis,10andidentifymechanicalissue11Researcherstechniquesforgeneratingestimated-
statevariablesusingConditionalGANsforindividualdronesdroneswarms.127T.D.Barfoot,Stateestimationforrobotics.CambridgeUniversityPress,2017.8Guangyuan,NguyenVanHuynh,HongyangDinhThaiHoang,DusitNiyato,KunZhu,JiawenKang,ZehuiJamalipour,andDongInKim.“GenerativeUnmannedVehicleSwarms:Challenges,ApplicationsandOpportunities.”arXiv,February28,/10.48550/arXiv.2402.18062.
9Chai,andZ.approachstateestimationgenerativeadversarialnetwork,"inIEEEInternationalConferenceonMan(SMC),2019,2248-/document/891458510Wang,Vinogradov,“ImprovingtheperformanceconvolutionalhistoryensembleforunsupervisedearlydetectionwithacousticemissionsignalsSci.,(5)(2023),p.3136,10.3390/APP1305313611Zheng,Farahat,andC.Gupta,“GenerativeAdversarialNetworksforFailurePrediction.”arXiv,04,2019.Accessed:Mar.15,2024.[Online].Available:/abs/1910.0203412He,C.X.Tian,andW.Zeng,"AtwofoldSiamesenetworkforreal-timetracking,"inProceedingstheIEEEconferenceoncomputervisionpattern
2018,pp.4834-4843./document/8578606AnomalydetectionOpportunity:Improveanalysisofdronesensordataidentifyabnormalconditions.
Moreaccurate,multi-dimensionalsystem-staterecordscanalsohelpidentifyanomaliesrelevant
dronehealth.Forexample,VAEscanimprovefaultdetectionisolation.Theycanidentifythewarningsignsofstressinsystems,prioritizepredictiveschedules.Typically,machine-learningclassificationalgorithmstrainedonclasses
(suchastagged“faulty”faulty”).onclasses,suchasdata,isoften
scarceinpublicdatasetsandtherealworld.thesecases,GenerativeAdversarialNetworks
(GANs)canbevaluablesynthesizingtheseclasses—making“faulty”datathatlooks
conditions.addition,researchersareexploringhowcouldhelpbettercontextualizeoperatenewenvironments.13Forexample,DriveLLMcombinesLLMtraditionalautonomousnavigationalgorithmssupportbetterreasoningdecision-makingwhenrespondingedgecases.14Researchers
foundthemethodcouldimproveproactivedecision-makingunexpectedcircumstances.
Anotherapplication,TypeFly,enhancescommunicationbetweenhumansanddronesthrougha
naturallanguageinterface15.Suchlargelanguagemodelsmay,however,manifestsurfacebias,inaccuracies,and
hallucinationissuesthatrequireadditionalsafeguards.Similarly,MicrosoftResearchdiscusses
theiradvancementsinintegratingwithroboticsmakerobotcontrolintuitive
throughnaturallanguage.They'veenabledunderstandexecutetasksin
physicalenvironments,whichfacilitateseasierhuman-robotinteractionwithouttheneedfor
complexprogrammingknowledge.Theteamhasdevelopeddesignprinciplesfor
languagemodelsroboticstasks(involvingspecialpromptingstructureshigh-level
APIs),andtheyhavedemonstratedhowcanhandletasksoperatingdronesrobotarmsthroughuser-friendlycommandsfeedback.Thedevelopersemphasizethe
importanceofsafetyandsimulationtestingbeforereal-worldapplication16AdaptabilityOpportunity:Improvetranslationofautonomoussystemssoftwarerunacrossdifferenthardwaremakes,models,andconfigurations.Autonomoussystemcontrollersmustbetrainedforaspecificmodelconfiguration.Thiscan
createchallengeswhenupgradingindividualcomponentsoradoptingnewmodels.RTXisarobot
controlthattranslatecontrolpoliciesmanagedifferentroboticarmswithoutthecontrolalgorithmsforthelatesthardware.sometests,leveragingtheexperienceofcontrollers,producedcontrolpoliciestheoutperformedthebestcontrolscustom-builtforan
individuala1713Lei,surveyonlargelanguagemodelautonomousagents."FrontiersComputerScience18.6(2024):1-26./article/10.1007/s11704-024-40231-114“DriveLLM:ChartingthePathTowardFullAutonomousDrivingWithLargeLanguageModels,”IEEETransactionsIntelligentVehicles,vol.pp.1450–Jan.2024,doi:10.1109/TIV.2023.3327715.15Chen,Guojun,XiaojingYu,Zhong."TypeFly:FlyingwithLargeLanguageModel."arXivpreprintarXiv:2312.14950(2023)./pdf/2312.14950
16Vemprala,Sai,et"Chatgptforrobotics:Designandmodelabilities."arXivarXiv:2306.17582(2023)./abs/2306.17582
17X.-Collaborational.,“OpenX-Embodiment:RoboticDatasetsandRT-XarXiv,Dec.17,2023.10.48550/arXiv.2310.08864.EarlyLLMs,likeGPT3.5,weretrainedonlargebodiesoftextscrapedfromtheInternet.These
modelslackedreal-worldexperiencethatcouldreflecthowvariousconfigurationsofrobotsotherautonomoussystemsmakeandexecutedecisions.Researchintoroboticaffordances
exploreshowconstrainrobotmodelactionsthatarefeasibleandappropriateforcapabilities.18ThisprovidesaframeworkforguidingLLMdevelopmentbasedonmorecomplete
knowledgeofanoperationorprocedure.Atthesametime,thefunctiontranslateshigh-levelknowledgeintoexecutionbyaparticularrobotmodelinaspecifictargetenvironment.PredictivemaintenanceOpportunity:Predictthependingbreakdownofdronecomponentsoptimizemaintenance,repair,part-replacementschedules.Properlyandanalyzed,thedrone’ssensoroperationaldatarevealmechanicalproblemsbeforebreakdownsoccur.Predictivealgorithmsmaintenanceandrepair
crewsestablishregularschedules,prioritizemaintenance,andstayaheadofpartsinventories.
advancenoticeandplanning,evenmajorrepairsandreplacementscanbeperformedduring
routineservice.Theprobabilitiesofcostlybreakdownsand,worse,catastrophicfailuressharply.Partscanbereplacedjust-before-needed,withservicelifecalculatedasafunctioninstalled-partquality,servicetime,andoperationalprofile—slashingthecostsofreplacing
perfectlysoundonafixedschedule.TraditionalMLalgorithmsoftenlieattheheartofpredictivemaintenance.Forexample,metrics
likeRemainingUsefulLife(RUL)HealthIndicatorscanidentifymotoranomalies.But
syntheticdatageneratedbyGANsandotherGenAIalgorithmscanimprovethealgorithms’
performance.MultipleMLtechniques,includingGenAIalgorithms,canbecombinedimprovefaultandpredictivemaintenanceworkflow19Forexample,techniquesbeenappliedacousticsignalsfrommachineryidentifyandpredictfaultsnotidentifiedbyothermethods.20
GANshavealsobeengeneratesyntheticmonitoring-datasetshelpotheralgorithmsimprovefailure-predictionandoptimizemaintenanceschedules.21GAN-FP,genetic
adversarialnetworksforfailureprediction,specializeingenerating,balancing,andtrainingdataimproveperformanceofotherMLalgorithms22.FaultmanagementOpportunity:Identifyfaults,makedynamicadjustments,andeffectasafelandingwhenrequired.18M.Ahn“DoICan,ISay:GroundingLanguageinRoboticAffordances.”arXiv,16,2022.Accessed:Mar.22024.[Online].Available:/abs/2204.0169119Z.Mianal.,“Aliteraturereviewoffaultensemblelearning,”EngineeringApplicationsofArtificialIntelligence,vol.Jan.2024,10.1016/j.engappai.2023.107357.Wang,VinogradovImprovingperformanceofconvolutionalhistory-stateensembleunsupervisedearlyfaultdetectionwithacousticemissionsignals
Sci.,(5)(2023),p.3136,10.3390/APP13053136H.Wang,J.Zhao,andX.“AMaintenance-predictionMethodforAircraftusingGenerativeAdversarialNetworks,”inIEEE5thInternational
onComputerandCommunications(ICCC),Dec.2019,pp.225–doi:10.1109/ICCC47050.2019.9064184.
Zheng,Farahat,andC.Gupta,“GenerativeAdversarialNetworksforFailurePrediction.”arXiv,04,2019.Accessed:Mar.15,2024.[Online].Available:
/abs/1910.02034GenAImodelscantransformdataforotherMLimprovefaultdetectioninautonomous
systems.Forexample,VAEscanhelpcompressoperationaldatarepresentationsforlongshort-termmemorynetworks(LSTN),atypeofrecurrentneural
netwo23Spatio-temporaltransformercancapturetrendsanddimensionsacross
differenttimescalesimprovebatterydiagnosisfailureprognosis,enhancingpredictive
maintenanceforUAVs.Forexample,thesystemcanspotsubtlechangesinvisibleearlierMLtechniques)thatsignalimpendingbatteryfailureasmuchas24hours
beforethebatteries.24GANshavebeenusedgeneratetrainingsamplesandbuildinferencenetworksforaircraft-
enginemonitoringimprovefailurepredictionsofotherMLalgorithms.25
ResearchershavecombinedVAEsandLSTMsupportcontinuouslearningfromvehiclesensor
data,generatingsyntheticdataforwiderrangesoffaultscenarios.BytrainingotherMLalgorithms
onthisofsyntheticdata,Sadhuetachievedaccuracyindetectingfaultsaccuracyinclassifyingthem.Demandsforcomputingpowerrelativelyslowexecutionspeedaretopconcernswhen
runningthesekindsofalgorithmsonlow-costhardware.Oneisportthecomputations
whicharemorepower-efficientthanGPUs.ThisishowSadhuetal.achievedaspeedup(athalfthepowerconsumption)forVAE-LSTMfaultdetectionalgorith26
VAEscanalsobeusedtrainmodelsthatidentifynormaloperation.Usingthistechnique,Dhakl
etal.achieveda95.6%accuracyindetectingdeviationsindicativeoffaultsandanomaliesrepresentedinthetrainingset.27Whenaproblemarisesinadroneoritscommunicationsnetwork,thedronemustlandsafelyavoidsecondarydamage.thisMonteCarloalgorithmshavebeencalculate“targetlevelsofsafety”(levelsofacceptablerisk)forvariouslandingzone28
Techniqueslikethiscouldbecombinedwithtransformersmakecontext-awaredecisionswhen
afaultforcesaUAVsystemselectanappropriatelandingzone.thefuture,itmayalsobepossibleGenAItechniquestransformersletself-healinresponsehardwarefailures,softwarebugs,ornetworkdisruption.Forexample,
Khlaisamniangetal.proposedaforusingdetectanomalies,generate
code,debugit,andcreatereportsoncomputersystems.29Althoughstillinitsearlystages,this
worksuggestsdirectionsforfutureresearchonotherautonomoussystems.23Han,A.L.Ellefsen,F.T.Holmeset,H.Zhang,“FaultDetectionWithLSTM-BasedVariationalAutoencoderMaritimeComponents,”IEEESensorsvol.19,pp.2190321912,2021,doi:10.1109/JSEN.2021.3105226.24J.Zhao,Feng,J.Wang,Lian,M.Ouyang,andF.Burke,"Batteryfaultdiagnosisandfailureprognosisforelectricvehiclesspatio-temporaltransformernetworks,"
AppliedEnergy,vol.25H.Wang,J.Zhao,andX.“AMaintenance-predictionMethodforAircraftusingGenerativeAdversarialNetworks,”inIEEE5thInternational
onComputerandCommunications(ICCC),Dec.2019,pp.225–doi:10.1109/ICCC47050.2019.9064184.26Sadhu,andD.Pompili,“On-BoardDeep-Learning-BasedUnmannedAerialFaultCauseDetectionandClassificationviaFPGAs,”IEEETransactionsvol.no.4,pp.33193331,2023,doi:10.1109/TRO.2023.3269380.27R.Dhakal,C.Bosma,Chaudhary,andN.Kandel,"UAVfaultandanomalydetectionusingautoencoders,"inProceedingsIEEE/AIAADigitalAvionicsSystems
Conference.IEEE,2023,pp.1-8.28Tong,Gan,L.Yu,andH.Zhang,“EvaluationTargetofUnmannedVehicleinFusionAirspace,”inIEEEInternationalConferenceArtificialIntelligenceandComputerApplications(ICAICA),Jun.pp.37510.1109/ICAICA54878.2022.9844489.29Khlaisamniang,P.Khomduean,Saetan,Wonglapsuwan,GenerativeAIforSelf-HealingSystems.2023,10.1109/iSAI-NLP60301.2023.10354608.FleetGenAIcanhelpimproveswarmintelligence,swarmcoordination,andtherobustness,security,
andefficiencyofunderlyingcommunicationsnetworksattheleveloforswarmsof
autonomousthings.thiscontext,ZeroTrustsecurity,safety,andresilienceintoplay—
protectingdronefleets,improvingtheintegritysharedsensing,facilitatingbettercoordination,
andreducingtheimpactofacompromiseddroneonthefleetasawhole.SwarmintelligenceOpportunity:Improvetrustworthyfusionofsensordatafrommultipleindividualsinaswarm.
Overthepastdecade,researcherspublishedtensofthousandsofpapersonsynthesizing
unifiedsituationalviewsbyfusingfromnumerousdronestogether.Supposethata
dronefleetissurveyingalarge-scaledisasterafloodorafire.Ideally,trustworthyswarm
intelligencecaninformationfrommemberoftheswarmpaintacomprehensive
pictureofthesituation.GenerativeKnowledge-SupportedTransformers(GKSTs),forexample,canfuseimageryfrom
differentviewsofatargetobject,producingmoremeaningfulimagesfromamovingvehicle.30
Furtherenhancementsofthismulti-viewapproachmightimprovetheinterpretationofcollectedfromthedifferingperspectivesofmembers.importantappreciatethattheremanywaysofrepresentingtheworld,different
approachesmaybesuitablefordifferentpurposes.Forexample,similarityclassification
algorithmscanhelptheobjectfeaturesinimagery.contrast,semantic
categorizationalgorithmslabeltheseasmembersofspecificcategoriesorclasses.
SA-SIAM(atwo-foldSemantic-Appearanceneuralnetwork)sharesinformationacross30Yu,W.Liao,C.Qu,Q.andZ.Xu,“UAVCooperativeSearchMulti-agentGenerativeAdversarialImitationin2022InternationalConferenceon
Learning,CloudComputingandIntelligentMining(MLCCIM),Aug.pp.441446.doi:10.1109/MLCCIM55934.2022.00081.separateneuralnetworks,onetrainedonSemanticinformation,theotheronAppearancedata.31
TheSemanticsideusesattentionmechanismhelpinterpretdatabasedonthetargetadditionalcontextualinformation.Althoughthisnotafull-generativeAIimplementation,itshows
howamoretargetedattentionmechanismintransformerscanbeappliedotherusecasesina
moretargetedway,whichmaybeefficientthanafull-blownLLMimplementation.ConditionalGANs(CGANs)—variantsofGenerativeAdversarialNetworksbasedspecific
conditions—beenusedformotionprediction.Thesepredictionsconsidereachobject's
relativemotionitschangingorientationrelativeaUAV.32Theseworkinconjunctiona
Siamesenetwork,inwhichneuralaresharedacrossapairofcomplementarynetworks.Thoughtheinitialresearchfocusedonindividualdrones,theworksuggestsafuture
pathforsynthesizing3Dviewsofsituationscontributionsacrossafleet.2016,researchersexploreda“socialpooling”layercouldhelpautonomousagentsmodel
theinteractionsofpeopleproximity,usingseparateLSTMnetworkspredictperson’s
motion33thiscase,theresearcherswerelookingatanautonomouscouldplanitspaththroughgroupsofindependentlymovinghumanbeings.Futureresearchexplorehowsocialpoolingcouldextendimprovemodelsthatallowdronesunderstand
currentlocationsandpredictfuturepositionsofnearbydrones,bystanderdrones,andoutsideradversarydrones.Text-to-image-baseddiffusionmodelsalsobeenusedgeneraterealisticimagesofUAVs
invaryingscenarios,improvingalgorithmsfordetectingUAVsby12%.Theresearcherscombined
normalizedmod
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 二零二五年度全新转让合同:网络直播平台运营权出让协议
- 2025年度汽车租赁代驾与网约车平台合作协议
- 二零二五年度亲友间代持房产购房协议
- 二零二五年度施工安全责任协议书(含风险评估)
- 二零二五年度实习生实习期间企业文化建设合作合同
- 2025年度购房合同解除及物业管理协议
- 妇幼保健员职业定位试题及答案
- 二零二五年度征收国有土地房屋拆迁安置合同
- 二零二五年度城市中心区三方合租公寓租赁协议
- 二零二五年度土地承包经营权权属变更合同
- 宋代农书研究出版对宋代农业研究的价值4篇
- 5.2《稻》教案-【中职专用】高二语文同步教学(高教版2023·拓展模块下册)
- 2025年超长期特别国债申报工作及成功案例
- 电梯困人培训课件
- 熔化焊接与热切割作业题库题库(1455道)
- 2025年中国中煤华东分公司招聘笔试参考题库含答案解析
- 铁路运输碳排放分析-洞察分析
- 第16课数据管理与编码(教案)四年级全一册信息技术人教版
- HPV分型检测介绍课件
- 超全自考英语二词汇表-含音标4500-个单词
- 外卖骑手交通安全课件
评论
0/150
提交评论