揭開人工智能在交通網(wǎng)絡(luò)安全中的神秘面紗（英文）

上傳人：策*** IP屬地：山西上傳時間：2024-12-29 格式：DOCX 頁數(shù)：128 大?。?.38MB 積分：19.9 舉報 版權(quán)申訴

已閱讀5頁，還剩123頁未讀，繼續(xù)免費(fèi)閱讀

版權(quán)說明：本文檔由用戶提供并上傳，收益歸屬內(nèi)容提供方，若內(nèi)容存在侵權(quán)，請進(jìn)行舉報或認(rèn)領(lǐng)

文檔簡介

DEMYSTIFYINGARTIFICIALINTELLIGENCEINTRANSPORTATIONCYBERSECURITY

URBANJONSON,SVPITANDCYBERSECURITYSERVICES,SERJON

URBANJONSON

ujonson@

Current

SVPInformationTechnologyandCybersecurityServices,SERJON,LLC

USFBIInfraGardTransportationSubjectMatterExpert

FBIAutomotiveSectorSpecificWorkingGroup(SSWG)

BoardofDirectors,CyberTruckChallenge

ProgramCommittee,ESCARUSA

SAEVehicleElectricalSystemSecurityCommitteeMember

Technology&MaintenanceCouncil(TMC)S.5andS.12StudyGroupMember

Experience

Over35yearsofexperienceinITandCybersecurity,includingstrategicplanning,assessments,projectmanagement,andprogrammanagement

Variouspapers,talks,andresearchonhacking,aswellasdefendingtrucksandtransportationingeneral

Abusinganddefendingsystemssincethe1980s

AGENDA

?AIOverview

?AITaxonomy

?ChallengeswithAI

?CommonAImistakes

?TransportationApplications

?AttackingAI

?DefendingAI

?Wrap-up

ImagegeneratedbyBingImageCreator

AIOVERVIEW

HISTORICALOVERVIEW

?1950s–BirthofAI

?AlanTuringandotherslaidthegroundworkformachineintelligence

?1960s–EarlyApplications

?Problem-solvingandsymbolicreasoning,e.g.playingchess

?1980s–ExpertSystems

?Programsdesignedtomimichumanexpertise,e.g.taxsoftware

?1990s–MachineLearningResurgence

?NeuralnetworksandnewalgorithmsrevitalizeinterestinAI

HISTORICALOVERVIEW

?2000s–RiseofBigData

?Availabilityoflargedatasetsandimprovedcomputingpower

?2010s–DeepLearningDominance

?Multi-layeredneuralnetworkscapableofimagerecognition,speechrecognition

?2020s–GenerativeAI

?NewLargeLanguageModels(LLM)builtonmassivedatacloudplatformscapableofgeneratingimages,code,andother

content(ChatGPT,BingImageCreator,etc.)basedoninputprompts

TECHNICALOVERVIEW

UNDERSTANDABLEVSPREDICTIVEPOWER

Image:NationalCyberSecurityCentre(UK)-Principlesforthesecurityofmachinelearning

[.uk/collection/machine-learning]

MACHINELEARNINGAPPLICATIONS

Commonmachine

learninganalytic

applications

Image:

/science/article/pii/S0951832021003835

ARTIFICIALNEURALNETS

Thecorecomponent

ofneuralnetsisthe

artificialneuron

Conceptuallycanbethoughtofasamini

linearregressionmodel

Image:

/blog/artificial-neural-networks-basics-guide/

TRAININGMETHODS

Image:

/science/article/pii/S0951832021003835

MODELTRAINING

Simplesupervisedlearning

Imagesareconvertedintonumericaldatausuallybyflatteningintoavector

Image:

/@MITIBMLab/estimating-information-flow-in-deep-neural-networks-b2a77bdda7a7

AITAXONOMY

OVERVIEW

NISTAIUseTaxonomy*:

?Decomposescomplexhuman-AItasksintoactivitiesthatareindependentoftechnologicaltechniques(e.g.,neural

network,largelanguagemodel,reinforcementlearning)anddomains(e.g.,finance,medicine,law).

?ProvidesaflexiblemeansofclassifyinganAIsystem’scontributiontoaspecifiedhuman-AItask.

?Intendedtobealivingdocumentthatisupdatedperiodicallywithfeedbackfromstakeholders,suchasthoseintheAI

evaluationandhumanfactorscommunities.

*NISTTrustworthyandResponsibleAINISTAI200-1,AIUseTaxonomy:AHuman-CenteredApproach,byTheofanos,Choong,andJenson,March2024,

/10.6028/NIST.AI.200-1

AITAXONOMY-TRANSPORTATION

?Connectionist–Learningalgorithmsbasedonneuralnetworks

?Bayesians–Probability-basedinferencesystems

?Symbolists–Logic-basedalgorithmssuchasrules-based

programming,decisiontrees,fuzzylogic,andrationalagents

?Analogizers–Similarity-basedclassifiers,suchassupportvectormachines

?Optimizations–Algorithmsperformingiterativeupdatesand

comparisonstodiscoveroptimumsolutions,e.g.GeneticAlgorithm(GA)

JonPerez-Cerrolaza,JaumeAbella,MarkusBorg,CarloDonzella,JesúsCerquides,FranciscoJ.Cazorla,CristoferEnglund,MarkusTauber,GeorgeNikolakopoulos,andJoseLuisFlores.

2024.ArtificialIntelligenceforSafety-CriticalSystemsinIndustrialandTransportationDomains:ASurvey.ACMComput.Surv.56,7,Article176(July2024),40pages.

/10.1145/3626314

CHALLENGESWITHAI

NON-DETERMINISTIC

?Thesameinputswillnotalwaysgeneratethesameoutputs

?Randomdataselectionbasedonprobabilitycurves

?Anytimeyouaddrandomdataselectioninconnectionistmodels,youruntheriskofnon-deterministicoutcomes

?Ifyourmodelcontinuestolearn,forexamplelinearregression,theoutputswillvaryovertimeasmodellearns

?Expertsystemsgenerallydonotsufferfromthisproblem,butanything

thathasaneuralnetworkwillruntherisk

INSCRUTABLE

?ThemathandcodeforAIiscompletelyunderstandable…..and“human-ish”readable

?Thedatathatyouusetotrainthemodelis

understandable(hopefully,ifyouhavedoneyourjobright)

?Theproblemiswhenyouusethecodetogenerateamodelbasedonthedata

?Duetohowthemodellearns(developingacomplexwebofprobabilisticweights)andisexpressed,itisnot

possibletolookatthemodelandunderstandhowitworks

UNEXPLAINABLE

?Sincethemodelisnon-deterministicandinscrutable,itisnoteasilyunderstood

?Makesexplaining“why”amodelproducedtheexactoutputexceedinglydifficultfor

neuralnetworks

?ExplainableandTrustworthyAIisanareaofintenseresearch

?TrustworthyAIcangenerateatrusted

explanationthathumanscanunderstand

WHYTHEPROBLEM?

?Layeredneuralnetworks

?Randomlygeneratedvalues

?Probabilisticevaluations

?Deeplearningisstatisticswithlinearalgebra

Image:

/tutorial/introduction-to-deep-neural-networks

DATAPROBLEM

?Ourmodelsareonlyasgoodasourdata

?Transportationdatasetsareintheirinfancy

?Wearestillinthegreat“dataownership”battle

?Ourvehicleplatformslackthesensorstocollectthenecessaryinformation(possible

exception…Tesla)

?Modelsareverylimitedinwhattheycando

TRANSPORTATIONDATASAMPLES

?Lackofinstrumentation

?Teslaprobablyhasbestdataset

?VehicleISasensorplatform

?Designedtocollectalldata

?Driversprovidingexperience

?Robottaxifleetdata2ndplace

?Cruise

?Waymo

?ProprietaryDataSources

?OEMdata

?Telematicsdata

?Vehicle/Fleetoperatordata

?OpenData

?PIVOT(

)

?EUDataAct

?ColoradoStateUniversity

TRAININGDATAPROBLEM

?DeeplearningandLLMsrequiremassivedatasetsforlearningandvalidation

?LLMs,suchasChatGPT,haveusedagreatdealofinternetcontent

?Manyimages,text,books,etc.usedinlearningmodelsarecopyrightedmaterials

?IsgeneratinganAImodelbasedonsomeoneelse’swork

a“fairuse”ofcopyrights?

?WhatiftheresultingAIismonetized?

?Howdoyouremoveonepartorsegmentoftrainingdataonceamodelhasbeencreated?

EDGECASES

?Edgecasesarestatisticaloutliereventsthatarenotpartofthetrainingdata

?Thoughtheymayberare,theycan

resultinunexpectedandundesirableoutcomes

?Edgecasesarewheretragedylives

Source:Projectguru.in

UBER

?UberAutonomousCrashMarch2018

?Pedestrianwalkingbicycleacrosstheroad

?Vehicleidentifiesandtrackspedestrian

?Vehicledoesnotbreak

?Safetydriverwasdistractedanddidnotact

?Factoryauto-breakingsystemdisabledsoasnottointerferewith

automateddrivingsoftware

?Exampleoflimitationsofdataandwhathappenswhennotfollowingfunctionalsafetybestpractices

CRUISE

?CruiseaccidentOctober2023

?Pedestriancrossesroadagainstdonotwalksignal

?Pedestriangetshalfwayacrossbeforecrosstrafficforcespedestriantowalkback

?Pedestrianhitbyacarandthrownintothepathofcruisevehicle

CRUISE

?Cruisevehicleisacceleratingeventhoughit“sees”pedestrian

?Vehicledoesnotrecognizescenario(edgecase)

?PedestriangetstrappedundertheCruisevehicle

?Vehiclesystemrecognizessomethingiswrong

?Insteadofstopping,thevehicledrivesforwardandpullsover,draggingthepedestrianunderthecar

?Whynotstop?

?Exampleoflimitationsofdataandwhathappenswhennotfollowingfunctionalsafetybestpractices

COMMONAIMISTAKES

LACKOFUNDERSTANDINGMODELS

?Mathishard,andlibrariesareeasy

?ThereareTONSofdifferentAImodelsandapproaches

TensorFlowiseasyandrequiresalmostnothinking

AnacondaextendsaccesstoeveryonewhocanPythonAmodelorimplementationinsearchofaproblem

Sometimes,neuralnetworksarenotthebestsolution

OftenseeproblemssolvedwithMLthatshouldbesolvedbyexpertsystems

Lackofunderstandingoflimitations

DisconnectbetweenAIandfunctionalsafety

DATASCIENCEDISCIPLINE

?Traditionalprogrammingisbasedonrequirements

?AIisbasedonDATA,whichmakesdatasciencecritical

?“Garbagein,garbageout”x1000

?Acommonerroristhatdatasciencelifecyclenotfollowed

?Smalldatasetsforlearningandvalidationareproblematic

?Lackofproperlysizeddatasetsleadto:

?Overfitmodelschasingaccuracy

?Falseminimaandmaximaforoptimizationproblems

?Higherprobabilityeventsarenotincludedinthemodel

?Significantmisclassificationscomparedtoreal-worlddata

Experimentationandprediction

DATASCIENCELIFECYCLE

preparation

Data

Explorationandvisualization

?Datacollectionandstorage

?Defineprojectobjectives

?Collectdata

?Normalizestorageandformat

?Explorationandvisualization

?Statisticaldataanalysis

?Datalabeling

?Datapreparation

?Missingorinconsistentdata

?Cleaningandaugmentingdata

?Removingduplicates

?Normalization

?Datalabeling

?Datatypeconversation

?Graphsandchartsforunderstanding

?Experimentationandprediction

?Trydifferentmodelsandapproaches

?Identifydatatrendsandpatterns

?Discoverinsights

?Buildmodel

Data

collection/storage

Source:

/blog/what-is-data-science-the-definitive-guide

TRANSPORTATIONAPPLICATIONS

ASSISTINGDRIVERS

?Safety-related“assistant”applicationstoreduceimpairedordistracteddriving

?Lane-keepingassist

?Intelligentcruisecontrol

?Automaticemergencybraking

?Limitcellphoneusewhiledriving(cellphone“drivingmode”)

?Limitcellphonedistractionsviapredictivebehaviororvehicleintegration

?Bettermusicplaylistprediction

?Bettermapsanddirections

?Recognizeimpaireddriving

?Inallthesescenarios,thedriverisstillthemaincontrollingactor

ANOMALY/ERRORDETECTION

?Vehiclepredictivemaintenance

?Batterylife

?Tirewear

?Many,manymore…..

?MotorfreightcarriersandTSPs

?Analyzebatteryvoltages

?Exhaustsensors

?Noiseandvibrationsensors

?CANbusmessages

?IntermittentDTCs

ANOMALY/ERRORDETECTION

?Vehiclecybersecurity[atscale]

?VSOC

?Faultpatterns

?Geolocationtrends

?IndividualvehicleIDSstillproblematic

?Lackoftrainingandvalidationdata

?Rule-basedexpertsystemsareprobablymoreeffective

?CompanionroleforML(seefunctionalsafety“safetybag”examples)

GENERALTRANSPORTATION

?TherearemanyapplicationsofAIinTransportationManagementSystems(TMS)andTrafficManagementSystems(TMS)

?Freightmovementoptimization

?Fuelconservation

?Mostefficientroutecalculations

?Trafficmanagement

?Parkingefficiencyandoptimization

FUNCTIONAL-SAFETY

?Functionalsafetyiswell-knownpracticewithspecificrulesandknownapproachestoachievingsafety

?Mitigationtechniquestodealwithuncertainty

?Safetybag(thinkinput/outputparametervalidation)

?Safetymonitors

?Diagnostics

?Formalmethods

?Functionalsafetysystemslikecrashavoidanceandlanedepartureassistthatcontainclassifiermodelsarenotprimarysafetysystems

?Driverremainstheprimarysafetysystemincontrolofthevehicle

FUNCTIONAL-SAFETY

?Neuralnetwork-basedAIisapoorchoiceforfunctionalsafetysystems

?Testingmassivelycomplexnon-deterministicsystemsisalmostimpossible

?Caveat:FormalmethodscombinedwithML

?Impossibletoexplainwhyamodelbehavesinacertainway

?Introducessafetyrisksandmassivelegalliabilities

AUTONOMOUSVEHICLES

?ExistingMLmodelsareunsuitableforSAELevel3–5automation

?Wedonothavethedatatousethemeffectively

?CurrentMLmodelsarenotexplainableortrustworthy

?MoreadvancedMLmodelsarenon-deterministic

?MLissuitableforclassifiersandpreceptorsbutnotatthe

accuracyrequiredforfunctionalsafety

?Closedandcontrolledenvironmentsarepossible

?Real-worldpublicroadsandadversarialenvironmentsaretoocomplexwheresafetystandardscannotbemet

Image:createdbymonkik

ATTACKINGAI

?Traditionaltechniquesarestillapplicable

?Denialofserviceattacks

?Softwarestackvulnerabilitiesandexploits

?Hostingandruntimeenvironmentexploitation

?Socialengineering

?TherearenewattackmethodstargetingAI

?MITREAdversarialThreatLandscapeforAI

?OWASPTop10MachineLearningRisks

Systems(ATLAS?)

CLASSIFICATIONINPUTMANIPULATION

?Modificationofaninput(e.g.image,sensorvalue)tocause:

?Misclassification

?Triggererrorconditions

?Alterintendedbehavior(inference-basedsystems)

?Acommonexampleisstopsign“modification”:

?Addingtapetocausemisclassification

?Shiningbrightlightsalteringgradientanalysis

?Alargenumberofacademicpapersonhowtomessuptrafficsignalinputs

?FewMLmodelsareimmune

?Distinctfrompromptmanipulation(coveredlater)

EXPLOITEDGECASES

?Limitationsofavailabledataallowedgecaseexploitation

?Analyzethemodelanddeterminelow-probabilityinputs

?UseanAItofuzzanotherAImodeltodeterminelimitations

?Causethemodeltobehaveincorrectlyorevencrash

?Especiallyeffectiveifinputandoutputvaluesarenotvalidatedandboundschecked

?Maycausesystemfailureorsoftwarestackmalfunction

?Errorconditionscancauseremotecodeexecutionordataexfiltrationopportunities

PROMPTINJECTION/MANIPULATION

?ApplicabletoLLMmodelswhichgenerateoutputbasedonprompts

?OnewaytothinkofthisisasSQLInjectionattacks,butinsteadoftargetingaSQLDB,theunderlyingmodelistargeted

?Injection/promptattackscancause:

?Datadisclosure

?Modelcorruption

?Hostingsystemcorruption

?Bypasssafetyorcontentrestrictions

?Wecanalsousesocialengineeringtrickstogetthemodeltodothingsitisnotsupposedtodo(hardastrickinga4yrold)

TRAININGDATAPOISONING

?AImodelsarebuiltfromthetrainingdata

?Duetolackoftrainingdata,manytrainingsetsarebasedonpublicdatasources

?Poisoningapublicdatasetcanintroduce

?Backdoors

?Remotecodeexecutionerrors

?Anynumberofmalwarescenarios

?Feedingmaliciousdataintoacontinuouslearningmodelcancausemodeldriftandeventualmodelfailure

?ContinuouslylearningIDSsystems

?Self-optimizingmodels

HACKING

EXAMPLE

HACKINGADASMODEL

?BlackhatAsia2024-TheKeytoRemoteVehicleControl:AutonomousDrivingDomainController

?ShupengGao,SeniorSecurityResearcher,Baidu

?YingtaoZeng,SeniorSecurityResearcher,Baidu

?JieGao,SeniorSecurityResearcher,Baidu

?YimiHu,SeniorSecurityResearcher,Baidu

?Analyzedover30ADASdevices

?50%hadSHHenabled

?Somedeployedwithoriginalmodelfiles(*.onnx)

?Littleornodiskencryption

/asia-24/briefings/schedule/index.html#the-key-to-remote-vehicle-control-autonomous-driving-domain-contr

oller-38089

HACKINGADASMODEL

?ADASunitcanbeapathwaytototalvehiclecompromiseasitneedstobeaccessibleonCANnetworkandupdateable

?Researchers:

?Abletooffloadentiremodelfiles,includingoriginalmodelfiles

?Abletoreadthemodelanddeployina$50toycar

?Possibletoupdatemodelonoriginaldevice

?PoorlysecuredADASmodulecanleadtototalvehiclecontrol

?Steering

?Braking

?Powertrain

DEFENDINGAI

AI-WHATISTHESAME?

Themorethingschange,themoretheyremainthesame:

?Softwarestackvulnerabilities

?Operatingsystemvulnerabilities

?Softwaresupplychainattacks

?Hardwarefirmware

?PlatformOS

?Browservulnerabilities

?ITandDevSecOpsbestpracticesstillapply

?Sanitizeandvalidateinputsandoutputs

AI-WHAT’SDIFFERENT?

?Systemsarebuiltgroundupfromdata,notrequirements

?Awholenewmindsetforadversarialattackvectors

?Increasedsupplychaincomplexity

?Inputvalidationbecomesharderandmoreimportant

?Datahandlingproceduresaremoreimportant

?Erroneous,mislabeled,orincompletedatacanhavebigimpact

?DevSecOpsneedstoincorporatedataandmachinelearningmodels

?Datascientistsneedtounderstandcybersecurity

AI–DEFENSIVEBESTPRACTICES

Whilenotanexhaustivelist,herearesomebestpractices:

?ProtectsystemboundariesbetweenITandAI

?Identifyandprotectallproprietarydata

?StrongaccessandauthorizationcontrolsforfinalAImodelweights

?Hardenthedeploymentenvironment

?Applyversionlabelstomodels(changestoweights)

?Validateallinputsforedgecasesandattacks

?Validatealloutputstoensureoperationinsideboundaries(safetybag)

DEFENSIVERESOURCES

ThereareseveralrecentgoodpublicationsoncybersecuritybestpracticesfordeployingmachinelearningandAIingeneral:

?NationalCyberSecurityCentre(UK)-Principlesforthesecurityofmachine

learning

.uk/collection/machine-learning

?NationalCyberSecurityCentre(UK)–GuidelinesforsecureAIsystem

development

.uk/collection/guidelines-secure-ai-system-development

?JointCybersecurityInformation-DeployingAISystemsSecurely

/2024/Apr/15/2003439257/-1/-1/0/CSI-DEPLOYING-AI-

SYSTEMS-SECURELY.PDF

WRAP-UP

AIFUTURE

?Improvementsincustomerservice

?Improvementsinoperationalefficiency

?Developingbetterdesigns

?Assistingindevelopmentofnewmaterials

?Inspectingandevaluatinginfrastructure

?Improvingsafetythroughnewdriverassistancefeatures

?Increasefleetuptime

……ExplainableandtrustworthyAIwillbringmoreapplications

FURTHERREADING

IfyouarenewtoAI/ML,Icanhighlyrecommendthefollowingbookasagoodstartingpoint

?Ozdemir,S.,Kakade,S.,&Tibaldeschi,M.(2018).PrincipalsofDataScience(2nded.).PacktPublishing.

/product/principles-of-data

-science-second-edition/9781789804546

Ifyouarelookingforamorein-depthfoundationalbook,Iwouldrecommendthefollowingbook:

?Kelleher,J.D.,MacNamee,B.,&D’Arcy,A.(2015).FundamentalsofMachineLearningforPredictiveDataAnalytics.TheMITPress.

FURTHERREADING

Forthosewhowanttogettotheheartofthemathandbuildyourownmodels,includingdeepregressionlearning,Iwouldrecommendthefollowingbook:

?Goodfellow,I.,Bengio,Y.,&Courville,A.(2017).DeepLearning.TheMITPress.Iwouldalsorecommendthefollowingpaperformoreinformationaboutsafety-criticalAIapplicationsintransportationtohelptietogethertheuseofAIintransportation:

?JonPerez-Cerrolaza,JaumeAbella,MarkusBorg,CarloDonzella,Jesús

Cerquides,FranciscoJ.Cazorla,CristoferEnglund,MarkusTauber,George

Nikolakopoulos,andJoseLuisFlores.2024.ArtificialIntelligenceforSafety-CriticalSystemsin

人人文庫> 全部分類> 應(yīng)用文書 > 研究報告

溫馨提示

1. 本站所有資源如無特殊說明，都需要本地電腦安裝OFFICE2007和PDF閱讀器。圖紙軟件為CAD,CAXA,PROE,UG,SolidWorks等.壓縮文件請下載最新的WinRAR軟件解壓。
2. 本站的文檔不包含任何第三方提供的附件圖紙等，如果需要附件，請聯(lián)系上傳者。文件的所有權(quán)益歸上傳用戶所有。
3. 本站RAR壓縮包中若帶圖紙，網(wǎng)頁內(nèi)容里面會有圖紙預(yù)覽，若沒有圖紙預(yù)覽就沒有圖紙。
4. 未經(jīng)權(quán)益所有人同意不得將文件中的內(nèi)容挪作商業(yè)或盈利用途。
5. 人人文庫網(wǎng)僅提供信息存儲空間，僅對用戶上傳內(nèi)容的表現(xiàn)方式做保護(hù)處理，對用戶上傳分享的文檔內(nèi)容本身不做任何修改或編輯，并不能對任何下載內(nèi)容負(fù)責(zé)。
6. 下載文件中如有侵權(quán)或不適當(dāng)內(nèi)容，請與我們聯(lián)系，我們立即糾正。
7. 本站不保證下載資源的準(zhǔn)確性、安全性和完整性, 同時也不承擔(dān)用戶因使用這些下載資源對自己和他人造成任何形式的傷害或損失。

揭開人工智能在交通網(wǎng)絡(luò)安全中的神秘面紗（英文）

文檔簡介

溫馨提示

最新文檔

評論

揭開人工智能在交通網(wǎng)絡(luò)安全中的神秘面紗（英文）

文檔簡介

溫馨提示

最新文檔

評論

相關(guān)文檔