AI組織責(zé)任-核心安全責(zé)任 AI Organizational Responsibilities Core Security Responsibilities_第1頁(yè)
AI組織責(zé)任-核心安全責(zé)任 AI Organizational Responsibilities Core Security Responsibilities_第2頁(yè)
AI組織責(zé)任-核心安全責(zé)任 AI Organizational Responsibilities Core Security Responsibilities_第3頁(yè)
AI組織責(zé)任-核心安全責(zé)任 AI Organizational Responsibilities Core Security Responsibilities_第4頁(yè)
AI組織責(zé)任-核心安全責(zé)任 AI Organizational Responsibilities Core Security Responsibilities_第5頁(yè)
已閱讀5頁(yè),還剩94頁(yè)未讀, 繼續(xù)免費(fèi)閱讀

下載本文檔

版權(quán)說(shuō)明:本文檔由用戶(hù)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(qǐng)進(jìn)行舉報(bào)或認(rèn)領(lǐng)

文檔簡(jiǎn)介

AIOrganizationalResponsibilities:

CoreSecurityResponsibilities

ThepermanentandofficiallocationfortheAIOrganizationalResponsibilitiesWorkingGroupis

/research/working-groups/ai-organizational-responsibilities

?2024CloudSecurityAlliance–AllRightsReserved.Youmaydownload,store,displayonyour

computer,view,print,andlinktotheCloudSecurityAllianceat

subjectto

thefollowing:(a)thedraftmaybeusedsolelyforyourpersonal,informational,noncommercialuse;(b)thedraftmaynotbemodifiedoralteredinanyway;(c)thedraftmaynotberedistributed;and(d)thetrademark,copyrightorothernoticesmaynotberemoved.Youmayquoteportionsofthedraftas

permittedbytheFairUseprovisionsoftheUnitedStatesCopyrightAct,providedthatyouattributetheportionstotheCloudSecurityAlliance.

?Copyright2024,CloudSecurityAlliance.Allrightsreserved.2

Acknowledgments

LeadAuthors

JerryHuangKenHuang

Contributors/Co-Chairs

KenHuang

NickHamiltonChrisKirschkeSeanWright

Reviewers

CandyAlexanderIlangoAllikuzhiErayAltili

AakashAlurkarRomeoAyalinRenuBedi

SauravBhattacharya

SergeiChaschinHongChen

JohnChiu

SatchitDokras

RajivGunja

HongtaoHao,PhDGraceHuang

OnyekaIlloh

KrystalJackson

ArvinJakkamreddyReddySimonJohnson

GianKapoor

BenKereopa-Yorke

ChrisKirschke

MaduraMalwatte

MadhaviNajana

RajithNarasimhaiah

GabrielNwajiaku

GovindarajPalanisamyMeghanaParwate

PareshPatel

RangelRodrigues

MichaelRoza

LarsRuddigkeit

DavideScatto

MariaSchwengerMj

BhuvaneswariSelvadurai

HimanshuSharmaAkshayShetty

NishanthSingarapuAbhinavSingh

Dr.ChantalSpleissPatriciaThaine

EricTierling

AshishVashishthaPeterVentura

JiewenWangWickeyWang

UdithWickramasuriyaSounilYu

CSAGlobalStaff

MarinaBregkouSeanHeide

AlexKaluza

ClaireLehnertStephenLumpe

?Copyright2024,CloudSecurityAlliance.Allrightsreserved.3

TableofContents

Acknowledgments 3

TableofContents 4

ExecutiveSummary 6

Introduction 7

AISharedResponsibilityModel 7

KeyLayersinanAI-EnabledApplication 7

FoundationalComponentsofaData-CentricAISystem 9

Assumptions 13

IntendedAudience 13

ResponsibilityRoleDefinitions 14

ManagementandStrategy 14

GovernanceandCompliance 15

TechnicalandSecurity 16

OperationsandDevelopment 16

NormativeReferences 18

1.IncorporatingDataSecurity&PrivacyinAITraining 19

1.1DataAuthenticityandConsentManagement 19

1.2AnonymizationandPseudonymization 20

1.3DataMinimization 21

1.4AccessControltoData 22

1.5SecureStorage&Transmission 23

2.ModelSecurity 24

2.1.AccessControlstoModels 24

2.1.1AuthenticationandAuthorizationFrameworks 24

2.1.2.ModelInterfacesRateLimiting 25

2.1.3.AccessControlinModelLifecycleManagement 25

2.2.SecureModelRuntimeEnvironment 26

2.2.1.Hardware-BasedSecurityFeatures 26

2.2.2.NetworkSecurityControls 27

2.2.3.OS-LevelHardeningandSecureConfigurations 28

2.2.4.K8sandContainerSecurity 29

2.2.5.CloudEnvironmentSecurity 29

2.3VulnerabilityandPatchManagement 30

2.3.1MLCodeIntegrityProtections 30

2.3.2VersionControlSystemsforMLTrainingandDeploymentCode 31

2.3.3CodeSigningtoValidateApprovedVersions 32

2.3.4InfrastructureasCodeApproaches 32

2.4MLOpsPipelineSecurity 33

2.4.1.SourceCodeScansforVulnerabilities 33

?Copyright2024,CloudSecurityAlliance.Allrightsreserved.4

2.4.2.TestingModelRobustnessAgainstAttacks 34

2.4.3.ValidatingPipelineIntegrityatEachStage 35

2.4.4.MonitoringAutomationScripts 36

2.5AIModelGovernance 37

2.5.1.ModelRiskAssessments 37

2.5.2.BusinessApprovalProcedures 37

2.5.3.ModelMonitoringRequirements 38

2.5.4.NewModelVerificationProcesses 39

2.6SecureModelDeployment 39

2.6.1.CanaryReleases 40

2.6.2.Blue-GreenDeployments 40

2.6.4.RollbackCapabilities 41

2.6.5.DecommissioningModels 41

3.VulnerabilityManagement 42

3.1.AI/MLAssetInventory 42

3.2.ContinuousVulnerabilityScanning 43

3.3.Risk-BasedPrioritization 44

3.4.RemediationTracking 45

3.5.ExceptionHandling 45

3.6.ReportingMetrics 46

Conclusion 48

Acronyms 49

?Copyright2024,CloudSecurityAlliance.Allrightsreserved.5

ExecutiveSummary

ThiswhitepaperisaworkingdraftthatfocusesontheinformationsecurityandcybersecurityaspectsoforganizationalresponsibilitiesinthedevelopmentanddeploymentofArtificialIntelligence(AI)and

MachineLearning(ML)systems.Thepapersynthesizesexpert-recommendedbestpracticeswithincoresecurityareas,includingdataprotectionmechanisms,modelvulnerabilitymanagement,Machine

LearningOperations(MLOps)pipelinehardening,andgovernancepoliciesfortraininganddeployingAIresponsibly.

Keypointsdiscussedinthewhitepaperinclude:

●DataSecurityandPrivacyProtection:Theimportanceofdataauthenticity,anonymization,pseudonymization,dataminimization,accesscontrol,andsecurestorageandtransmissioninAItraining.

●ModelSecurity:Coversvariousaspectsofmodelsecurity,includingaccesscontrols,secureruntimeenvironments,vulnerabilityandpatchmanagement,MLOpspipelinesecurity,AImodelgovernance,andsecuremodeldeployment.

●VulnerabilityManagement:DiscussesthesignificanceofAI/MLassetinventory,continuousvulnerabilityscanning,risk-basedprioritization,remediationtracking,exceptionhandling,andreportingmetricsinmanagingvulnerabilitieseffectively.

Thewhitepaperanalyzeseachresponsibilityusingquantifiableevaluationcriteria,theResponsible,

Accountable,Consulted,Informed(RACI)modelforroledefinitions,high-levelimplementationstrategies,continuousmonitoringandreportingmechanisms,accesscontrolmapping,andadherenceto

foundationalguardrails.ThesearebasedonindustrybestpracticesandstandardssuchasNISTAIRMF,NISTSSDF,NIST800-53,CSACCM,andothers.

Byoutliningrecommendationsacrossthesekeyareasofsecurityandcompliance,thispaperaimstoguideenterprisesinfulfillingtheirobligationsforresponsibleandsecureAIdesign,development,anddeployment

?Copyright2024,CloudSecurityAlliance.Allrightsreserved.6

Introduction

Thiswhitepaperfocusesonwhatwedefineasanenterprise's"coresecurityresponsibilities"around

ArtificialIntelligence(AI)andMachineLearning(ML),datasecurity,modelsecurity,andvulnerability

management.AsorganizationshavedutiestoupholdsecureandsafeAIpractices,thiswhitepaperandtwoothersinthisseriesprovideablueprintforenterprisestofulfillsuchorganizationalresponsibilities.

Specifically,thiswhitepapersynthesizesexpert-recommendedbestpracticeswithincoresecurityareas-dataprotectionmechanisms,modelvulnerabilitymanagement,MLOpspipelinehardening,and

governancepoliciesfortraininganddeployingAIresponsibly.TheothertwowhitepapersinthisseriesdiscussadditionalaspectsofsecureAIdevelopmentanddeploymentforenterprises.Byoutlining

recommendationsacrossthesekeyareasofsecurityandcomplianceinthreetargetedwhitepapers,thisseriesaimstoguideenterprisesinfulfillingtheirobligationsforresponsibleandsecureAIdesign,

development,anddeployment.

AISharedResponsibilityModel

TheAISharedResponsibilityModeloutlinesthedivisionoftasksbetweenAIplatformproviders,AIapplicationowners,AIdevelopersandAIusage,varyingbyservicemodels(SaaS,PaaS,IaaS).

ThesecureoperationofAIapplicationsinvolvesacollaborativeeffortamongmultiplestakeholders.InthecontextofAI,responsibilitiesaresharedbetweenthreekeyparties:theAIserviceusers,theAIapplicationownersanddevelopers,andAIplatformproviders.

WhenevaluatingAI-enabledintegration,itiscrucialtocomprehendthesharedresponsibilitymodelanddelineatethespecifictaskshandledbyeachparty.

KeyLayersinanAI-EnabledApplication

1.AIPlatform:

○ThislayerprovidestheAIcapabilitiestoapplications.Itinvolvesbuildingand

safeguardingtheinfrastructurethathostsAImodels,trainingdata,andconfigurationsettings.

○Securityconsiderationsincludeprotectingagainstmaliciousinputsandoutputs

generatedbytheAImodel.AIsafetysystemsshouldprotectagainstpotentialharmfulinputsandoutputslikehate,jailbreaks,andsoon.

○AIPlatformLayerhasfollowingtasks:

■Modelsafetyandsecurity

■Modeltuning

■Modelaccountability

■Modeldesignandimplementation

■Modeltrainingandgovernance

■AIcomputeanddatainfrastructure

?Copyright2024,CloudSecurityAlliance.Allrightsreserved.7

2.AIApplicationLayer:

○TheAIapplicationlayerinterfaceswithusers,leveragingtheAIcapabilities.Itscomplexitycanvarysignificantly.Attheirmostbasiclevel,standaloneAIapplicationsserveasa

conduittoacollectionofAPIs,whichprocesstextualpromptsfromusersandrelaythemtotheunderlyingmodelforaresponse.MoresophisticatedAIapplicationsarecapableofenrichingthesepromptswithadditionalcontext,utilizingelementssuchasapersistencelayer,asemanticindex,orpluginsthatprovideaccesstoabroaderrangeofdatasources.ThemostadvancedAIapplicationsaredesignedtointegrateseamlesslywithpre-existingapplicationsandsystems,enablingamulti-modalapproachthatencompassestext,

audio,andvisualinputstoproducediversecontentoutputs.

○AsanAIapplicationowner,youensureseamlessuserexperiencesandhandleany

additionalfeaturesorservices.TosafeguardanAIapplicationfromharmfulactivities,itisessentialtoestablisharobustapplicationsafetysystem.AGenerativeAI(GenAI)systemshouldthoroughlyexaminethecontentutilizedinthepromptdispatchedtotheAImodel.Additionally,itmustscrutinizetheexchangeswithanyadd-onslikepluginsandfunctions,dataconnectors,andinteractionswithotherAIapplications,aprocessreferredtoasAIorchestration.ForthosedevelopingAIapplicationsonanInfrastructure-as-a-Service

(IaaS)orPlatform-as-a-Service(PaaS)services,integratingadedicatedAIcontent

safetyfeatureisadvisable.Dependingonspecificrequirements,additionalfeaturesmaybeimplementedtoenhanceprotection.

○AIApplicationhasthefollowingtasks:

■AIpluginsanddataconnections

■Applicationdesignandimplementation

■Applicationinfrastructure

■AIsafetysystem

3.AIUsage:

○TheAIusagelayeroutlinestheapplicationandconsumptionofAIfunctionalities.GenAIintroducesaninnovativeuser/computerinteractionmodel,distinctfromtraditional

interfaceslikeAPIs,commandprompts,andGUIs.Thisnewinterfaceisinteractiveandadaptable,moldingthecomputer’scapabilitiestotheuser’sintentions.Unlikeearlierinterfacesthatrequireduserstoconformtothesystem’sdesignandfunctions,the

generativeAIinterfaceprioritizesuserinteraction.Thisallowstheusers’inputstosignificantlyshapethesystem’soutput,emphasizingtheimportanceofsafety

mechanismstosafeguardindividuals,data,andcorporateresources.

○SecurityconsiderationsforAIusageareakintothoseforanycomputersystem,relyingonrobustmeasuresforidentityandaccessmanagement,devicesecurity,monitoring,datagovernance,andadministrativecontrols.

○Giventhesignificantimpactuseractionscanhaveonsystemoutputs,agreaterfocusonuserconductandresponsibilityisnecessary.Itisessentialtorevisepoliciesfor

acceptableuseandtoinformusersaboutthedistinctionsbetweenconventionalIT

applicationsandthoseenhancedbyAI.ThiseducationshouldcoverAI-specificissuesconcerningsecurity,privacy,andethicalstandards.Moreover,it’simportanttoraiseawarenessamongusersaboutthepotentialforAI-drivenattacks,whichmayinvolvesophisticatedlyfabricatedtext,audio,video,andothermediadesignedtodeceive.

○AIusagelayerhasthefollowingtasks:

■Usertrainingandaccountability

■Acceptableusagepolicyandadmincontrols

■IdentityandAccessManagement(IAM)anddevicecontrols

?Copyright2024,CloudSecurityAlliance.Allrightsreserved.8

■Datagovernance

Rememberthatthissharedresponsibilitymodelhelpsdemarcaterolesandensuresaclearseparationofduties,contributingtothesafeandeffectiveuseofAItechnologies.Thedistributionofworkload

responsibilitiesvariesbasedonthetypeofAIintegrationbasedonservicemodels.

1.SoftwareasaService(SaaS):

○InSaaS-basedAIintegrations,theAIplatformproviderassumesresponsibilityfor

managingtheunderlyinginfrastructure,securitycontrols,andcompliancemeasures.

○Asauser,yourprimaryfocusliesinconfiguringandcustomizingtheAIapplicationtoalignwithyourspecificrequirements.

2.PlatformasaService(PaaS):

○PaaS-basedAIplatformsofferamiddleground.WhiletheprovidermanagesthecoreAIcapabilities,youretainsomecontroloverconfigurationsandcustomization.

○YouareresponsibleforensuringthesafeuseoftheAImodel,handlingtrainingdata,andadjustingmodelbehavior(e.g.,weightsandbiases).

3.InfrastructureasaService(IaaS):

○InIaaSscenarios,youhavegreatercontrolovertheinfrastructure.However,thisalsomeanstakingonmoreresponsibilities.

○Youmanagetheentirestack,includingtheAImodel,trainingdata,andinfrastructuresecurity.

FoundationalComponentsofaData-CentricAISystem

Thefoundationalcomponentsofadata-centricAIsystemencompasstheentirelifecycleofdataandmodelmanagement.ThesecomponentsworktogethertocreateasecureandeffectiveAIsystemthatcanprocessdataandprovidevaluableinsightsorautomateddecisions.

●RawData:Theinitialunprocesseddatacollectedfromvarioussources.

●Datapreparation:Theprocessofcleaningandorganizingrawdataintoastructuredformat.

●Datasets:Curatedcollectionsofdata,readyforanalysisandmodeltraining.

●DataandAIgovernance:PoliciesandprocedurestoensuredataqualityandethicalAIusage.

●MachineLearningalgorithms:Thecomputationalmethodsusedtointerpretdata.

●Evaluation:Assessingtheperformanceofmachinelearningmodels.

●MachineLearningModels:Theoutputofalgorithmstrainedondatasets.

●Modelmanagement:Overseeingthelifecycleofmachinelearningmodels.

●Modeldeploymentandinference:Implementingmodelstomakepredictionsordecisions.

●Inferenceoutcomes:Theresultsproducedbydeployedmodels.

●MachineLearningOperations(MLOps):PracticesfordeployingandmaintainingAImodels.

●DataandAIPlatformsecurity:Measurestoprotectthesystemagainstthreats.

DataOperations:Involvestheacquisitionandtransformationofdata,coupledwiththeassuranceofdatasecurityandgovernance.TheefficacyofMLmodelsiscontingentupontheintegrityofdata

pipelinesandafortifiedDataOpsframework.

ModelOperations:EncompassesthecreationofpredictiveMLmodels,procurementfrommodel

marketplaces,ortheutilizationofLargeLanguageModels(LLMs)suchasthoseprovidedbyOpenAIor

?Copyright2024,CloudSecurityAlliance.Allrightsreserved.9

throughFoundationModelAPIs.Modeldevelopmentisaniterativeprocessthatnecessitatesasystematicapproachtodocumentandevaluatevariousexperimentalconditionsandoutcomes.

ModelDeploymentandServing:Entailsthesecureconstructionofmodelcontainers,theisolatedandprotecteddeploymentofmodels,andtheimplementationofautomatedscaling,ratelimiting,and

surveillanceofactivemodels.Italsoincludestheprovisionoffeaturesandfunctionsforhigh-availability,low-latencyservicesinRetrievalAugmentedGeneration(RAG)applications,aswellastherequisite

featuresforotherapplications,includingthosethatdeploymodelsexternallytotheplatformorrequiredatafeaturesfromthecatalog.

OperationsandPlatform:Coversthemanagementofplatformvulnerabilities,updates,model

segregation,andsystemcontrols,alongwiththeenforcementofauthorizedmodelaccesswithinasecurearchitecturalframework.Additionally,itinvolvesthedeploymentofoperationaltoolsforContinuous

Integration/ContinuousDeployment(CI/CD),ensuringthattheentirelifecycleadherestoestablishedstandardsacrossseparateexecutionenvironments—development,staging,andproduction—forsecureMLoperations(MLOps).

Table1alignstheoperationswiththecoreaspectsofadata-centricAIsystem,highlightingtheirrolesandinterdependencies

FoundationalComponent

Description

DataOperations

Ingestion,transformation,security,andgovernanceofdata.

ModelOperations

Building,acquiring,andexperimentingwithMLmodels.

ModelDeploymentandServing

Securedeployment,serving,andmonitoringofMLmodels.

OperationsandPlatform

Platformsecurity,modelisolation,andCI/CDforMLOps.

Table1:MappingData-CentricAISystemComponentsandTheirInterconnectedRoles

?Copyright2024,CloudSecurityAlliance.Allrightsreserved.10

Table2providesasynthesizedviewofthepotentialsecurityrisksandthreatsateachstageofanAI/MLsystem,alongwithexamplesandrecommendedmitigationstoaddresstheseconcerns.

SystemStage

System

Components

Potential

SecurityRisks

Threats

Mitigations

DataOperations

RawData,DataPrep,Datasets

Dataloss:Unauthorizeddeletionorcorruptionofdata.Datapoisoning:

Deliberatemanipulationofdatatocompromisethemodel’sintegrity.

Compliancechallenges:Failuretomeet

regulatoryrequirementsfordataprotection.

Compromise/poisoningofdata:Attackersmayinjectfalsedataoralterexistingdata.

Implementrobustdatagovernanceframeworks.Deployanomaly

detectionsystems.Establishrecovery

protocolsandregulardatabackups.

Model

Operations

MLAlgorithms,

Model

Management

Modeltheft:Stealingofproprietarymodels.

Unauthorizedaccess:Gainingaccessto

modelswithoutpermission.

AttacksviaAPIaccess:ExploitingAPI

vulnerabilitiestoaccessormanipulatemodels.Modelstealing

(extraction):Replicatingamodelfor

unauthorizeduse.

Strengthenaccesscontrolsand

authentication

mechanisms.SecureAPIendpointsthrough

encryptionandratelimiting.Regularlyupdateandpatchsystems.

Model

Deploymentand

Serving

ModelServing,InferenceResponse

Unauthorizedaccess:Accessingthemodel

servinginfrastructurewithoutauthorization.Dataleakage:Exposingsensitiveinformationthroughmisconfiguredsystems.

Modeltricking

(evasion):Altering

inputstoreceivea

specificoutputfromthemodel.Trainingdata

recovery(inversion):Extractingprivate

trainingdatafromthemodel.

Securedeployment

practices,including

containerizationand

networksegmentation.Activemonitoringandloggingofmodel

interactions.Implementratelimitingand

anomalydetection.

OperationsandPlatform

MLOperations,

DataandAI

PlatformSecurity

Inadequatevulnerabilitymanagement:Not

addressingknown

vulnerabilitiesinatimelymanner.Modelisolationissues:Failureto

properlyisolatemodels,leadingtopotential

cross-contamination.

AttackingMLsupply

chain:Introducing

vulnerabilitiesor

backdoorsinthird-partycomponents.Model

contamination

(poisoning):Corruptingtrainingdatatocausemisclassificationor

systemunavailability.

Continuousvulnerabilitymanagementand

patching.CI/CD

processesforconsistentdeployment.Isolation

controlsandsecurearchitecturedesign.

Table2:AI/MLSecurityRiskOverview

?Copyright2024,CloudSecurityAlliance.Allrightsreserved.11

Weanalyzeeachresponsibilityinthefollowingdimensions.

1.EvaluationCriteria:WhendiscussingAIresponsibility,considerquantifiablemetricsforassessingthesecurityimpactofAIsystems.Byquantifyingtheseaspects,stakeholderscanbetterunderstandthe

associatedrisksofAItechnologiesandhowtoaddressthoserisks.Organizationsmustfrequently

evaluatetheirAIsystemstoensuresecurityandreliability.Theyshouldassessmeasurablethingslikehowwellthesystemhandlesattacks(adversarialrobustness),whetheritleakssensitivedata,howoftenit

makesmistakes(false-positiverates),andwhetherthetrainingdataisreliable(dataintegrity).Evaluatingandmonitoringthesecriticalmeasuresaspartoftheorganization'ssecurityplanwillhelpimproveoverallsecuritypostureofAIsystems.

2.RACIModel:ThismodelhelpsclarifywhoisResponsible,Accountable,Consulted,andInformed

(RACI)regardingAIdecision-makingandoversight.ApplyingtheRACImodeldelineatesrolesand

responsibilitiesinAIgovernance.ThisallocationofresponsibilitiesisessentialforsecureAIsystems.Itisimportanttounderstandthatdependingonanorganization'ssizeandbusinessfocus,thespecificrolesandteamsdelineatedinthiswhitepaperareforreferenceonly.Theemphasisshouldbeonclearly

outliningthekeyresponsibilitiesfirst.Organizationscanthendeterminetheappropriaterolestomaptothoseresponsibilities,andsubsequently,theteamstofillthoseroles.Theremaybesomeoverlapping

responsibilitiesacrossteams.TheRACIframeworkdefinedhereinaimstoprovideinitialroleandteam

designationstoaidorganizationsindevelopingtheirowntailoredRACImodels.However,implementationmayvaryacrosscompaniesbasedontheiruniqueorganizationalstructuresandpriorities.

3.High-levelImplementationStrategies:ThissectionoutlinesstrategiesforseamlesslyintegratingcybersecurityconsiderationsintotheSoftwareDevelopmentLifecycle(SDLC).Organizationsmust

prioritizetheenforcementofCIAprinciples—ensuringtheconfidentiality,integrity,andavailabilityofdataandsystems.Accesscontrolmechanismsshouldbeimplementedrigorouslytomanageuserpermissionsandpreventunauthorizedaccess.Robustauditingmechanismsmusttracksystemactivityandpromptlydetectsuspiciousbehavior.Impactassessmentsshouldevaluatepotentialcybersecurityrisks,focusingonidentifyingvulnerabilitiesandmitigatingthreatstosafeguardsensitiveinformationinAIsystems

4.ContinuousMonitoringandReporting:ContinuousMonitoringandReportingensurestheongoingsecurity,safety,andperformanceofAIsystems.Criticalcomponentsincludereal-timemonitoring,alertsforpoormodelperformanceorsecurityincidents,audittrails/logs,andregularreporting,followedby

actiontoimplementimprovementsandresolveissues.ContinuousMonitoringandReportinghelpsorganizationsmaintaintransparency,enhanceperformanceandaccountability,andbuildtrustinAI

systems.

5.AccessControl:AccesscontroliscrucialforsecuringAIsystems.ThisincludesstrongAPI

authentication/authorizationpolicies,managingmodelregistries,controllingaccesstodatarepositories,overseeingcontinuousintegrationanddeploymentpipelines(CI/CD),handlingsecrets,andmanaging

privilegedaccess.BydefininguserrolesandpermissionsforvariouspartsoftheAIpipeline,sensitivedatacanbesafeguarded,andmodelscan'tbetamperedwithoraccessedwithoutproperauthorization.

ImplementingstrongidentityandaccessmanagementnotonlyprotectsintellectualpropertybutalsoensuresaccountabilitythroughoutAIworkflows.

?Copyright2024,CloudSecurityAlliance.Allrightsreserved.12

6.AdherencetoFoundationalGovernance,RiskandCompliance,Security,Safety,andEthicalGuardrails:Emphasizeadherencetoguardrailsbasedonindustrybestpracticesandregulatory

requirementssuchasthefollowing:

●NISTSSDFforsecuresoftwaredevelopment

●NISTArtificialIntelligenceRiskManagementFramework(AIRMF)

●ISO/IEC42001:2023AIManagementSystem(AIMS)

●ISO/IEC27001:2022InformationSecurityManagementSystem(ISMS)

●ISO/IEC27701:2019PrivacyInformationManagementSystem(PIMS)

●ISO31700-1:2023ConsumerprotectionPrivacybydesignforconsumergoodsandservices

●OWASPTop10forLLMApplications

●NISTSP800-53Rev.5SecurityandPrivacyControlsforInformationSystemsandOrganizations

●GeneralDataProtectionRegulation(GDPR)ondataanonymizationandpseudonymizationandguidance

●Guidancefortokenizationoncloud-basedservices

Assumptions

Thisdocumentassumesanindustry-neutralstance,providingguidelinesandrecommendationsthatcanbeapplicableacrossvarioussectorswithoutaspecificbiastowardsaparticularindustry.

IntendedAudience

Thewhitepaperisintendedtocatertoadiverserangeofaudiences,eachwithdistinctobjectivesandinterests.

1.ChiefInformationSecurityOfficers(CISOs):ThiswhitepaperisspecificallydesignedtoaddresstheconcernsandresponsibilitiesofCISOs.ItprovidesvaluableinsightsintointegratingcoresecurityprincipleswithinAIsystems.PleasenotethattheroleofChiefAIOfficer(CAIO)isemerginginmanyorganizations,andit'santicipatedthatamajorityofrelatedresponsibilitiesdefinedinthiswhitepapermayshiftfromCISOtoCAIOinthenearfuture.

2.AIresearchers,engineers,dataprofessionals,scientists,analystsanddevelopers:ThepaperofferscomprehensiveguidelinesandbestpracticesforAIresearchersandengineers,aidingthemin

developingethicalandtrustworthyAIsystems.ItservesasacrucialresourceforensuringresponsibleAIdevelopment.

3.Businessleadersanddecisionmakers:Forbusinessleadersanddecision-makerssuchasCIO,CPO,CDO,CRO,CEOandCTOthewhitepaperoffersessentialinformationandawarenessforcybersecuritystrategiesrelatedtoAIsystemdevelopment,deployment,andlifecyclemanagement.

?Copyright2024,CloudSecurityAlliance.Allrightsreserved.13

4.Policymakersandregulators:Policymakersandregulatorswillfindthispaperinvaluableasit

providescriticalinsightstohelpshapepolicyandregulatoryframeworksconcerningAIe

溫馨提示

  • 1. 本站所有資源如無(wú)特殊說(shuō)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(qǐng)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(qǐng)聯(lián)系上傳者。文件的所有權(quán)益歸上傳用戶(hù)所有。
  • 3. 本站RAR壓縮包中若帶圖紙,網(wǎng)頁(yè)內(nèi)容里面會(huì)有圖紙預(yù)覽,若沒(méi)有圖紙預(yù)覽就沒(méi)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(kù)網(wǎng)僅提供信息存儲(chǔ)空間,僅對(duì)用戶(hù)上傳內(nèi)容的表現(xiàn)方式做保護(hù)處理,對(duì)用戶(hù)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(duì)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(qǐng)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(shí)也不承擔(dān)用戶(hù)因使用這些下載資源對(duì)自己和他人造成任何形式的傷害或損失。

最新文檔

評(píng)論

0/150

提交評(píng)論