1

Abstract

Thenext-generation6Gmobilesystemaimstosupportandimplementintelligentconnectionofeverything,intelligentassociationofservicesandtasks,andinclusiveintelligence.Itwillbecomeapowerfulintelligentinfrastructureplatformthatsupportsandempowersawidevarietyofindustries.Withtheacquisitionofconsumersatalllevelsandthein-depthdevelopmentofdigitalandintelligentapplicationsinvariousindustries,thereisanincreasinglyurgentneedtoprovidenativeAIcapabilitiesandservicesinnew6Gnetworks.Thiswhitepapersystematicallyanalyzesandsummarizesthetechnicalrequirementsandimpacts(intermsofcapability,service,architecture,standardization,etc.)ofnativeAIdesignbasedontherequirementsandtrendsoffuture6Gbusinessforms,newindustryscenariosandusecases,andnewtechnologiestoeffectivelysupportthedesignandimplementationofnativeAI-relatedtechnicalsolutionsfornew6Gnetworks.

Theorganizationsthathaveparticipatedinwritingthiswhitepaperincludebutarenotlimitedto:ZTE,CICTMobile,Huawei,ChinaMobile,ChinaTelecom,ChinaUnicom,ShanghaiTechUniversity,ChongqingUniversityofPostsandTelecommunications,NokiaShanghaiBell,vivo,PurpleMountainLaboratories,AsiaInfo,Ericsson,DalianMaritimeUniversity,andPengChengLaboratory.Manythankstoalltheseorganizationsfortheircontributions.

2

Contents

1Introduction 4

2ApplicationsIntegratingAIand5GMobileSystems 5

2.1StatusQuoofAIfor5G 5

2.2StatusQuoof5GforAI 5

2.3StatusQuoof5GNetworkAI 6

2.4ChallengesinIntegratingAIand5GMobileSystems 10

3MechanismofNativeAIintheNew6GMobileSystem 11

3.1NecessityAnalysis 11

3.2FeasibilityAnalysis 11

3.3GainAnalysis 12

3.4AnalysisofNewFeaturesandNewParadigms 13

4TechnicalRequirementsofNativeAIintheNew6GMobileSystem 15

4.1CapabilityRequirements 15

4.1.1ComputingPowerRequirements 15

4.1.2AlgorithmRequirements 16

4.1.3DataRequirements 17

4.1.4OtherRequirements 17

4.2ServiceRequirements 18

4.2.1ComputingServiceRequirements 18

4.2.2AlgorithmServiceRequirements 19

4.2.3DataServiceRequirements 20

4.2.4OtherServiceRequirements 21

4.3ArchitectureRequirements 22

4.3.1ComputingArchitecture 22

4.3.2AlgorithmArchitecture 25

4.3.3DataArchitecture 26

4.3.4OtherArchitectures 28

5SummaryofTechnicalRequirementsandPrinciples 29

References 30

AcronymsandAbbreviations 31

3

Authors

Contributor

Organization

LiYang,FengXie,HonghuiKang,YanXue,MenghanWang,JiaohongNiu

ZTECorporation

MingAi,XiaoyanDuan,WanfeiSun,MinShu

CICTMobile

ChenghuiPeng,ZheLiu,JunWang,FeiWang

HuaweiTechnologiesCo.,Ltd.

GangLi,ZiruiWen

ChinaMobile

XuXia,MenghanYu,HengWang,WenQi

ChinaTelecom

BingmingHuang,JunLiao

ChinaUnicom

YangYang,LiantaoWu,KaiLi,ChenyuGong,MuleiMa

ShanghaiTechUniversity

ChengchaoLiang,LunTang,RongChai,GuozhongWang

ChongqingUniversityofPostsand

Telecommunications

GangShen,ChenhuiYe,KaibinZhang,

FangfangGu

NokiaShanghaiBell

YannanYuan,BuleSun

vivoMobileCommunicationCo.,Ltd.

LanlanLi,JianjieYou

PurpleMountainLaboratories

YeOuyang,DaWang,YangBai,YanZhao

AsiaInfoTechnologiesCo.,Ltd.

LingSu,DandanHao

Ericsson

TingtingYang,JiahongNing,ZhengqiCui

DalianMaritimeUniversity/PengCheng

Laboratory

1Introduction

4

NativeAIisconsideredtobeoneofthecorearchitecturefeaturesofthefuture6Gmobilesystem[1].WiththerapiddevelopmentofAItechnologiesandapplications,theirscientificconcepts,paradigms/modes,algorithms/models,andmethods/meansneedtobemorecloselyanddeeplyembeddedintothearchitecture,NEs,andfunctionalprocessesofthenew6Gmobilesystem.Inaddition,theyneedtoleveragetheadvantagesofthenew6Gmobilesystemasapowerfulandubiquitoustelecominfrastructureplatformsoastofurthermanifesttheperformance/effectivenessgainsandbenefitsofAI,andcomprehensivelyhelprealizethevisionsoftheintelligentconnectionofeverything,intelligentassociationofservicesandtasks,andinclusiveintelligenceinthe6Gera.ThiswhitepaperbrieflyintroducesthebackgroundandmotivationsofnativeAI,andsystematicallyelaboratesnumeroustechnicalrequirements(intermsoffunction,performance,service,architecture,etc.)involvedinthedesignofnativeAIforthenew6Gmobilesystem,coveringthreebasicAIelements:AIcomputingpower,AIalgorithm,andAIdata.Further,thispaperfurtheranalyzesandsummarizesthecomprehensiveimpactsandrequirementsofnativeAIonthearchitectureandstandardizationofthenew6Gmobilesystembasedonthelatestresearch.

2ApplicationsIntegratingAIand5GMobileSystems

5

AI,especiallymachinelearning(ML)anddeeplearning(DL),wasfirstintegratedwiththe5Gsystem(5GS)accordingto3GPPspecifications.However,AIserviceapplicationsandrelatedcapabilitiesandservices,forexample,howtooptimizeaspecificcommunicationworkingmechanismbasedonnewAIparadigms,werenotfullyconsideredduringthedesign(referredtoasthenativephase)ofthe5GS.The5GSmobilesystemhasshowntendencytowardcloudnative,software-definednetworking(SDN),andservicevirtualization.Forexample,itsupportsServiceBasedArchitecture(SBA)CNandcloud-basedCN.Inaddition,allianceorganizationssuchasO-RANandOpenRANhavebeenactivelypromotingtheopenness,cloudification,andvirtualizationofwirelessnetworks.Nevertheless,theNextGenerationRadioAccessNetwork(NG-RAN)subsystemofthe5GSretainsthetraditional"siloedCT-basedbasestationarchitecture"andrelativelyrigidRANprotocolstackmodelsduetorestrictionsintermsoftechnologymaturity,security,andsystemO&Mcomplexity.WiththegradualpenetrationandprovenbenefitsofAIfunctionsandservices,the5GScanaddthenetworkdataanalyticsfunction(NWDAF)—anewlogicalfunctionnode—andvariousmodule-level"externalplug-in/add-on"AIfunctionstotheCNbasedontheexistingsystemarchitectureandprotocolstacktofurtherenhancesystemperformanceandexternalservicecapabilities.Each"externalplug-in/add-on"AIfunctionalmoduleisdevelopedforspecificknowncommunicationissue,suchasslicequalityassurance,qualityofexperience(QoE)optimization,mobilityprediction,faultlocating,andnetworkplanning,optimization,andO&M.TheseAIfunctionalmodulesareaddedtoimprovetheperformanceofthe5GS,optimizewirelesstransmissionefficiency,andsimplifynetworkmanagementandO&M.

2.1StatusQuoofAIfor5G

ManypracticeshaveproventhatAI,ML,andDLarefeasibleandeffectivemethodsforsolvingmulti-dimensionalcomplexandcomputing-intensiveproblemsintraditionalwirelesscommunication.Currently,AI,ML,andDLhavebeenpreliminarilyappliedandverifiedatmultiplelevelsandinmultiplebusinessdomainsofthe5GS.Prominentexamplesinclude:1)Anintelligentplatformhelpssimplify5GSO&M,optimizefunctionssuchasservicescenarioidentification,networkanomalydetection,andfaultrootcauseanalysis,andincreaseenergysaving.2)Anintelligentmoduleoptimizesthedeploymentofnetworkpoliciesandresourcesandimprovestheaccuracyofparametersettings.3)Intelligentchannelstateinformation(CSI)compressionisusedtoprovidefeedbackonresourceoverheads.4)Intelligentmodulationandcodinghelpimproveairinterfaceresourceutilization.5)Userexperienceisimprovedbyintelligentlyidentifyingandpredictingnetworktrafficdistribution,usertrajectory,anduserbehavior[2].AlltheprecedingexamplescanbecalledAIforNetwork(AI4NET).TheydemonstratethevalueofAIasanadvancedtechnologytoempowerandoptimizethe5GSinspecificaspects.AIalgorithms/modelsinthe5GSareappliedtosolvespecificknowncommunicationissues.Thatis,thesemodelshaveundergonealotofdedicatedofflinetrainingandverification.Althoughtheyonlyhelpmitigateorsolvespecificissues,theyalsoimprovecommunicationserviceperformanceandreducesystemO&Mcoststoacertainextent.Ingeneral,thecurrentAI4NETstillneedsimprovementintermsofsystematicness,thoroughness,globality,andexplainability.TheAIcapabilityextensibility,iterationenhancement,andAImodelgeneralizabilityaregreatlyrestricted.Inaddition,variousAIresourcesandcapabilities(coveringAIcomputingpower,algorithms,anddata)inthecurrent5GSarenotopenorservice-oriented.Mostofthemareusedonlywithinthesystem.

2.2StatusQuoof5GforAI

Currently,variousmobileAIapplications(e.g.,AIrecognition,classification,andprocessingofvoice,images,videos,anddata)areprovidedforendusersinlocalizedorcentralizedcloudAIservicemode(e.g.,thepubliccloudsrepresentedbyAmazonAWS,MicrosoftAzure,GoogleCloud,andAlibabaCloud).Therefore,the5GSactsmoreasanunderlyingdatatransmissionpipe.The5GStransmitsdataflowsrelatedtoAImodeltrainingandAIserviceapplicationsascommonuserservicedata.ThelargeamountsofsampledatarequiredforcloudAImodeltrainingaretransmittedintheformofapplication-layerdatainthe5GSinanE2Emanner.Inaddition,3GPPRelease16hasdefinedtheNWDAFandrelatedinteraction

6

interfacestoimplementintelligentapplicationsinsidethenetworkandempowerAIinternallyandexternally,therebyreducingthedependencyontraditionalcloudAI.AlltheprecedingexamplescanbecalledNetworkforAI(NET4AI),whichreflectsthebenefitsofthe5GStoAIbusinessesandservices.ForNET4AIintelligentapplications,therequirementsonKPIsrelatedtonetworkconnectionperformanceinspecificscenarioshavebeenspecifiedfrom3GPPTS22.261[3].However,intheprecedingNET4AImobileapplicationexamples,alloperationsinupper-layercloudAImodearealmosttransparenttothe5GS,andcross-layerdeepintegrationorcooperationbetweenupper-layerAIapplicationsandlower-layernetworkpipesisnotrealized.Inaddition,thecommunication,sensing,andcomputingresources,capabilities,anddatainthe5GSarenotfullyopentoandutilizedbyintelligentNEssuchastheupper-layercloudAIapplicationserversorNWDAFs,andarenotwelladaptedtoupper-layerAIserviceapplications.Thecomputingpower,algorithm,anddataresourcesownedbyeachlogicalNEnodeinthe5GSarenotfullyscheduledorutilizedbyexternalAIfunctionentities.Mostoftheseresources,suchaslargeamountsofbasebandcomputingpowerandwirelessdatainbasestations,serveonlytraditionalcommunication—ifnotleftidle.Itisexpectedthatthenew6Gmobilesystemwillnativelysupportthecapabilitiesoffullyutilizingthecomputingpower,algorithm,anddataresourcesandcapabilitiesofnew6Gnetworks,maximizingthevalueoftheseresourcestobusinessesandservices,andfurtherexpandingtheprofitabilitysystemofmobileoperators.

2.3StatusQuoof5GNetworkAI

AlthoughAIwasnotnativelyconsideredduring5Gnetworkdesign,5GNetworkAI,thatis,integratingAIand5Gnetworksformutualbenefits,hasbeencontinuouslyexploredandcontinuestoevolve.Standardsorganizationsandforumssuchas3GPP,ITU-T,ETSI,andTMFhaveprogressedintheexplorationof5GNetworkAI.

3GPPhasmadepreliminaryexplorationsinto5GNetworkAI.Thestatusandprogressareasfollows:

In3GPPRelease16,anewlogicalfunctionentity,NWDAF,wasaddedtothenetworkarchitectureofthe5GS,asshowninFigure2-1.Fordetails,see3GPPTS23.288[4].NWDAFsinteractwithotherfunctionentitiesinthe5GC(e.g.,AF,PCF,AMF,andSMF)toprovidemultipletypesofnetworkdataanalysisservices.Theofferedservicesincludereceivingnetworkdataanalysisrequestsfromeachnetworkfunction(NF)entityintheCN,collectingtherequestednetworkdata,leveragingAIalgorithmsfordataanalysisandinference,andreturningthenetworkanalysisresultinformationtotherequestingNFentity.EachNFmonitorstheoperatingstatusesofthe5GSnetworkanddevicesbasedonthenetworkanalysisresultinformationprovidedbyNWDAFs,andperformsclosed-loopcontrolandoptimizationoncommunicationservices.Asof3GPPRelease17,NWDAFscananalyzenetworkserviceexperience,networkperformance,sliceload,NFload,UEmobility/communication/anomalyevents,qualityofservice(QoS)sustainability,anduserdatacongestion.

Figure2-1NWDAF-based5Gnetworkdataanalysisarchitecture(See3GPPTS23.288.)

Thearchitectureandfunctionsof5GSnetworkdataanalysiswereenhancedin3GPPRelease17,including:logicalfunctionsplittingandinteractionwithinanNWDAF,collaborativedatatrainingandmodelsharingamongmultipleNWDAFinstances,andadditionofnewfunctionentitiestoimprovedatacollectionefficiencyandenhancereal-timeperformance.TheNWDAFspecifiedin3GPPRelease17hasthefollowingfeatures:

7

.NWDAFentitydeploymentismoreflexible.Thecentralized,distributed,andhybrid(centralized+distributed)deploymentmodesaresupported.

.MultipleNWDAFentitiescancollaboratewitheachother(e.g.,duringanalysisaggregation,analysistransfer,andAIdata/modelsharing).

.TheNWDAFcanbefurtherdecomposedintomodeltraininglogicfunction(MTLF)andanalyticlogicfunction(AnLF).AnMTLFcanprovideMLmodelsforNWDAFentitiesotherthantheNWDAFtowhichtheMTLFbelongs.

.Thedatacollectioncontrolfunction(DCCF),analyticdatarepositoryfunction(ADRF),andrelateddatacollectionoptimizationprocessesareadded.

.NWDAFentitiescancollectdatafromUEs.

.Edgecomputingserviceexperienceandnetworkperformancecanbeanalyzed.

.UEsandsessionscanbeanalyzedfromthefollowingaspects:sliceloads,discretedistributionofdata/signaling,WLANperformance,userplaneperformance,sessioncongestioncontrol,andredundanttransmission.

3GPPRelease17alsoresearchedonandstandardizedtheintelligenceofthe5GSmanagementplane.Themanagementdataanalytics(MDA)functionwasadded,asshowninFigure2-2.Fordetails,see3GPPTS28.104[5].MDAcollectsdatarelatedtonetworkandserviceeventsandstatus.Suchdataincludesnetworkperformancemeasurementdata,trace/MinimizationofDriveTests(MDT)/QoEreports,alarms,configurationdata,networkanalysisdata,andAFserviceexperiencedata.MDAperformsdataanalysisbasedonspecificAIalgorithms,generatesanalysisresultreports,andperformsnetworkmanagementoperationsbasedontheanalysisresultreports,therebyimplementingautomaticandintelligentnetworkmanagementandO&M.

Figure2-2MDAfunctionandserviceframework(See3GPPTS28.104.)

3GPPRelease17isalsoresearchingonRANintelligence.Fordetails,see3GPPTR37.817[6].Thepotentialapplicationscenariosincludenetworkenergysaving,loadbalancing,andUEmobilitymanagement.

8

3GPPRelease18continuestheresearchontheperformancerequirementsofAI/MLmodeldatatransmissioninthe5GS.Fordetails,see3GPPTS22.261[3].The5GSprovidesQoSguaranteeforAI/MLmodeldatatransmissionbasedonAI/MLserviceorapplicationrequirements.The5GScanalsomonitorthetransmissionstatus(e.g.,transmissionrate,latency,andreliability)ofAI/MLmodeldataandreportthestatustoAIapplicationserverssothattheserversadjustAIapplication-layerparametersaccordingly.

Currently,3GPPRelease18isstillintheinitiationphasein3GPPSA2.Itconsidersnetworkintelligencefromthefollowingtwoaspects:

.AI4NET:focuseson5GCNE-relatedanalysisandresearchonpotentialarchitectureenhancements,newscenarios,etc.TheresearchonAI4NETincludeswhetherandhowtoenhancethe5GCarchitecturetosupportfederatedlearningandonlinelearning,reportUPFservicedatatotheNWDAFforintelligentanalysis,utilizetheanalysissuggestionsfromtheNWDAF,enhancedatacollectionandstorage,supportNWDAF-assistedUErouteselectionpolicy(URSP),etc.

.NET4AI:focusesontheperformancerequirementsofAI/MLmodeldatatransmissionin3GPPRelease18inSA1,explores5GS-assistedAI/MLservicetransmission,supportsAI/MLmodeldistribution,transfer,andtraining,andappliestodifferentAIapplications,suchasvideo/speechrecognition,robotcontrol,andautomobiles.Theresearchalsocoversthepossiblearchitectureandfunctionextensionsthatsupportapplication-layerAI/ML,possibleQoSandpolicyenhancements,andhowthe5GSassistsfederatedlearningbetweenUEclientsandASs.

ITU-TSG13hasalsoexploredtheintegrationof5GnetworksandAI.InNovember2017,itestablishedtheFocusGrouponMachineLearningforFutureNetworksincluding5G(FGML5G).InJuly2020,theFGML5Gconcludedtheworkofthesecondphaseandsubmitted10technicalspecificationstoSG13,includingspecificationsonAIusecases,architecturalframework,intelligencelevels,dataprocessing,MLfunctionorchestrator,serviceframework,etc.Inaddition,theFGML5GproposedanML-orientedmanagementsubsystem,andtheconceptofmulti-domain,multi-cloud,andmulti-levelMLworkflowsbasedonthefunctionsrequiredindifferentphasesoftheMLlifecycle.

ETSIhasalsoexploredtheintegrationofICTsystemsandAI.AsearlyasFebruary2017,ETSIestablishedtheExperientialNetworkedIntelligence(ENI)IndustrySpecificationGroup(ISG).TheENIISGdefinesanAIenginethatprovidesintelligentservicesforapplicationssuchasnetworkO&M,serviceorchestration,andnetworkassurance.Figure2-3showsthefunctionalarchitectureofENI[7].

9

Figure2-3FunctionalarchitectureofENI

Currently,theENIsystemcontainsknowledgemanagement,modelmanagement,policymanagement,andothermodules.AfterdataisprocessedbyAImodules,theENIsystemcanautomaticallyprovideserviceoperationandassurance,slicemanagement,andresourceorchestrationfornetworks.ENIfunctionsarecontinuouslyevolving.Forexample,theENIsystemnowsupportsintent-drivennetworks.

TMFfocusesonOSSs/BSSswhenexploringtheintegrationof5GnetworksandAI.Currently,TMFiscarryingoutanAIandDataAnalytics(AI&DA)project,whichfocusesonthearchitecture,usecases,AIterminology,dataprocessing,andAItraining.Table2-1describestheresearchdirectionsandspecificworkofthisproject[8].

Table2-1MajorworkinvolvedintheAI&DAprojectofTMF

10

2.4ChallengesinIntegratingAIand5GMobileSystems

The5GSnetworkisexpectedtobetterutilizeAIcapabilitiestoenhancethenetworkitselfandsupportAIserviceapplications,firstintheCNandthenintheRAN.TheNWDAFintroducedin3GPPRelease16aimstoimproveAIdatacollectionandanalysiscapabilities.Forexample,itcanprovideanalysisresultinformationforothercoreNFsandUEs,helpingoptimizenetworkserviceprovisioning.TheNWDAFalsosupportsdatacollectionfromthenetworkO&Mandmanagementsysteminthe5GS,andprovidesdedicatedservicesforregisteringNFsandopeningmetadata.Evenso,theintegrationofthe5GSandAIfacesthefollowingchallenges:

.Limiteddatasources:NWDAFentitiesmainlycollectandanalyzethedatareceivedby5GCNFs.However,inabroadsense,thedatafromthewirelessinfrastructure,environment,devices,andvarioussensorsisnotfullyconsidered.Asaresult,AIdatasamplesareincomplete.

.Transmissionbandwidthconsumption:AlargenumberoftransmissionbandwidthresourcesareconsumedtocollectAIdataregardlessofcentralizedordistributedNWDAFdeployment.IfdatasourcesarefarfromNWDAFentities,theremaybelatencyindataupdates.

.Lackofdataprivacyprotection:TheNWDAFcollectsandanalyzesdatainacentralizedmanner.Datasourcesusuallycomefromthesameservicedomain.Therefore,dataprivacyprotectionisnotfullyconsideredinarchitecturedesign,anduserprivacymaybecompromised.

.LackofsupportforexternalAIservices:Asaninternalfunctionofthe5GC,theNWDAFismainlyusedtoenhanceandoptimizethe5GS.ExternalAIapplicationscannotbedirectlyservedbyorbenefitfromtheAIfunctionsofthe5GCorRAN.

.Insufficientinfrastructureutilization:Key5Gfunctions,suchasnetworkslicing,ultra-reliablelow-latencycommunication(URLLC),andmassivemachine-typecommunications(mMTC),aredesignedtomeetverticalindustryrequirementsonperformance,function,andoperation.However,thesupportfornativeAI(datagovernanceandservices,distributedarchitecture,etc.)isnotspecificallyconsideredduringthearchitecturedesignofthesekey5Gfunctions.Variousresources(thathavelowutilizationorareinanidlestate)inthenetworkinfrastructurearenotfullyutilizedbyserviceapplicationsandthevalueoftheseresourcesisnoteffectivelyrealized.

.Lackofdatagovernanceandservices:AIinvolvesmorethandatacollectionandtraining&analysisandinference.Tosupport6GnativeAI,thearchitectureofAIdatagovernanceandservicesneedstobesystematicallydesigned.Thisapproachwasnotconsideredduringthedesignofthe5GS.Thevalueofdataservicesneedstobefurthermanifestedinthefuture.

3MechanismofNativeAIintheNew6GMobileSystem

11

Theoperationdatainformationcommunicationtechnology(ODICT)industrybelievesthatnativeAIwillbecomeoneofthecorefeaturesofthenew6Gmobilesysteminthefuture[1].Therefore,deepintegrationwithAIiscomprehensivelyconsideredintherequirementandarchitecturedesignphasesofthenew6Gmobilesystem.DifferentfromaddingAIfunctionstothe5GSbymeansofadd-ons,patches,andplug-ins,thegoalofnativeAIwillbringmanyprofoundimpactsandchallengestothedesignofthenew6Gmobilesystem.Thissectionwillfirstelaboratethemotivationsandreasonsforimplementing6GnativeAI(e.g.,whynativeAIcanbetteradapttonew6Gscenariosandusecasesinthefuture,andhowtoformnewbusinessesandgeneratenewvalue).

3.1NecessityAnalysis

Withtheevolutiontowardfuturemobilenetworks,networkmanagementandO&Mneedtotransformfromlocalintelligentoperationsforlowercostandhigherefficiency,toE2Ehigh-levelnetworkautonomy.However,theR&DofexistingAIusecasesisgenerallyperformedbymeansofpatches,externalplug-ins,andsiloedadd-ons,andlacksaunifiedsystemframework.MostAImodelslackthemeanstopre-verifytheeffectofapplyingthemodelsandQoSassurance.AImodeltrainingandlearningaredecoupledfromanalysisandinferenceinthattheyarenotperformedatthesametime.TheeffectofanAImodelcanbeverifiedonlyafterAIinferenceisfinished,andthiscanhaveagreatimpactonthelivenetwork.Asaresult,high-levelnetworkautonomyisimpossible,andAImodelscannotbepre-verified,evaluatedonline,orfastoptimizedinanautomaticclosed-loopmanner.Inaddition,AImodeltraining/retrainingrequiresalargeamountofsampledata,butitisdifficulttocollectdatainacentralizedmanner.Thisleadstoanumberofissues,forexample,highnetworktransmissionoverheadandtrainingoverhead,longiterationandupdatetime,slowconvergence,andpoorgeneralizabilityofAImodels.Therefore,thenew6GmobilesystemneedstofurtherenhancethetrainingandapplicationperformanceofAImodelssoastoimprovethenetworkautonomylevel.

ThekeydrivingforcesofnativeAIincludeassistingvariousverticalindustriesinintelligentdigitaltransformation,exploringnewintelligentbusinessmodels,andprovidingnew6Gscenariosandcapabilities.Onthepremiseofprotectingdataprivacyofverticalindustriesandpreventingdatafrombeingtransferredoutofcampuses,thenew6Gmobilesystemmustbeabletoprovidedistributedandregionalcomputingresources,platforms,andservicestoflexiblyprovisionintelligencecapabilitiesondemandanytime,anywhere,andenabledata-centriccomputing.ComparedwithtraditionalcloudAIserviceproviders,6GnativeAIcanprovideintelligencecapabilitiesandserviceswithhighertimeliness,privacy,andperformance.Inaddition,6GnativeAIcanprovideinter-industryfederatedintelligencetofacilitatecross-domainandcross-industryintelligenceconvergenceandintelligentdigitalsharing.

Withtheevolutionofintelligentdevicesinthefuture,massivedeviceswillgeneratemoredata,andthecomputingandintelligencecapabilitiesofdeviceswillbecomeincreasinglystronger.6GnativeAIneedstocoordinatenetworkAIanddeviceAItoprovideToCuserswiththeultimateserviceexperienceandhigher-valuenoveldatainformationcommunicationtechnology(DICT)services.Ensuringthesecurityandtrustworthinessoffuturenetworksisalsoanimportantresearchtopic.NativeAIcanpromotetherealizationofnativenetworksecurityandtrustworthiness,andautonomouslydetectanddefendagainstvarioustypesofpotentialattacksandthreats.

3.2FeasibilityAnalysis

Thesuccessfulintegrationofthe5GSandAIhasproventhefeasibilityofimplementingnetworkintelligencetosomeextent.ThissectionanalyzesthefeasibilityofdeeplyintegratingnativeAIandthenew6GmobilesystemfromtheperspectiveofthreekeyAIelements.

.Computingpower:Duetotherequirementsonlatency,reliability,datasecurity,andprivacyprotection,deployingcomputingcapabilitiesandresourcesonthenetworkedgeshasbecomeamajortrendinthe5Gera.Inthe6Gera,dataconnectionandcomputingmaybefurtherintegrated.

12

Forexample,adual-infrastructurethatintegratesbothconnectionandcomputingmayemerge,whichcanprovideacomputingservice-relatedbasisforintegratednativeAIdesign.

.Algorithm:AlthoughcentralizeddataprocessingontheAIcloudhasitsadvantages,issuesrelatedtodataprivacy,highperformance,andcomputingenergyconsumptionremainchallenging.Ifthenew6Gmobilesystemcanintegratealgorithms/modelsandintelligencecapabilitiesintothenetwork,th