新加坡《合成數(shù)據(jù)生成指南》

PRIVACYENHANCINGTECHNOLOGY(PET):

PROPOSEDGUIDEONSYNTHETICDATA

GENERATION

Published15July2024VersionNumber1.0

JOINTLYDEVELOPEDWITHSUPPORTEDBY

2

TABLEOFCONTENTS

I.IntroductiontoPrivacyEnhancingTechnology(PET) 3

II.SyntheticData 4

WhatisSyntheticData? 5

UnderWhatCircumstancesisSyntheticDataUseful? 6

CaseStudies 8

III.Recommendations 10

AnnexA:HandbookonKeyConsiderationsandBestPracticesin

SyntheticDataGeneration 11

AnnexB:DataDictionaryFormat 24

AnnexC:ExamplesofMethodsofSyntheticDataGeneration 27

AnnexD:Re-identificationRisks 33

AnnexE:ExamplesofApproachesinEvaluationofRe-identification

Risks 35

ACKNOWLEDGEMENTS 41

3

I.IntroductiontoPrivacyEnhancingTechnology(PET)

PrivacyEnhancingTechnologies(PETs)areasuiteoftoolsandtechniquesthatallowtheprocessing,analysis,andextractionofinsightsfromdatawithoutrevealingtheunderlyingpersonalorcommerciallysensitivedata.ByincorporatingPETs,companiescanmaintainacompetitiveedgeinthemarketthroughleveragingtheirexistingdataassetsforinnovationwhilecomplyingwithdataprotectionregulations,reducingtheriskofdatabreachesanddemonstratingacommitmenttodataprotection.PETsarenotjustadefensivemeasure;theyareaproactivesteptowardsfosteringacultureofdataprotectionandsecuringacompany'sreputationinthedigitalage.

PETscangenerallybeclassifiedintothreekeycategorie

s1:

dataobfuscation,encrypteddataprocessing,andfederatedanalytics.PETscanalsobecombinedtoaddressvaryingneedsoforganisations.ThefollowingTable1mapsoutthecurrenttypesofPETsinthemarketandtheirkeyapplications.

Table1.TypesofPETsandtheirapplications

Categoriesof

PETs

PETs

Examplesofapplications(non-exhaustive)

Data

obfuscation

Anonymisation/pseudonymisationtechniques

?Securestorage

?Datasharingandretention

?Softwaretesting

Syntheticdatageneration

?Privacy-preservingAImachinelearning

?Datasharingandanalysis

?Softwaretesting

Differentialprivacy

?Expandingresearchopportunities

?Datasharing

Zeroknowledgeproofs

?Verifyinginformationwithout

requiringdisclosure(e.g.,ageverification)

Encrypteddataprocessing

Homomorphicencryption

?Securedatastoredincloud

1AdaptedfromOECD,“EmergingPrivacyEnhancingTechnologies:CurrentRegulatoryandPolicyApproaches,”O(jiān)ECDDigitalEconomyPapers(OECD,2023).

4

?Computingonprivate

datathatisnotdisclosed

Multi-partycomputation

(includingprivatesetintersection)

?Computingonprivate

datathatisnotdisclosed

Trustedexecutionenvironments

?Computingusing

modelsthatneedtoremainprivate

?Computingonprivate

datathatisnotdisclosed

Federatedanalytics

Federatedlearning

?Privacy-preservingAImachinelearning

Distributedanalysis

II.SyntheticData

Thisguidefocusesontheuseofsyntheticdata

2

togeneratestructureddata.Whilesyntheticdataisgenerallyfictitiousdatathatmaynotbeconsideredpersonaldataonitsown,itisnotinherentlyrisk-freeduetopossiblere-identificationrisk

s3.

Assuch,thisguideproposesgoodpracticesthatorganisationsmayadopttogeneratesyntheticdatatominimisesuchrisksforasetofcommonusecasearchetypes.Theguidealsoincludesasetofgoodpracticesandriskassessments/considerationsforgeneratingsyntheticdataaswellasgovernancecontrols,contractualprocess,andtechnicalmeasurestomitigateresidualrisks.

ThetargetaudienceforthisguideareCIOs,CTOs,CDOs,datascientists,dataprotectionpractitioners,andtechnicaldecision-makerswhomaydirectlyorindirectlybeinvolvedinthegenerationanduseofsyntheticdata.

Syntheticdataisatechnologythatisbeingactivelyresearchedanddevelopedatthetimeofpublication.Hence,thisguideisnotintendedtoprovideacomprehensiveorin-depthreviewofthetechnologyoritsassessmentmethods.Theguideisintendedtobealivingdocument,andwillbeupdatedtoensureitsrecommendationsremainrelevant.

2Therearetwotypesofsyntheticdata:fullysyntheticdataandpartiallysyntheticdata.Thisguidediscussestheuseoffullysyntheticdata.

3Inthisguide,wegenerallyrefertoprivacyrisksasre-identificationrisks.

5

WhatisSyntheticData?

Syntheticdataiscommonlyreferredtoasartificialdatathathasbeengeneratedusingapurpose-builtmathematicalmodel(includingartificialintelligence(AI)/machinelearning(ML)models)oralgorithm.Itcanbederivedbytrainingamodel(oralgorithm)onasourcedatasettomimicthecharacteristicsandstructureofthesourcedata.Goodqualitysyntheticdatacanretainthestatisticalpropertiesandpatternsofthesourcedatatoahighextent.Asaresult,performinganalysisonsyntheticdatacanproduceresultssimilartothoseyieldedwithsourcedata.

Characteristicsofsyntheticdata

Figure1showsanexampleofhowsyntheticdatamaylooklikeascomparedwiththesourcedata.Generatedsyntheticdatawillgenerallyhavedifferentdatapointsfromthesourcedata,asseenfromthetabulardata.However,thesyntheticdatawillhavestatisticalpropertiesthatareclosetothatofthesourcedata,i.e.,capturingthedistributionandstructureofthesourcedataasseenfromthetrendlinesinFigure1.

Figure1:Sourcedataversussyntheticdata

.4

Assuch,syntheticdatamaynotalwaysbeinherentlyrisk-freeasinformationaboutanindividualinthesourcedataset,orconfidentialdata,canstillbeleakedduetotheresemblanceofthesyntheticdatatothesourcedata.Therewillalsobetrade-off

s5

betweendatautilityanddataprotectionrisksinsyntheticdatageneration.However,suchriskscanbeminimisedbytakingdataprotectionintoconsiderationduringthesyntheticdatagenerationprocess.

4DiagramtakenwithmodificationfromKhaledElEmam,LucyMosquera,andRichardHoptroff,PracticalSyntheticDataGeneration(O’ReillyMedia,Inc,2020).

5Trade-offbetweendatautilityanddataprotectionrisksisfurtherdiscussedinAnnexA:Step1andStep

3inthisguide.

6

UnderWhatCircumstancesisSyntheticDataUseful?

SyntheticdatacanbeusedinavarietyofusecasesrangingfromgeneratingtrainingdatasetsforAImodelstodataanalysisandcollaboration.Theuseofsyntheticdatanotonlycanaccelerateresearch,innovation,collaboration,anddecision-makingbutalsomitigateconcernsaboutcybersecurityincidentsanddatabreaches,enablingbettercompliancewithdataprotection/privacyregulations.Table2discussesafewcommonusecasearchetypes,theirkeybenefits,andgoodpracticesthatorganisationscanfocusonwhengeneratingsyntheticdata.

Table2.Usecasearchetypesforsyntheticdata.

TypesofUseCases

KeyBenefits

GoodPracticesto

GenerateSyntheticData

Usecasearchetype1:GeneratingtrainingdatasetforAImodels

Augmenting

dataforAI/MLmodels

?Syntheticdataaddressesthechallengeoftheuserhavingtoobtainlargevolumesof

labelleddataneededfor

trainingandtestingAI/MLmodelsduetocosts,legalregulations,andproprietaryrights.

?Augmentingtrainingdatasetswithsyntheticallygeneratedlabelleddatacanbemore

cost-effective,especially

whenthesourcedatasetsaresparse.

?Addnoise*toorreducegranularityofthe

syntheticdatapoints.

?Suchfictitiousnewdatapointswillgenerallynotbeconsideredpersonaldata.

*Ifthestatistical

properties/characteristicsofthesyntheticdatais

representativeofthe

populationinquestionandnotsignificantlyskewed

towardsaspecific

individual/groupof

individualsusedassourcetrainingdata,addingof

noisemightnotbe

necessaryasre-

identificationrisksaregenerallylow.

Increasing

datadiversityforAI/ML

models

?Syntheticdatacanbeusedtosimulaterareeventsor

augmentunder-representedgroupsintrainingAImodels.

?DiversedatasetscanbeusefulinimprovingperformanceofAI/MLmodels

Usecasearchetype2:Dataanalysisandcollaboration

Datasharingandanalysis

?Underlyingtrendsorpatterns,andbiasesofthedataare

usefulfordataanalytics

regardlessofwhetherthedatasourceisrealorsynthetic.

?Balancethetrade-offs

betweendatautilityanddataprotectionby

incorporatingdataprotectionmeasures

7

?Syntheticdatacanenabledatasharingforanalysisespeciallyinindustriesandsectors,e.g.,healthcare,wherethesourcedatacanbesensitive.

throughoutthesyntheticdatagenerationprocess,forexample:

Datapreparation

?Removeoutliersfromsourcedata

?Pseudonymisesourcedata

?Employdata

minimisationand

generalisegranulardata

Syntheticdatageneration

?Addnoisebeforeoraftersyntheticdatageneration

Postsyntheticdatageneration

?Incorporatetechnical,

contractual,and

governancemeasurestomitigateanyresidualre-identificationrisks

Previewing

datafor

collaboration

?Syntheticdatacanbeusedindataexploration,analysis,andcollaborationtoprovide

stakeholderswitha

representativepreviewofthesourcedatawithoutexposingsensitiveinformation.

?Thisenablesstakeholdersto

exploreandunderstandthe

structure,relationships,and

potentialinsightswithinthe

datatogainassuranceofthedataqualitybeforefinalisinganyagreementor

collaboration.

Usecasearchetype3:Softwaretesting

System

development/software

testing

?Organisationscanuse

syntheticdatainsteadof

productiondatatofacilitatesoftwaredevelopment.

?Useofsyntheticdatacanhelporganisationsavoiddata

breachesintheeventofthedevelopmentenvironmentbeingcompromised.

?Focusongenerating

syntheticdatathat

followssemanticse.g.,format,min/maxvaluesandcategories,of

sourcedatainsteadofthestatistical

characteristicsandproperties.

RefertoAnnexAforproposedconsiderationsandgoodpracticestogeneratesyntheticdata.

8

CaseStudies

(A)TrainingAImodelforfrauddetectioninthefinancialsecto

r6

Problem:Sincethenumberoffraudulenttransactionsinthesourcedataissmallcomparedtonormal,non-fraudulenttransactions,thesourcedatadidnottrainmodelsverywellforfrauddetection.

Solution:J.P.Morgansuccessfullyusedsyntheticdataforfrauddetectionmodeltraining.AImodelswereprovidedwithsamplesofnormalandfraudulenttransactionstounderstandthetell-talesignsofsuspicioustransactions.

Benefit:Syntheticdataprovedtobemoreeffectiveintermsoftrainingmodelstodetectanomalousbehaviour.Thisisbecausethesyntheticdatausedwasdesignedtocontainahigherpercentageoffraudulenttransactions.

(B)TrainingAImodelforresearchintoAIbia

s7

Problem:Multi-labelclassificationandregressionmodelsarefrequentlyutilisedatMastercardforvariousapplications,includingfraudprevention,anti-moneylaunderingandmarketingusecasesforportfoliooptimisation.Thesemodels,whilepowerful,requirecarefulattentiontoproxiesofdemographicattributeswithintheirtrainingdata,whichcouldlearnunintendedbiases.Ensuringtheaccuracyandfairnessofthesemodelsiscomplexduetotheirmulti-labelsetting,theconfidentialityofthedemographicattributes,andthechallengesinaccessingthetrainingdatasetformodeldevelopment.

Solution:MastercardpartneredwithresearcherstodevelopnewAIbiastestingmethodsadaptedtomulti-labelsettings.Toprotecttheprivacyofthedatasharedexternally,syntheticdatawascreatedtosupportmodeltrainingandmethodologicalresearchintofairmulti-labelmodels.

Benefit:Syntheticdatawasmeasuredtobesufficientlyprivatetobesharedwithexternalresearcherswhilecapturingrealrelationshipswithinthesourcedata.Syntheticdataenablednewinsightsthatwouldnothavebeenpossiblewithouttheprivacyprotectingcharacteristicsinherenttosyntheticdata.

6J.P.Morgan,“SyntheticDataforRealInsights,”TechnologyBlog,n.d.,

/

technology/technology-blog/synthetic-data-for-real-insights

7ContributedbyMastercard

9

(C)Safeguardingpatientdatafordataanalysis

8

Problem:Priortoutilisingsyntheticdata,Johnson&Johnson(J&J)allowedexternalresearchersorconsortiatoaccesshealthcaredataforresearchproposalsvalidatedbyJ&J.Tosafeguardpatientprivacy,thedatawastransformedintoanonymisedhealthcaredata.However,feedbackreceivedindicatedthattheoverallusefulnessoftheanonymiseddata,whichreliedontraditionalanonymisationtechniques,wasnotalwayssatisfactoryanddidnotalwaysmeettherequirementsoftheresearchersorconsortia.

Solution:J&Jhasintroducedhigh-qualityAIgeneratedsyntheticdataasanadditionaloptiontoprocesstheirhealthcaredata.

Benefit:Researchersandclientshaveexperiencedsignificantlyimprovedanalysis.Whenemployedproperly,thisformofsyntheticdatacaneffectivelyrepresentthetargetpopulationandoffervariousanalyticalandscientificbenefits.

(D)Facilitatingdata

collaboration9

Problem:Apharmaceuticalcompanywantedtopurchaseheart-relatedhealthdatafromaresearchinstitutetotestoutanewhypothesis.Thehealthdata,whichwascollectedbytheresearchinstitutefromconsentingsubjects,washostedunderahighlyregulatedenvironmentasrequiredofthehealthcaresector.However,thispresentssignificantchallengesformanydataengagementactivities.

Solution:A*STARwasengagedbythepharmaceuticalcompanytobuildapipelinetocreatesyntheticcopiesoftheactualdata,whichcanthenbebroughtoutsideofthisregulatedenvironment.

Benefit:Thisallowedthepharmaceuticalcompanytopreviewthedataandbeassuredofthedataqualitypriortothehigh-valuepurchaseandaccesstotheactualdata.

8ContributedbyJohnson&Johnson(J&J)

9ContributedbyA*STAR

10

III.Recommendations

SyntheticdatahasthepotentialtodrivethegrowthofAI/MLbyenablingAImodeltrainingwhileprotectingtheunderlyingpersonaldata.ItalsoaddressesdatasetrelatedchallengesforAImodeltraining,suchasinsufficientandbiaseddata,throughenablingtheaugmentationandincreaseddiversityoftrainingdatasets.

Inaddition,syntheticdatacanbeusedtofacilitateandsupportorganisations’dataanalytics,collaborationandsoftwaredevelopmentneeds.Anaddedbenefitofusingsyntheticdatainplaceofproductiondatatofacilitatesoftwaredevelopmentisthatdatabreachescanbeavoidedintheeventthedevelopmentenvironmentiscompromised.

PDPCrecommendsasetofgoodpracticesandriskassessments/considerationsforgeneratingsyntheticdataandtoreduceanyresidualrisksfromre-identificationthroughgovernancecontrols,contractualprocess,andtechnicalmeasures(refertoAnnexA).

11

AnnexA:HandbookonKeyConsiderationsandBestPracticesinSyntheticDataGeneration

Inthishandbook,wedescribethekeyconsiderationsandbestpracticesfororganisationstoreducere-identificationrisksofsynthetictabulardatathroughafive-stepapproach.

Foranyothercomplexsyntheticdatasetsthatareunstructured,organisationsareadvisedtoconsiderhiringsyntheticdataexperts,datascientistsorindependentriskassessorstoassessandmitigatetherisksofthegeneratedsyntheticdata.

Overviewoffive-stepapproachtogeneratesyntheticdata

Step1:Knowyourdata

Beforeembarkingonanysyntheticdataproject,itisnecessarytohaveaclearunderstandingofthepurposeandusecasesofthesyntheticdataandthesourcedatathatthesyntheticdataistomimic.Thiswillhelptodeterminewhetheruseofsyntheticdatamightberelevantandidentifythepossiblerisksofusingthesyntheticdata.Someoftheconsiderationsmayinclude:

?Wheregeneraltrends/insightsofsourcedataaresensitive,organisationshouldtakenotethattheuseofsyntheticdatawillnotofferanyprotectiontothetrends/insightssincetheywillbereplicatedinthesyntheticdata.

?Wherethesyntheticdataisintendedtobereleasedpublicly,organisationsmayhavetoprioritisedataprotectionoverdatautilityinsuchcircumstances.

12

?Whererelevant,organisationsshouldalsoputinplacepropercontractualobligationsonrecipientsofsyntheticdatawherenecessarytopreventre-identificationattacksonthedata.

Withthisknowledge,themanagementanddataowner,withthehelpofrelevantstakeholderssuchasthedataanalyticsteam,shouldestablishobjectivespriortosyntheticdatagenerationtodetermineanacceptableriskthreshol

d10

ofthegeneratedsyntheticdataandtheexpectedutilityofthedata.Thiswillhelpprovideorganisationswiththeappropriatebenchmarkstoassessanytrade-offsbetweendataprotectionrisksanddatautility.

Thesebenchmarksmaybeadjustedappropriatelytomeetthebusinessobjectives,takingintoconsiderationanytrade-offsbetweendatautilityanddataprotectionrisksafterthesyntheticdatagenerationprocess,aswellassafeguardsandcontrolstomitigateorloweranyresidualrisksposedbythegeneratedsyntheticdata.Theacceptancecriteriashouldbeincorporatedintotheorganisation'sriskassessments(e.g.,enterpriseriskmanagement

framework11

ifapplicable)oraDataProtectionImpactAssessment(“DPIA”

)12.

Step2:Prepareyourdata

Whenpreparingthesource

data13

forgeneratingsyntheticdata,itisimportanttoconsiderthefollowing:

?Whatarethekeyinsightsthatneededtobepreservedinthesyntheticdata?

?Whicharethenecessarydataattributesforthesyntheticdatatomeetthebusinessobjectives?

10There-identificationriskthresholdrepresentsthelevelofre-identificationriskthatisacceptableforagivensyntheticdataset.Thereiscurrentlynouniversallyacceptednumericalvalueforriskthreshold.ForfurtherdetailsrefertoStep4(Assessre-identificationrisks).

11OrganisationsmayrefertoISO27001formoreinformationondevelopinganenterpriseriskmanagementframework.

12AnexampleofthisisPDPC’sGuidetoDataProtectionImpactAssessments.ADPIAisapplicableinthecasewherepersonaldataisinvolved.TheDPIAmaynotberelevantinsituationswherethesyntheticdatagenerationdoesnotinvolvepersonaldataprocessing.

13Thisstepassumesthatthesourcedatahasbeenproperlycleaned(suchasfixingorremovingincorrect,corrupted,incorrectlyformatted,duplicate,orincompletedata)andisofacceptablequalityforthegenerationofsyntheticdata.

13

Understandingkeyinsightstobepreserved

Toensurethatthesyntheticdatacanmeetthebusinessobjectives,organisationsneedtounderstandandidentifythetrends,keystatisticalproperties,andattribute-relationshipsinthesourcedatathatneedtobepreservedforanalysise.g.,identifyrelationshipsbetweendemographiccharacteristicsofpopulationandtheirhealthconditions.

Organisationsshouldconsider,atthispoint,whetheroutliertrendsandinsightsarenecessarytobepreservedforthebusinessobjectives.Keyconsiderationscouldincludethefollowing:

?Ifoutliersarenotnecessarytomeetthebusinessobjectivesandtheriskofre-identificationishigh,organisationsshouldconsiderremovingtheoutliers.Thiscanbedonepriortosyntheticdatagenerationoratsubsequentstagesofthesyntheticdatageneration.

?Iftheobjectiveistomimicthecharacteristicsofthesourcedataascloselyaspossible,includingoutliers,thentheorganisationmayhavetopreservetheoutliertrend/insighttomeetthebusinessobjectives.Insuchinstance,theorganisationshouldnotethatthere-identificationrisksofindividualsintheoutlierdatamaybehighandhenceputinplaceriskmitigationmeasures.

?Ifthebusinessobjectiveistobalancethenumberofdatapointsindifferentdatacategories,thenthesyntheticdatagenerationprocessitselfcanhelpmitigatetheissueofoutlierssimplybygeneratingmoreoutliers.Forexample,inadataset,thenumberofoutlierdatapointscomprisingmaleindividualsmaybebalancedwithoutlierdatapointscomprisingfemaleindividuals.

Selectingdataattributes

Basedonthekeyinsightsneeded,organisationsshouldapplydataminimisationtoextractonlytherelevantdataattributesfromthesourcedata.Thereafter,removeorpseudonymisealldirectidentifier

s14

fromtheextracteddata.

Wheregranularinformationisnotnecessary,organisationsmaygeneraliseorfurtheraddnoisetothedataatthispointoratalatersteptoreducetheriskofre-identification.Forexample,organisationscangeneraliseexactheightandweight

14RefertoPDPC’sGuidetoBasicAnonymisationonhowtoidentifydirectidentifiersinadataset.

14

informationintoheightandweightbandstoreducethepossibilityofheightandweightcombinationsbeingusedtoidentifyanyoutliers.

Organisationsshouldalsostandardiseanddocumentthedetailsoneachdataattribute(suchasdatadefinitions,standards,metricsetc.)inadatadictionary.Thisenablestheorganisationtosubsequentlyvalidatetheintegrityofthegeneratedsyntheticdatatodetectanomaliesandfixanydatainconsistencies.RefertothefollowingchecklistinTable3forkeyconsiderations.

Table3:Checklistfordatapreparation

DataPreparationChecklist

Understandkeyinsights

i.

Identifytrendsandentityrelationshipstobepreservedforsyntheticdatageneration.

ii.

Removeoutliersifsuchtrends/insightsarenotnecessary.Thiscanbeperformedpostgeneration.

Selectdataattributes

iii.

Applydataminimisationtoselectonlydataattributesthatarenecessarytomeetbusinessneeds.

iv.

Removeorpseudonymisedirectidentifiers(e.g.,name,nationalidentificationnumbers).

v.

Generalisegranulardataoraddnoise(e.g.,usingdifferentialprivac

y15)

tothe

data/modelifsuchdetailedinformationisnotnecessary.Thiscanalsobeperformedpostgeneration.

vi.

Standardiseanddocumentformat,constraints,andcategoriesofsourcedataindatadictionary(refertoAnnexBforareferencetemplate):

Format

?Standardisestringstolowerorpropercase

?Datatypes,columnnames,structures,relationships

?FrequencyofdatarecordConstraints

?Constraintsofvaluesforeachdatatype,e.g.,min-maxvalues,non-negativevalues,non-nullvalues

Category

?Typesofdatacategories

?Expectedorvalidvaluesfordataattributeswithineachdatacategory.Exampleofadatacategoryis“country”.

15Theuseofdifferentialprivacytoaddnoisetosyntheticdataiswidelydiscussedasamechanismtoreducere-identificationrisks.However,thereiscurrentlynouniversalstandardonhowtoimplementdifferentialprivacy.Moreover,thenoiseaddedmayalsoreducetheutilityofthesyntheticdata,makingitlessaccurateorusefulforcertaintypesofanalysis.

15

Step3:Generatesyntheticdata

Therearemanydifferentmethods

16

togeneratesyntheticdata,forexample,sequentialtree-basedsynthesisers,copulas,anddeepgenerativemodels(DGMs).Organisationsneedtoconsiderwhichmethodsaremostappropriate,basedontheirusecases,dataobjectives,andtypesofdata.PleaserefertoAnnexCformoreinformationonthesesyntheticdatagenerationmethods.Thereafter,organisationsmayconsidersplittingthesourcedataintotwoseparatesetse.g.,80%astrainingdataset,and20%ascontrol

dataset17

forassessingre-identificationrisksofthesyntheticdata.

Aftergeneratingsyntheticdata,itisagoodpracticefororganisationstoperformthefollowingchecksonthequalityofthegeneratedsyntheticdata:

?Dataintegrity

?Datafidelity

?Datautility

Dataintegrity

Dataintegrityensurestheaccuracy,completeness,consistency,andvalidityofthesyntheticdataascomparedwiththesourcedata.Organisationscanvalidatetheintegrityofthegeneratedsyntheticdataagainstthedictionaryofthesourcedata.

Datafidelity

Datafidelityexaminesifsyntheticdatacloselyfollowsthecharacteristicsandstatisticalattributesofthesourcedata.Thereareafewmetricsformeasuringdatafidelityandtheyaretypicallydonebystatisticallycomparingthegeneratedsyntheticdatadirectlywiththesourcedata.Organisationsshouldusetheperformancemetric(s)fordatafidelit

y18

(seeTable4)thatbestmeettheirdataobjectives.

16ThisguidemaynotbecomprehensiveincoveringallothersyntheticdatagenerationmethodssuchasBayesianmodelandvariationalautoencoders(VAE).

17RefertoApproach2inAnnexEformoredetailsontheassessmentandevaluationframeworkforquantifyingre-identificationrisk.

18ThereareothergenericmetricsdescribedhereinadditiontothoselistedinTable4.SeeKhaledElEmametal.,

“UtilityMetricsforEvaluatingSyntheticHealthDataGenerationMethods:ValidationStudy,”

JMIRMedicalInformatics10,no.4(2022)

Table4:Performancemetricsfordatafidelity

Performancemetricsgenerallyusedforassessingdatafidelity

Histogram-basedsimilarity

Measuresthesimilaritybetweensourceandsyntheticdata’sdistributionsthroughahistogramcomparisonofeachfeature.Thisensuresthesyntheticdatapreservesimportantstatisticalpropertiessuchascentraltendency(mean,median),dispersion(variance,range),anddistributionshape(skewness,kurtosis).

Correlationalsimilarity

Measuresthepreservationofrelationshipsbetweenfeaturesinthesourceandsyntheticdatasets.Forexample,ifhighereducationtypicallyleadstohigherincomeinthesourcedata,thispatternshouldalsobeevidentinsyntheticdata.

Datautility

Datautilityreferstohowwellsyntheticdatacanreplaceoraddtosourcedataforthespecificdataobjectiveoftheorganisation.

Therearedifferentapproachestoevaluatetheutilityofsyntheticdata.Thetruetestofutilityishowitperformsinreal-worldtasks.OnecommonapproachtocheckthisisbytrainingidenticalAI/MLmodelsonsyntheticandtrainingdata.T