Contentslistsavailableat ARTICLEINPRESS1045926Xseefrontmatter2009ElsevierBVAllrightsreserved101016jjvlc200901010 CorrespondingauthorEmailaddressesrnr4csebuffaloeduRNRodriguesfe ID: 177215
Download Pdf The PPT/PDF document "Robustnessofmultimodalbiometricfusionmet..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
RobustnessofmultimodalbiometricfusionmethodsagainstspoofattacksRicardoN.Rodrigues,LeeLuanLing,VenuGovindaraju Contentslistsavailableat ARTICLEINPRESS1045-926X/$-seefrontmatter2009ElsevierB.V..Allrightsreserved.10.1016/j.jvlc.2009.01.010 Correspondingauthor.E-mailaddresses:rnr4@cse.buffalo.edu(R.N.Rodrigues)fee.unicamp.br(L.L.Ling) JournalofVisualLanguagesandComputing]]]] Pleasecitethisarticleas:R.N.Rodrigues,etal.,Robustnessofmultimodalbiometricfusionmethodsagainstspoofattacks,JournalofVisualLanguageandComputing(2009),doi:10.1016/j.jvlc.2009.01.010 spoofed,whichhasneverbeeninvestigatedbefore.Infact,thehypothesisthatanimpostormayspoofabiometricmodeisnotmodeledinanyproposedfusionmethodsofar.Somerecentworkshaveproposedtheintroductionofauxiliaryinformation,likebiometricsamplequalityuality…anduserspeci“cparameterss,inthefusionmethodtobuildmoreadaptableandreliablebiometricsystems.Thegeneralideainthesemethodsistoweight(directlyorindirectly)thecontributionofeachunimodalbiometricbasedontheauxiliaryinformation.Weexploreasimilarconceptbyincorporationinthefusionprocessanauxiliaryinformationthatindicateshowsecureeachbiometricmodeis.Ourmainobjectiveistodevelopbiometricfusionstrategiesthatarerobustagainstspoofattacksandthatarecapableofcombiningbiometricsystemswithdifferentlevelsofsecuritywithoutcompro-misingtheoverallsecurityofthemultimodalsystem.Weproposetwonovelmultimodalbiometricfusionschemesthatconsiderthespoo“nghypothesisandtakeintoaccountthesecurityofeachbiometricsystembeingfused.The“rstschemeisanextensionoftheLLRandthesecondismodeledusingfuzzylogic.Bothmodelssharethesamebasicideas,butdifferindetailsandTheproposedfusionschemesaredescribedinSection2.InSection3wedescribetheexperimentsconductedtoanalyzethefusionschemesrobustnessagainstbadqualitysamplesandspoofattack.Theresultsoftheseexperi-mentsareshowninSection4.TheconclusionsarepresentedinSection5.2.Proposedfusionschemes2.1.GeneralconceptsInthisworkweconsideronlytheveri“cationtask(i.e.theuserhasclaimedanidentityandthesystemneedstodecideiftheuserisgenuineorimpostor).Letbethenumberofbiometricsystemstobefused.Fig.1illustratesthegeneralmultimodalarchitecturewhen2.Thebiometricinformationisfusedatthematchingscorelevel.Thismeansthateachbiometricsystemindividuallyperformsamatchingbetweentheenrolledsampleandthetestsample,andcomputesasimilaritybetweenthetwosamples.Weconsiderthatforeachbiometricsystem,thereexistsanexpertthat,givenatestsample,canprovideascorethatmeasuresthequalityofthebiometricsample.Theset...formstheinputforthefusionschemethatprocesstheseinputsandproducesasinglescalaroutputsuchthathighervaluesofindicatethattheuserisgenuine(orimpostor).Athresholdoperationisappliedtotheoutputfor“nalclassi“cationbetweenimpostororgenuine.Thesecurityofeachbiometricsystemismodeledbytheparameter,whichrepresentshowdif“cultitistospoofthebiometricsystem.Itshouldbenotedthatitisveryhard(ifnotimpossible)tomeasurethesecurityofabiometricsystem[10].Inthiswork,wemanuallysettheparameterbasedongeneralknowledgeaboutthesecurityofeachbiometric.Aqualitativeassessmentofthesecurityforsomebiometricscanbefoundinin.2.2.ExtendedLLRbeabinaryrandomvariablethatindicatesifauserisimpostor(1)orgenuine(0).Our“nalobjectiveistoevaluatetheLLRbetweenthegenuineandtheimpostordistributionasfollows: Sinceweareconsideringthefusionofdifferenttypesofbiometrics,wecanconsiderthatareindependent ARTICLEINPRESS Fig.1.Multimodalsystemarchitectureoverview. Pleasecitethisarticleas:R.N.Rodrigues,etal.,Robustnessofmultimodalbiometricfusionmethodsagainstspoofattacks,JournalofVisualLanguageandComputing(2009),doi:10.1016/j.jvlc.2009.01.010R.N.Rodriguesetal./JournalofVisualLanguagesandComputing]]]] ,thuswecanwriteUsually,theconditionaldistributionsinEq.(1)arelearntusingatrainingdatasetwhereisknowforasetofgivens.However,thisapproachdoesnotconsiderthefactthat,inpractice,animpostormayhavespoofedoneormorebiometricmodes,andthereforetheestimatedimpostordistributionmaynotberepresentative.Tosolvethisproblem,wecouldtrytocreateamultimodaldatasetwheresomesamples(forexample,5%ofthetotal)arespoofedsamplesandthenusestandardtechniquestolearntheimpostordistribution.However,thisapproachwouldrequireasigni“canteffortsinceitisnotalwayseasytospoofabiometricsample.Weproposeamodelthatcanestimatethetrueimpostordistributionwithouttheneedoftrainingspoofedsamples.Ourmodelreliesontheassumptionthatifabiometricsystemwassuccessfullyspoofedbyanimpostor,thesimilarityscore(inthiscase)willfollowagenuineprobabilitydistribution.Basedonthisassumption,we“rstintroducehiddenvariablesthatmodeltheprobabilityofaspoofattackandtheprobabilityofaspoofattacktobesuccessful;then,thesevariablesaremarginalizedtoobtainthetrueimpostordistribution.Themarginalizationoverthehiddenvariablescanbeinterpretedasifwewereconsideringeverypossiblesituationinwhichthemulti-modalsystemcanbespoofed.Westartbyde“ningasabinaryrandomvariablethatindicatesiftherewasanattempttospoofbiometricsystem1iftherewasanattemptofspoofingbiometricsystem0iftherewasnoattemptofspoofingbiometricsystemTheconditionaljointdistributionindicateshowoftenimpostorstrytospoofthebiometricsystems.Thisdistributiondependsontheapplicationsincesomeapplicationsmaybemorefrequentlytargetedbyforgersthanothers.Obviously,thereisnosenseinanauthenticusertrytospoofthesystem.Therefore,wecan1.Forimpostoruserswede“ne... isaparameterthatindicatestheprobabilitythatsomespoofattackhasbeenattempted.Notethatthereare1possiblecombinationsforspoo“ngthemultimodalbiometricsystems,beingthatweassignanequalprob-abilityforeachofthesecombinations.Wede“neanotherbinaryrandomvariableindicatesifagivenspoofattackwassuccessfulornot.Theprobabilitydistributionforthisvariableisconditionedonlybytheexistence(ornot)ofaspoofattempt.Obviously,iftherewasnospoofattackattempt(theprobabilityofasuccessfulspoofattack(1)iszero:0;thisimpliesthatTheprobabilityofaspoofattackbeingsuccessfulisdirectlyrelatedtohowsecureabiometricsystemis.Therefore,wede“netheprobabilityofaspoofattackbeingsuccessfulas.ThisimpliesthattheprobabilityofaspoofattackbeunsuccessfulisFig.2illustratetheproposedmodelforthecasewhen2.Wecanmarginalizetheintroducedhiddenvari-ablestoestimateasfollows:...............InordertoevaluateEq.(5)westillneedtoknowthe.Inthecasewhenthebiometricsystemwasnotspoofed(i.e.0),thisdistributioncanbelearntdirectlyfromthetrainingdata.Wenotethatthereisnoneedtoknowsincewehaveassumedthatagenuineuserwillnevertrytospoofthesystem.Usingtheunderlyingassumptionthatfollowsagenuinedistributionwhenabiometricsystemwassuccessfullyspoofed,weconsider.Now,givenanewtestinput,wecanuseEq.(5)toevaluatetheLLRinEq.(1).Weuseagammadistributiontomodelthegenuineandimpostordistributionwhen0.Thechoiceforthe ARTICLEINPRESS Fig.2.Graphicalmodelshowingtherelationbetweenthevariablesandparameters.Theshadedcirclesareobservedrandomvariables,whilethewhitecirclesarelatentrandomvariables.Theshadedsquaresrepresent“xedparameters.Ourobjectiveistoinferifanuserisimpostor(1)orauthentic(0)basedontheobservedvariables. Pleasecitethisarticleas:R.N.Rodrigues,etal.,Robustnessofmultimodalbiometricfusionmethodsagainstspoofattacks,JournalofVisualLanguageandComputing(2009),doi:10.1016/j.jvlc.2009.01.010R.N.Rodriguesetal./JournalofVisualLanguagesandComputing gammadistributionisduetotheempiricalevidencethatbiometricsimilarityscorestendtohavelongtailstails.Note,however,thatthisdistributiondependsonthematcherthatisbeingusedinthefusionandcanbeeasilybechangedforapplicationsthatusedifferentmatchers.representagammadistributionwherearetheshapeandinversescaleparameters,respectively.Thenwehavewheretheparametersarelearntviamaximumlikelihoodbasedontrainingdata.2.3.FuzzylogicfusionschemeTheprobabilisticfusionschemedescribedinSection2.2usesmanyheuristicsinthemodeling.Usually,onetriestousetrainingdatatolearnallprobabilitydistributions.However,inourcasethisisnotpossible,thusweareforcedtoheuristicallyde“nesomeprobabilitydistributions.Inthissectionwepresentafusionmethodbasedonfuzzylogicthatallowsustoexplicitlydescribetheheuristicsusinglinguisticexpressions.Forsakeofsimplicity,wedescribethefusionschemesforcombiningtwobiometricsystems(2)andindicatewhatshouldbemodi“edforthecaseofanarbitrarynumberofbiometricsystems.Thedevelopmentofafuzzylogicsysteminvolvesthreemainsteps:(i)de“ningfuzzyvariablesandtheirmember-shipfunctions(fuzzi“cationprocess);(ii)creatingthefuzzyrulesthatdescriberelationsbetweenthefuzzyvariables;(iii)establishinganappropriateddefuzzi“cationmethod[11].Thefuzzylogicsystemisbasedona“rstorderTakagi…Sugeno…Kangscheme[12]Fig.3showsablockdiagramforthefuzzyfusionscheme.Inthefuzzi“cationstep,eachoneofthesixinputs()ismodeledasafuzzyvariable.AmembershipfunctionmapseachfuzzyvariableintoarealnumberontherangethatdescribesthelinguisticexpressionhighparameterThesimilarityscorerangesmaybeverydistinctfordifferentbiometricmodesaswellasfordifferentmatch-ingalgorithmsthatusethesamebiometricmodality.Therefore,choosinganappropriatemembershipfunctioniscrucialforkeepingthelinguisticexpressionmeaningful.Forthehighsimilaritylinguisticexpressionweproposeamembershipfunctionwherethefuzzyregionistheregionwherethefalseacceptancerate(FAR)andfalserejectionrate(FRR)arenonzero:min1max0 arethepointsinthewheretheFRRandFARarezero,respectively.Fig.4illustratesthisfunctioninrelationtoanhypotheticalFARandFRRgraphs.ThevaluesforarelearntusingatrainingdatasetforeachbiometricsystemInfact,anymonotonicallyincreasingfunctionboundedbetweenzeroandonecouldbeusedasamembershipfunction.However,itshouldbecapableofinterpretingthehighsimilaritylinguisticexpressioninameaningfulway.Forthehighqualitylinguisticexpression,wechooseamin…maxfunction: qiminðQiÞmaxðQiÞminðQiÞ(9) ARTICLEINPRESS Fig.3.Fuzzyfusion. Pleasecitethisarticleas:R.N.Rodrigues,etal.,Robustnessofmultimodalbiometricfusionmethodsagainstspoofattacks,JournalofVisualLanguageandComputing(2009),doi:10.1016/j.jvlc.2009.01.010R.N.Rodriguesetal./JournalofVisualLanguagesandComputing]]]] whereminandmaxspecifytheminimumandmaximumvaluesfortheSincethesecurityparametersarealreadyconstrainedtotherangewede“neTheoutputofthefusionmodule,denotedby,isalsomodeledasafuzzyvariableandcanassumeoneofthethreelinguisticvalues:withthecorrespondingnumericvaluesas1;wherelargerthenumericalvalue,greateristheindicationthattheuserisgenuine.Oncealllinguisticvariablesareadequatelymappedintomembershipfunctions,theyareprocessedbyasetoffuzzyrules.Theserulesareelaboratedbasedonhumanexpertise.Duetospaceconstrains,wewillnotjustifyeachruleindividually.Themainideasfollowedinthecompilationofthesefuzzyrulesaresimple:lowsecuritybiometricsystemcannotfaithfullyperformtherecognitiontaskalone;andsimilarityscoreswithlowqualityshouldhavelowweightsinthe“naloutput.Thefuzzyrulesimplementedfortheproposedfuzzybimodalbiometricsystemarelistedbelow.ThesamerulesareillustratedinFig.5(1)Ifishighandishighthenishigh.(2)Ifisnothighandisnothighthenislow.(3)Ifishighandisnothighandisnothighandishighandisnothighthenismedium.(4)Ifishighandisnothighandishighandhighandisnothighthenislow.(5)Ifishighandishighandisnothighandhighandisnothighthenishigh.(6)Ifishighandishighandishighandishighisnothighthenismedium.(7)Ifisnothighandishighandisnothighthenislow.(8)Ifishighandisnothighandisnothighandisnothighandishighthenismedium. ARTICLEINPRESS 0 10 20 30 40 50 60 70 80 90 0 0.2 0.4 0.6 0.8 1 ThresholdError rate 0 10 20 30 40 50 60 70 80 90 0 0.2 0.4 0.6 0.8 1Similarity scoreHigh similarity FRR FAR ZeroFAR ZeroFRR Fig.4.Highsimilaritymembershipfunction. Fig.5.Fuzzyrules. Pleasecitethisarticleas:R.N.Rodrigues,etal.,Robustnessofmultimodalbiometricfusionmethodsagainstspoofattacks,JournalofVisualLanguageandComputing(2009),doi:10.1016/j.jvlc.2009.01.010R.N.Rodriguesetal./JournalofVisualLanguagesandComputing (9)Ifishighandishighandisnothighandnothighandishighthenislow.(10)Ifishighandisnothighandishighandnothighandishighthenishigh.(11)Ifishighandishighandishighandisnothighandishighthenismedium.(12)Ifisnothighandisnothighandishighthenislow.operationinthefuzzyrulesisimplementedthroughthemultiplicationbetweenthetwoinvolvedmembershipvalues.Theoperationisde“nedas,whereisthecorrespondingmembershipThedefuzzi“cationprocessisresponsibleforcombin-ingtheresultsproducedbyallfuzzyrulesandproducingonesinglescalaroutputsuitableforthe“nalclassi“ca-tion.Inourmethod,thedefuzzi“cationprocessiscarriedoutbyaweightedaverageasfollows: istheactivationvalueforrulecorrespondentoutputvalue.Forexample,ifweconsiderthesecondrule,wehave0.Afterthedefuzzi“cationstep,theoutputwillbeintherange,beingthatthebiggeritsvaluethebiggeristheindicationthattheuserisgenuine.Themainchallengeinextendingthisfusionschemeforanarbitrarynumberofbiometricsystems(2)isinthefuzzyrulesde“nition.Thisisbecausethenumberofrulesgrowexponentiallywith.In[13]wepresentapreliminaryanalysisfor3.ExperimentsTheperformanceoftheproposedfusionschemeswereevaluatedusingtwobiometricsystems:afacerecognitionsystemimplementedusingeigenfaces[14]andapubliclyavailable“ngerprintsystemdevelopedbyNIST[15].Thequalityofa“ngerprintsamplewascomputedusingtheNFIQsoftware[16],alsodevelopedbyNIST.Thequalityforafaceimagewasmanuallyassignedbasedonthefacerotation,illuminationandfacialexpression.Inapracticalapplication,thefaceimagequalitycouldbeautomaticallycalculatedusingmethodsdescribedin[17,18].Weusethesubscripttorefertothefaceand“ngerprintsystems,respectively(e.g.referstothesimilarityscorefor“ngerprint).Wesetthesecurityparametersas3and7.FortheextendedLLRweset01(i.e.theprobabilityofaspoofattackis1%).AmultimodaldatasetwascreatedbyrandomlycombiningusersfromtheFVC2004-DB1dataset[19]usersfromtheFERET-bseriesfacedatasetdataset,creatingamultimodaluser[21].Thebiometricsamplesforeachvirtualuserarerandomlyanduniquelycombinedtocreatemultimodalsamples.TheFVC2004-DB1datasetcontains1000“ngerprintsfrom100differentusers(10“ngerprintsperuser),whiletheFERET-bseriescontains2200facesfrom200users(11faceimagesperuser).Weruntheexperiments10times,beingthatanewmultimodaldatasetisrandomlycreatedateachtime.Theresultspresentedherearebasedontheaveragefromthese10runs.Werandomlychoose40usersfromthemultimodaldatasettotrainthefusionmodels,andusetheother60userstorunthetests.Threedifferentexperi-mentswereperformedusingthemultimodaltestdataset:I:Theobjectiveofthisexperimentistotestthefusionmethodsundernormaloperationconditions,wherenopoorqualitysamplesarepresented.Onlythosesampleswith6and6areusedintheII:Inthisexperiment,allsamplesareused,independentlyofitsquality.ComparingtheresultsfromthisexperimentwiththeresultsfromExperimentI,wecanevaluatewhichfusionmethodismorerobustinnoisyenvironments.III:Whathappenswhenoneoftheunimodalbiometricsinamultimodalsystemissuccess-fullyspoofed?Toanswerthisquestion,thisexperimentsimulatesascenariowherethefacesystemisspoofed.Inthisexperiment,themultimodalimpostorcomparisonswereperformedusinganimpostor“ngerprintsample(asusual)togetherwithagenuinefacesample.Weassumethatthereisnodifferencebetweenagenuinefacesampleandasuccessfullyspoofedone.AsinExperimentI,weonlyusesamplesthathavehiqhface6andhiqhfingBesidethefusionschemesproposedinSection2,wealsoruntheexperimentsonthefollowingfusionfusion:Eq.(1)describetheLLRfusion.Here,weusethetraditionalapproachtoestimatethegenuineandimpostordistributions(i.e.wedonotconsiderthepossibilityofaspoofattack).NotethatthisisequivalenttotheextendedLLRfusionschemedescribedinSection2.2wheretheprobabilityofaspoofattackiszero(i.e.Weightedsum:Somearticlesarticleshavereportedthatthesumrulefusionhavepresentedverygoodresults,evenwhencomparedwithsophisticatedmethodslikeneuralnetworkssanddecisiontreestrees.Theweightedsumfusionperformsalinearcombinationbetweenthesimilarityscoresasfollows:istheweightofbiometricsystem.Inourcase,sincewearecombiningonlytwobiometricsystems,wecanwrite .Inourexperiments,theparameterwasfoundthroughanextensivesearchforanoptimalvalueusingthetrainingdata.4.ResultsandanalysisTheexperimentresultsareanalyzedwiththeuseofthereceivingoperatingcurve(ROC)OC).Thiscurveis ARTICLEINPRESS Pleasecitethisarticleas:R.N.Rodrigues,etal.,Robustnessofmultimodalbiometricfusionmethodsagainstspoofattacks,JournalofVisualLanguageandComputing(2009),doi:10.1016/j.jvlc.2009.01.010R.N.Rodriguesetal./JournalofVisualLanguagesandComputing]]]] obtainedbyvaryingthedecisionthresholdvalueandplottingthegenuineacceptancerate(GAR)versustheFARforthesame(implicit)threshold.TheGARistheprobabilityofagenuineuserbeingcorrectlyacceptedasgenuineandtheFARistheprobabilitythatanimpostoruserbeingmistakenlyacceptedasgenuine.NotethatFRR,whereFRRisthefalserejectionrate.Fig.6showstheROCcurvesobtainedinExperimentI.Theresultsshowthatinthisscenario(nobadqualitysamples),allfusionmethodshaveimprovedcomparedtotheunimodalsystems.TheLLRandweightedsumgivethebestresults.Thisisexpectedandcanbejusti“ed.Insummary,thishappensbecausetraditionalmethodsdonotconsiderthehypothesisofspoofattack.FortheLLR,wecanuseNeyman…Pearsonlemmatoprovethatanyfusionmethodthatconsiderthepossibilityofspoofattackwillnotbebetter(onaverage)thantheLLRinExperiment1:LetHandHbetwohypothesesthatrepresentthepresenceorabsenceofspoofattack,respectively.Thenwecanwrite,whererepresentthesimilarityscoresandisalinearcombinationbetweenFARandFRRthatdependsonthechosenthreshold.InthetraditionalLLRfusionmethodonlyhypothesisHisconsideredduringtraining(i.e.isminimized),soitisguaran-teedthatwhenHistrueitwillhavethelowesterrorratepossible.Incontrast,ourmethodminimizesConsideringthatthefusionschemethatminimizesisdifferentthantheonethatminimizesandthat0,wehave(onaverage):,whererepresenttheerrorrateforanymethodthatconsiderthepossibilityofspoofattackandtheerrorforthetraditionalLLRfusionmethod.Fig.7showstheROCcurvesfortheExperimentII.Inthisscenario(wherebadqualitysampleswereintro-duced),wenotethatthefusionmethodsthatusesamplequalityscorearenotsoaffectedastheweightedsum.Theseresultssuggestthattheuseofsamplequalityindeedincreasetherobustnessoffusionmethodsasexpected.ThemoresurprisingresultswereobservedinExperimentIII,showninFig.8.Inthisscenario(wherethefacetraitwassuccessfullyspoofed),allfusionstrategieshadworseresultthan“ngerprintalone.However,theintroductionofsecurityparameterintheprobabilisticandfuzzyfusionmethodsresultedinamoresecuremulti-modalsystemwhencomparedwiththeLLRandweightedAnimportantparameterfortheproposedextendedLLR,whichrepresentsthepriorprobabilityofspoofattack.Thisparameterisapplicationdependentandmayevenvaryforthesameapplication(e.g.forsomeapplications,theprobabilityofaspoofattackmaybehigheratnight).Wecanunderstandtheeffectofbyconsidering,where.Forlowervaluesof,thewillhavealowerweightduringtheminimizationof.Therefore,theaveragetesterrorrateswillbehigherwhenHistrue(ExperimentIII)butlowerwhenHistrue(ExperimentI).Inpractice,abiometricsystemhastooperatewithauniquethresholdforallsituations.Forexample,wedonothaveaprioriknowledgeaboutthespoofattackonfacesystem,sowecannotsetaspeci“cthresholdforthissituation.However,whenplottingtheROCcurveforeachexperimentseparately,thethresholdremainsim-plicit,causingthelossofreferencewhencomparingthesamebiometricsystemindifferentexperiments.Therefore,weuseExperimentIROCcurveto“xsomereferencethresholdswhereFARis1%,0.1%and0.01%(referredas,respectively).Then,weevaluatetheFARandFRRfortheotherexperi-mentsusingthesethresholds.Ideally,theFARshouldbethesameasthereferencevaluesforallexperiments.Table1showstheerrorratesforthe“xedthresholds. ARTICLEINPRESS 104 103 102 101 100 101 102 0 10 20 30 40 50 60 70 80 90 100 False Acceptance Rate (FAR)%Genuine Acceptance Rate (GAR)% Fing Face ProbF Fuzzy LLR WSum Fig.6.ExperimentIROCcurves. Pleasecitethisarticleas:R.N.Rodrigues,etal.,Robustnessofmultimodalbiometricfusionmethodsagainstspoofattacks,JournalofVisualLanguageandComputing(2009),doi:10.1016/j.jvlc.2009.01.010R.N.Rodriguesetal./JournalofVisualLanguagesandComputing Analyzingthistable,thefollowingobservationscanbeInExperimentI,theweightedsumandLLRfusionmethodshavetheoverallbestresults.InExperimentII,theLLRfusionhadtheoverallbestresult.ThemethodsthatusethesamplequalityscorehadasmallerincreaseintheFRRwhencomparedwithExperimentI.Thissuggeststhattheintroductionofthesamplequalityscoreinthefusionschemecanindeedincreasethemultimodalrobustnessagainstnoisysamples.However,notethatsomemethodshadaslightincreaseintheFARaswell.InExperimentIII,both,weightedsumandLLRfusionmethodspresentedadramaticincreaseintheFARwhencomparedwiththereferenceFAR(ExperimentI).Forexample,whentheLLRfusionoperateswiththreshold,aforgerthathassuccessfullyspoofedthefacebiometricsystemhasachanceof42.09%ofbeingacceptedusinghisown“ngerprint;whileforthefuzzylogicfusion,hischancesare4.71%andfortheprobabilisticfusion,4.33%(boldednumbersinrow).Thissuggeststhatthetraditionalfusionmethodswhichdonotusethesecurityparametercanbecrackedbyspoo“ngonlyonebiometrictraitmoreeasilythantheproposedmethods.WeextendtheanalysisbetweenExperimentsIandIIIasfollows:foreachFARvalueinExperimentI(referredas)we“xthethresholdandevaluatetherespectiveFARinExperimentIII(referredasFAR).TheresultsareshownFig.9.ItcanbeseenthattheLLRfusion,whichhadthe ARTICLEINPRESS 102 101 100 101 102 0 10 20 30 40 50 60 70 80 90 100 False Acceptance Rate (FAR)%Genuine Acceptance Rate (GAR)% Fing Face ProbF Fuzzy LLR WSum Fig.7.ExperimentIIROCcurves. 102 101 100 101 102 0 10 20 30 40 50 60 70 80 90 100 False Acceptance Rate (FAR)%Genuine Acceptance Rate (GAR)% Fing Face ProbF Fuzzy LLR WSum Fig.8.ExperimentIIIROCcurves. Pleasecitethisarticleas:R.N.Rodrigues,etal.,Robustnessofmultimodalbiometricfusionmethodsagainstspoofattacks,JournalofVisualLanguageandComputing(2009),doi:10.1016/j.jvlc.2009.01.010R.N.Rodriguesetal./JournalofVisualLanguagesandComputing]]]] bestperformanceinExperimentsIandII,wasthemostaffectedbythespoo“ngofthefacebiometric.Theweightedsumwasalsosigni“cantlyaffected.Thefuzzylogicfusionpresentedthebestresults,whichforlowvaluesofFAR,weresimilartotheprobabilisticfusion.Thisresultindicatesthattheintroductionofthesecurityparametercanindeedresultinamoresecurefusionscheme.4.1.ValidationWevalidatetheresultsshownintheprevioussectionbyusingadifferentdataset:theNISTBiometricScoresRelease11,whichisapublicmultimodalscoredatasetthatincludesthesimilarityscoresforafacebiometricsystemanda“ngerprintbiometricsystem.Weusetheleftindex“ngerprintsetwiththeCfacerecognitionset.Sincetherearenosamplequalityscoresinthisdataset,we“xthequalityofallsamplesasone.Fig.10showthecomparisonbetweentheFARofExperimentsIandIIIforthisdatasetinthescenariowherethefacebiometricwassuccessfullyspoofed.Fig.11showthesamegraphbutinthescenariowherethe“ngerprintbiometricwasspoofed.TheresultssupporttheconclusionthattheproposedfusionschemesaremorerobustagainstspoofattackwhencomparedwiththeLLRandweightedsum. ARTICLEINPRESS Table1Errorratesforsome“xedreferencethresholds.ReferenceSystemThresholdsExperimentIExperimentIIExperimentIIIFAR(%)FRR(%)FAR(%)FRR(%)FAR(%)FRR(%)17.341.0117.7019.431.2718.83Face1.0028.0168.7372.3827.62ProbF1.001.5711.950.211.0014.021.001.1412.59WSum0.711.000.615.0124.070.1025.160.1227.6326.41Face0.1054.4954.17ProbF31.420.1018.0422.3118.830.1015.210.1621.464.7115.360.1011.630.1121.5411.69WSum0.1011.6038.5111.850.01Face0.0171.220.0028.5171.49ProbF185.940.010.0063.1231.700.0128.9110.570.0118.4829.1218.27WSum1.010.0118.870.0136.1312.2519.49 102 101 100 101 102 102 101 100 101 102 FAR Fing Face ProbF Fuzzy LLR WSum FAR1 % Fig.9.ComparisonbetweentheFARforExperimentI(FAR)andFARforExperimentIII(FAR)usingthesamethreshold. Pleasecitethisarticleas:R.N.Rodrigues,etal.,Robustnessofmultimodalbiometricfusionmethodsagainstspoofattacks,JournalofVisualLanguageandComputing(2009),doi:10.1016/j.jvlc.2009.01.010R.N.Rodriguesetal./JournalofVisualLanguagesandComputing 5.ConclusionInthiswork,wehaveanalyzedtheimpactofaspoofattackinmultimodalbiometricsystems.Ourexperimentsshowthatwhenusingtraditionalfusionschemes(i.e.LLRorweightedsum),aforgercandramaticallyincreasethechancesofcrackingamultimodalsystembyspoo“ngonlyoneofthebiometrics.Toreducethisweakness,weproposedtwonewfusionschemesthattakeintoaccountthesecurityofeachunimodalbiometricsystem.Theexperimentsindicatetheexistenceofatradeoffbetweenrecognitionaccuracyandrobustnessagainstspoofattacks.Theexperimentsalsoindicatethatthefuzzyfusionschemehadabetteroverallperformancewhencomparedwiththeprobabilisticfusionscheme.Inthefuture,wewillimplementatrainingprocesstoautomaticallyoptimizethemembershipfunctionsinthefuzzylogicfusion,andtestbothfusionschemeswithabroaderrangeofparameters.References[1]A.K.Jain,A.Ross,S.Prabhakar,Anintroductiontobiometricrecognition,IEEETransactionsonCircuitsandSystemsforVideoTechnology14(1)(2004).[2]C.Sanderson,K.K.Paliwal,Identityveri“cationusingspeechandfaceinformation,DigitalSignalProcessing14(2003)449…480.[3]J.Bigun,J.Fierrez-Aguilar,J.Ortega-Garcia,J.Gonzalez-Rodriguez,Multimodalbiometricauthenticationusingqualitysignalsin ARTICLEINPRESS 103 102 101 100 101 102 103 102 101 100 101 102 FAR Fing Face ProbF Fuzzy LLR WSum FAR Fig.10.ComparisonbetweentheFARforExperimentI(FAR)andFARforExperimentIII(FAR)whenthefacetraitwasspoofed. 103 102 101 100 101 102 104 103 102 101 100 101 102 FAR Fing Face ProbF Fuzzy LLR WSum FAR Fig.11.ComparisonbetweentheFARforExperimentI(FAR)andFARforExperimentIII(FAR)whenthe“ngerprintwasspoofed. Pleasecitethisarticleas:R.N.Rodrigues,etal.,Robustnessofmultimodalbiometricfusionmethodsagainstspoofattacks,JournalofVisualLanguageandComputing(2009),doi:10.1016/j.jvlc.2009.01.010R.N.Rodriguesetal./JournalofVisualLanguagesandComputing]]]] mobilecommunications,in:Proceedingsofthe12thInternationalConferenceonImageAnalysisandProcessing(ICIAP03),August2003.[4]J.Fierrez-Aguilar,J.Ortega-Garcia,J.Gonzalez-Rodriguez,J.Bigun,Discriminativemultimodalbiometricauthenticationbasedonqualitymeasures,PatternRecognition38(5)(2005)777…779.[5]K.-A.Toh,X.Jiang,W.-Y.Yau,Exploitingglobalandlocaldecisionsformultimodalbiometricsveri“cation,IEEETransactionsonSignalProcessing52(10)(2004)3059…3072.[6]A.Ross,A.K.Jain,Informationfusioninbiometrics,PatternRecognitionLetters24(2003).[7]Y.Wang,T.Tan,A.K.Jain,Combiningfaceandirisbiometricsforidentityveri“cation,in:LectureNotesinComputerScience,vol.2688,2003,pp.805…813.[8]K.Nandakumar,Y.Chen,S.C.Dass,A.K.Jain,Likelihoodratio-basedbiometricscorefusion,IEEETransactionsonPatternAnalysisandMachineIntelligence30(2)(2008)342…348.[9]S.Tulyakov,V.Govindaraju,Classi“ercombinationtypesforbiometricapplications,in:CVPRW06:Proceedingsofthe2006ConferenceonComputerVisionandPatternRecognitionWorkshop,2006,IEEEComputerSociety,Washington,DC,USA,pp.58,ISBN:0-7695-2646-2,doi:10.1109/CVPRW.2006.54[10]C.Soutar,Biometricsystemsecurity,2004http://www.bioscrypt.com/assets/security_soutar.pdf[11]W.Pedrycz,F.A.C.Gomide,AnIntroductiontoFuzzySets:AnalysisandDesign,ComplexAdaptiveSystems,MITPress,Cambridge,MA,1998.[12]M.Sugeno,K.T.Kang,Structureidenti“cationoffuzzymodel,FuzzySetsandSystems28(1991)191…212.[13]A.A.F.Ynui,L.L.Ling,Estudoeimplementacodeumsistematricomultimodalcomlogicanebulosa,UNICAMP,TechnicalReport,March2008.[14]M.A.Turk,A.P.Pentland,Facerecognitionusingeigenfaces,in:ProceedingsofComputerVisionandPatternRecognition,1991,pp.586…591.[15]Nist“ngerprintsoftware,NationalInstituteofStandardsandTechnology(NIST),Internetpage,March2006nist.gov/NFIS/[16]E.Tabassi,C.L.Wilson,C.I.Watson,Fingerprintimagequality,NationalInstituteofStandardsandTechnology(NIST),TechnicalReportNISTIR7151,August2004.[17]Y.Yao,B.Abidi,M.Abidi,QualityAssessmentandRestorationofFaceImagesinLongRange/HighZoomVideo,Springer,Berlin,2007,pp.43…60,(Chapter4).[18]Q.Xiong,C.Jaynes,Mugshotdatabaseacquisitioninvideosurveillancenetworksusingincrementalauto-clusteringqualitymeasures,in:AVSS03:ProceedingsoftheIEEEConferenceonAdvancedVideoandSignalBasedSurveillance,IEEEComputerSociety,Washington,DC,USA,2003,p.191.[19]Fvc2004:Thirdinternational“ngerprintveri“cationcompetition,UniversityofBologna,Internetpage,2006http://bias.csr.unibo.it/fvc2004/[20]NIST,Thefacialrecognitiontechnology(feret)database,NationalInstituteofStandardsandTechnology(NIST),TechnicalReportNISTIR7151,Augusthttp://www.itl.nist.gov/iad/humanid/feret/feret_master.html[21]R.Snelick,U.Uludag,A.Mink,M.Indovina,A.Jain,Large-scaleevaluationofmultimodalbiometricauthenticationusingstate-of-the-artsystems,IEEETransactionsonPatternAnalysisandMachineIntelligence27(3)(2005)450…455.[22]NIST,Biometricscoresset,NationalInstituteofStandardsandTechnology(NIST),TechnicalReportBSSR1,October2004www.itl.nist.gov/iad/894.03/biometricscores/ ARTICLEINPRESS Pleasecitethisarticleas:R.N.Rodrigues,etal.,Robustnessofmultimodalbiometricfusionmethodsagainstspoofattacks,JournalofVisualLanguageandComputing(2009),doi:10.1016/j.jvlc.2009.01.010R.N.Rodriguesetal./JournalofVisualLanguagesandComputing