Reliable information hiding based on support vector machine.docx

Reliable information hiding based on support vector machine.docx

《Reliable information hiding based on support vector machine.docx》由会员分享，可在线阅读，更多相关《Reliable information hiding based on support vector machine.docx（20页珍藏版）》请在冰豆网上搜索。

Reliableinformationhidingbasedonsupportvectormachine

ReliableInformationHidingBasedonSupportVectorMachine

Yong-GangFu1,3,Rui-MinShen1,Li-PingShen1andXu-ShengLei2

1Dept.ofComputerScienceandEngineering,ShanghaiJiaotongUniv.,Shanghai,China,200030

2Dept.ofAutomation,ShanghaiJiaotongUniv.,Shanghai,China,200030

3Dept.ofSoftware,XiamenUniv.,Xiamen,Chian,361000

E-mails:

{fyg,rmshen,lpshen,xushenglei}@

Abstract:

Inthispaper,areliableinformationhidingschemebasedonsupportvectormachineanderrorcorrectingcodesisproposed.Toextractthehiddeninformationbitsfromapossiblytamperedwatermarkedimagewithalowererrorprobability,informationhidingismodeledasadigitalcommunicationproblem,andboththegoodgeneralizationabilityofsupportvectormachineandtheerrorcorrectioncodeBCHareapplied.Duetothegoodlearningabilityofsupportvectormachine,itcanlearntherelationshipbetweenthehiddeninformationandcorrespondingwatermarkedimage;whenthewatermarkedimageisattackedbysomeintentionalorunintentionalattacks,thetrainedsupportvectormachinecanrecovertherighthiddeninformationbits.Thereliabilityoftheproposedschemehasbeentestedunderdifferentattacks.Theexperimentalresultsshowthattheembeddedinformationbitsareperceptuallytransparentandcansuccessfullyresistcommonimageprocessing,jitterattack,andgeometricaldistortions.Whenthehostimageisheavilydistorted,thehiddeninformationcanalsobeextractedrecognizably,whilemostofexistingmethodsaredefeated.Weexpectthisapproachprovideanalternativewayforreliableinformationhidingbyapplyingmachinelearningtechnologies.

Keywords:

Informationhiding;Supportvectormachine;Digitalwatermarking;BCHcoding

1.Introduction

DigitalpropertiesarereadilyreproducedandredistributedovertheInternetandothermedias.Howevertheseattractivepropertiesleadtoproblemsenforcingcopyrightprotectionissues.Asaresult,thecontributoranddistributorofthedigitalpropertiesarehesitanttoprovidetheaccesstotheirintellectualproperties.Itisrealizedthatconventionalcryptographicmeansarenotsufficientsincethedataiswithoutanyprotectionassoonasitisused,e.g.,decryptedanddisplayedinthecaseofimageorvideodata.Apotentialapproachtosolvethisproblemisinformationhidingordigitalwatermarking（Swansonetal.,1998）.Informationhidingistheimperceptibleembeddingofinformationbits（signature）intomultimediadata,wheretheinformationremainsdetectableaslongthequalityofthecontentitselfisnotrendereduseless.Asabranchofinformationhiding,itiscommonlyassumedthatdigitalwatermarkingisonlyoneofseveralmeasuresthathavetobecombinedtobuildagoodcopyprotectionmechanism（FuronandDuhamel,2000）.Asignificantmeritofdigitalwatermarkingovertraditionalprotectionmethodsistoprovideaseamlessinterface,sothatusersarestillabletoutilizetheprotectedmultimediatransparently.

Aninformationhidingschemeshouldatleastmeetthefollowingrequirements:

（1）Perceptualinvisible（ortransparent）.

（2）Difficulttoremovewithoutseriouslyaffectingtheimagequality.（3）Robustresistancetoimageprocessing,andattacks.

Developinganalgorithmcapableofproducingsignaturethatfulfillsalltheserequirementsisnotaneasytask.Ononehand,theinformationhidingprocessshouldnotintroduceanyperceptibleartifactsintothehostimage.Ontheotherhand,forhighrobustnessitisdesirablethatthemarkamplitudeisashighaspossible.Therefore,thedesignationofinformationhidingmethodalwaysinvolvesatradeoffbetweenimperceptibility（ortransparency）androbustness.Avarietyofwatermarkingorinformationhidingschemeshavebeenreportedrecentlyintheliterature,andsomenicereviewscanbefoundin（Fabienetal.,1999）.However,theresearchoncopyrightprotectionofimagesisstillinitsearlystageandnoneoftheexistingmethodsistotallyeffectiveagainstmaliciousattacks.

Thereareavarietyofschemesforhidinginformationintotheoriginalimage.Typicalschemesfortheinformationhidinginimagescanbebroadlyclassifiedintotwocategories:

（i）spatialdomainmethodswhichembedthedatabydirectlymodifyingthepixelvaluesoftheoriginalimage（NikolaidisandPitas,1998）;（ii）transformdomainmethodswhichembedthedatabymodulatingthecoefficientsofproperlychosentransformdomainsuchasDCT（Coxetal.,1997;Barnietal.,2000）,DFT（Barnietal.,2000）,andDWT（Xiaetal.,1998）.Manyofthespatialdomaintechniquesprovidesimpleandeffectiveschemesforembeddinganinvisiblewatermarkintoanimagebutarenotrobusttocommonattacks.Informationhidingtechniquescanbealternativelysplitintotwodistinctcategoriesdependingonwhethertheoriginalimageisnecessaryforthewatermarkextractionornot.Althoughtheexistenceoftheoriginalimagefacilitateswatermarkextraction（Coxetal.,1997;Swansonetal.1996;PodilchukandZeng,1998）togreatextent,sucharequirementraisestwoproblems:

（i）owneroftheoriginalimageiscompelledinsecurelytosharehisworkswithanyonewhowantstochecktheexistenceofthesignature（Barnietal.,1998）,（ii）ontheotherhand,thesearchingwithinthedatabasefortheoriginalimagethatcorrespondstoagivenwatermarkedimagewouldbeverytimeconsuming.Thus,methodscapableofrevealingtheinformationbitspresencewithoutcomparingthewatermarkedandoriginalimageswouldbepreferable.

Inordertodesignrobustinformationhidingscheme,Coxetal.consideredwatermarkingasaproblemofcommunicationwithsideinformation（Coxetal.,1999）.Also,somewatermarkingalgorithminliteratureappliederrorcorrectingcoding（ECC）toimprovethebiterrorrate（BER）performance,suchasBose-Chaudhuri-Hocquenghen（BCH）coding（Huangetal.1998;HuangandYun,2002）,Reed-Solomon（R-S）code（WuandHsieh,2000）,andTurbocode（PereiraandPun,2000）.Recently,effortsaremadetouseartificialintelligencetechniqueforwatermarkembeddingandextraction.Neuralnetworksareintroducedintowatermarkingin（Yuetal.,2001）,whichmakesthewatermarkextractionmorerobustagainstcommonattacks.GeneticalgorithmisproposedforselectionofthebestembeddingpositionsinblockbasedDCTdomainwatermarking（Shiehetal.,2004）.Wehavefirstlyintroducedthesupportvectormachineforthewatermarkembeddingandextractionin（Fuetal.,2004）,inwhichthewatermarkisembeddedintothehostbyapplyingthegoodlearningabilityofsupportvectorregressionmachine,andthewatermarkextractionisfinishedbytheaidsofthewelltrainedsupportvectormachine.Wecanexpectthatthecombinationofinformationhidingandmachinelearningtechniquesmightbeagoodsolutionforreliableinformationhiding.

Fromtheobservationsabove,inthispaperweproposeanovelblindreliableinformationhidingandrecoveringschemewhichmakesuseofsupportvectormachineandBCHcoding.Thisworkcanbeconsideredasanextensionofsomeexistingresearch（Kutteretal.,1998;Yuetal.,2001;Fuetal.,2004）.In（Kutteretal.,1998）,Kutterproposedaspatialdomainwatermarkingschemeforcolorimage.ThenYu（Yuetal.,2001）improvedKutter’smethodbyapplyingneuralnetworks.Duetothesupportvectormachine’sgoodlearningabilityintrainingprocess,itcanmemorizetherelationshipbetweentheembeddedinformationandcorrespondingwatermarkedimage.ApplyingSVM’sgoodgeneralizationabilitiesanderrorcorrectingabilityofBCHcoding,hiddeninformationextractioncanbefinishedwell.Experimentalresultsshowgoodrobustnessoftheproposedschemeagainstcommonimageprocessingandattacks.Thisresearchismuchdifferentfrommyearlywork.Inthisresearch,thesupportvectormachineisonlytrainedandappliedintheinformationextractionprocedure,whereas,in（Fuetal.,2004）,thesupportvectorregressionmachineisappliedbothinthewatermarkembeddingandextractionprocess.

Thepaperisorganizedasfollows:

insection2,basicconceptionsforsupportvectormachineareintroduced.Theembeddingandextractionalgorithmsofourmethodaredescribedinsection3.

Insection4,someexperimentalresultsareexhibited.Theconclusionisstatedinsection5.

2.AnoverviewofSupportVectorMachine

SupportVectorMachine（SVM）isauniversalclassificationalgorithmdevelopedbyVapnikandhiscolleagues（Vapnik,1995;Vapnik,1998）.Inrecentyears,therehavebeenalotofinterestsinstudyingtheapplicationsofSVMonfunctionapproximation,patternrecognitionproblemsandsoon（Campbell,2002;Christopher,1998）.

Givenatrainingdatasetofmsamples

where

istheithinputpatternand

istheithoutputpattern.

Thesupportvectormachinemethodsupposeswehavesomehyper-planesthatseparatethepositivesamplesfromnegativeones.Thepoint

whichliesonthehyper-planesatisfies

where

isnormaltothehyper-plane,

istheperpendiculardistancefromthehyper-planetotheorigin,and

istheEuclideannormof

.Forthelinearlyseparablecase,thesupportvectoralgorithmsimplylooksfortheseparatinghyper-planewiththelargestmargin.Thiscanbeformulatedasfollowing:

supposethatallthetrainingdatasatisfythefollowingconstraints:

（1）

Thiscanbecombinedintoonesetofinequalities:

（2）

Nowconsiderthepointsforwhichtheequalityin

（1）holds.Thesepointslieonthehyper-plane

withnormal

andperpendiculardistancefromtheorigin

.Similarly,thepointsforwhichtheequalityholdsin

（1）lieonthehyper-plane

withnormal

andperpendiculardistancefromtheorigin

.Hencethe