文献翻译——图像信息隐藏算法的研究和实现

外文文献资料Thedevelopmentofteleradiologybringstheconvenienceofglobalmedicalrecordaccessalongwiththeconcernsoverthesecurityofmedicalimagestransmittedovertheopennetwork.Withtheprevailingadoptionofthree-dimensional(3-D)imagingmodalities,itisvitaltodevelopasecuritymechanismtoprovidelargevolumesofmedicalimageswithprivacyandreliability.ThispaperpresentsthedevelopmentofanewandimprovedmethodofimplementingtamperdetectionandlocalizationbasedonafullyreversibledigitalwatermarkingschemefortheprotectionofvolumetricDICOMimages.Thistamperdetectionandlocalizationmethodutilizesthe3-Dpropertyofvolumetricdatatoachievemuchfasterprocessingtimeatbothsenderandreceiversideswithoutcompromisingtamperlocalizationaccuracy.TheperformanceoftheproposedschemewasevaluatedbyusingsamplevolumetricDICOMimages.Resultsshowthattheschemeachievedonaverageabout65%and72%reductioninwatermarkinganddewatermarkingprocessingtime,respectively.Forcaseswheretheimageshadbeentampered,itispossibletodetectandlocalizethetamperedareaswithimprovedlocalizationresolutionintheimagesusingthescheme.Keywords:Imageauthentication,Medicaldatasecurity,Tamperdetection,WatermarkingIntroductionDigitalmedicalimagehasgainedimportanceovertheyearsandhasbecometheprimarymediumformedicalimaginginthehealthcareindustry.Inparallelwiththeadventofdigitalmedicalimages,rapiddevelopmentofinformationandcommunicationtechnologyhasfacilitatedtheadvancementofteleradiologybyenablingfasterandwideraccesstoelectronicmedicalrecords,includingmedicalimages.Withincreasingdemandbymedicalexpertsatdifferentlocationstostudymedicalimagesfrompeopleallovertheglobeinteleradiology,anincreasingnumberofmedicalimagesarebeingtransmittedthroughtheopennetwork(e.g.,theInternet).Ingeneral,therehasbeenanincreasingdemandforfasterandgreateraccessibilityofmedicalimagesacrossthehealthcareindustry.Whenmedicalimageswhichcontainsensitivityinformationofpatientsaretransmittedovertheopennetwork,theybecomevulnerabletocorruptionbynoisytransmissionchannelsandattacksbyhackersorindividualswithmaliciousintends.Theseattacksmayincludeobtainingprivateinformationaboutthepatient,changingpatientinformationintheimageheader,andtamperingtheimagepixelcontent.Theseattackswillcausetheimagestoloseintegrity.Afterlosingintegrity,theuseofthesetamperedmedicalimagescouldpotentiallyleadtowrongdiagnosisofdiseases,whichcouldresultinadifferenttreatmentforthepatient.Intheworst-casesituation,itmightevenresultindeathofpatients.Consequently,securityofdigitalmedicalimagesbecomesanimportantproblemthatneedstobeaddressed.TheDigitalImagingandCommunicationinMedicine(DICOM)(thestandardformedicalimages)definesasetofsecurityprofilesthatapplicationentitiesshouldconformtoduringtheexchangeofmedicaldata.Theseprofilesincludesecureuseprofiles,securetransportconnectionprofiles,digitalsignatureprofiles,andmediastoragesecurityprofiles1.TheDICOMimagefileconsistsofaheaderthatstorespatientsinformationandimagedata.However,theauthenticityofDICOMmedicalimagescanbecompromisedasitispossibleforahackertochangetheheaderofanotherpersonsDICOMimagefiletohisownto,forexample,claimmonetarygainfrominsurancecompany.DICOMstandardproposestoverifytheauthenticityofmedicalimagesbyusingdigitalsignatureandembeddingthesignatureintheheader.However,digitalsignaturereliesonthestrengthofcryptographichashfunction.IthasbeendemonstratedthatitispossibletogeneratetwodatasetswithdifferentcontentbuthavingthesameMessage-Digestalgorithm5hash2.Moreimportantly,digitalsignaturedoesnotprovidetamperlocalizationcapability.Furthermore,imagepixeldataaloneisnotprotectedusingtheDICOMmechanismsifthepixeldataareseparatedfromtheDICOMheader.Digitalwatermarkingisaprotectiontechniquethathasthepotentialtoprotecttheintegrityofmedicalimages,providingtamperprotectionevenwhentheimagesleavethenetwork.Theadvantagesofusingwatermarkingincludeenablingauthenticationinformationsuchasthemetadatatobeembeddedintothemedicalimagesasvisuallyunperceivablewatermarkpayloadandprovidingcontinuousauthenticityandintegrityprotectionoftheimagesbeyondthepointofinternalnetwork.Anumberofdigitalwatermarkingtechniquesformedicalimageshavebeenreported215.Thesewatermarkingtechniquescanbebroadlycategorizedintonon-reversible3,4andreversiblewatermarking2,519intheliteraturewherereversiblewatermarkinghasbeenmorewidelyimplemented.Thisismainlybecauseoftheclinicalneedwhichrequiresoriginalmedicalimagestobestudiedduringdiagnosis.Asdigitalmedicalimagescanbeeasilymodified3,itisimportanttoidentifywhetherimageshavebeenalteredandbeabletolocalizeregionsthathavebeentampered.Tothisend,watermarkingtechniqueswithtamperdetectionandlocalizationcapabilitythatallowtherecipienttodetectwhethertamperingofthemedicalimagehasoccurredandtolocatethetamperedregionsintheimagehavebeenreported.ZainandFauzidevelopedatamper-detectionmethodwhichwasachievedthroughblock-basedinspectionofimageblockusingparitychecksinamulti-levelhierarchicalmanner5.GuoandZhuangreportedanewregion-basedlosslesswatermarkingschemewithatamperlocalizationcapability7.Intheirmethod,aregionofauthenticationwaspartitionedintosmallnon-overlappingregionsinamulti-levelhierarchicalmanner,andthedigitalsignaturesoftheseregionswerethenusedtorealizetamperlocalization.Toachievetamperdetection,Changetal.calculatedtheaveragepixelvalueofeachnon-overlappingblockandencodedthevaluebyusingasymmetrickeycryptosystem.Thisinformationwasthendecryptedattherecipientendtocheckfortampering15.Inadditiontodetectingtamperedregions,Wuetal.incorporatedimageinformationintheirtamperdetectionmethodtoenabletamperedregionstobepartiallyrecovered8.Ourearlierworkisadual-layerreversiblewatermarkingtechniqueinwhichtamper-detectionandlocalizationwasdevelopedbasedonverifyingcyclicredundancycheck(CRC)inablockbyblockmanner2.Overall,theseworksachievedverygoodtamper-detectionaccuracy.Ingeneral,tamper-detectionintheseworkswasachievedbycheckingintegrityinformationofsmallnon-overlappingblocksinanimage.Locatingblockswhichfailedinthecheckingwillgivethelocationsofthetamperedregions.Thesemethodsmainlydifferintheerrorcheckingcodes(e.g.,CRC,digitalsignatures,etc)thatwereusedtoverifytheintegrityoftheblock.Theaforementionedstudiesmainlyfocusedonimagesacquiredbytwo-dimensional(2-D)imagingmodalitiessuchasX-rayradiographyandultrasoundordirectlyapplyingthe2-Dtechniquesonindividualimageofthreedimensional(3-D)dataacquiredbyimagingmodalitiessuchascomputedtomography(CT)andmagneticresonanceimaging(MRI).However,withtheadventof3-Dmedicalimagingmodalitiesandbecause3-Dimagingmodalitiesoftenproducealargenumberofcross-sectionalimagesforavolumeofinterest(e.g.,over100imagesforakneeMRIdatasetinasinglestudy),directapplicationof2-Dtechniqueson3-Dvolumetricdatawouldbecomputationallyinefficientandtime-consuming.Thisrequiresthedevelopmentofnewtamper-detectionandlocalizationschemesforvolumetricmedicalimagestoimprovetheefficiencyoftheoverallwatermarkingtechnique.Thispaperpresentsanimprovedtamper-detectionandlocalizationmethodthatexploitsthe3-DpropertyofvolumetricDICOMimages.Thisnewtechniquewasdevelopedtosignificantlyreducetheprocessingtimeandtoimprovetamperlocalizationwithoutcompromisingitsabilitytoensureauthenticity.Tothisend,twotypesofwatermarks(i.e.,horizontalandverticalwatermarks)thatutilizethe3-Dpropertyofvolumetricimageswereintroduced.Theproposedtamper-detectionandlocalizationmethodwasdevelopedanddemonstratedusingthedual-layerreversiblewatermarkingschemethatwasdevelopedinourearlierstudy2.TheperformanceoftheproposedschemewasevaluatedbyusingsampleDICOMvolumetricimagesandcomparingtoourearlierdevelopedtamperdetectionandlocalizationscheme2whichwasconsideredasarepresentativeofadirect2-Dapplicationonvolumetricimages.MethodologyOneofthecomputationalintensivepartsofwatermarkingwithtamperdetectionandlocalizationcapabilityistheprocessing(calculation,embedding,andchecking)ofblock-basederror-detectingfunctions(e.g.,CRC,digitalsignatures,etc).Inthispaper,thetamper-detectionandlocalizationmethodproposedusestwodifferentmethodsforleadingimage(firstimageinastack)andnon-leadingimages(therestoftheimagesinastack),respectively.EachzonedblockoftheleadingimageinthevolumestackwasembeddedwithbothhorizontalCRCandverticalCRCwhereasthatofothernon-leadingimageswasembeddedwithonlyhorizontalCRC.Theaimistoachieveasignificantimprovementincomputationalefficiencyandabetterresolutionforexaminingthetamperedregionstoenabletheproposedschemetobeamorepracticalsolutiontoprotect3-Dvolumetricimages.Thereversiblewatermarkingalgorithmthatwasusedtoimplementthisnewtamperdetectionandlocalizationmethodisbasedonourearlierwork2.Althoughthemainfocusofthispaperisondevelopingaschemefortamperdetectionandlocalization,thereversiblewatermarkingmethodispresentedinthenextsectionforcompleteness.ReversibleWatermarkingAlgorithmInthissection,thereversiblewatermarkingalgorithmproposedinourearlierstudy2ispresented.Itconsistsoftwostages(i.e.,dataembeddinganddataextraction).Duringthedataembeddingstageatthesenderside,animageofWHpixelstobewatermarkedisfirstlydividedintonon-overlappingsquareblocksofsizeNN,where1Nmin(W,H).InthecasethatWorHisnotamultipleofN,theremainingpartattheendofeachblockroworblockcolumnwillbeappendedtotheblocktoitsleftorabove,respectively.Awatermarkingparameter,Q,isthendefinedtoensurethatenoughembeddingcapacityisavailable.EachNNimageblockissubsequentlyprocessedaccordingtotheflowchartillustratedinFig.1inarasterscanmanneruntiltheentirebitstreamisembedded.Fig.1ThewatermarkingflowchartDuringdataextractionstage(de-watermarking)atthereceiverdetectorside,thewatermarkedimageisfirstlydividedintonon-overlappingblocksinthesamewayasatthesenderside.EachblockissubsequentlyprocessedaccordingtotheflowchartshowninFig.2inarasterscanmanneruntiltheentireembeddedbitstreamisextracted.Inthisway,thewatermarkedimageisconvertedbacktotheoriginalimage.Fig.2ThedewatermarkingflowchartInourwatermarkingscheme,tamper-detectionandlocalizationinformationisembeddedasthesecondwatermarklayer(whichisthefocusofthispaper)aftermetadataisembeddedasthefirstwatermarklayer.IntegrityEmbeddingforTamper-DetectionandLocalizationIntegrityembeddingfortamper-detectionandlocalizationisachievedbyembeddingerror-detectiondataoneveryzonedimageregioninareversiblemanner.Toavoidtheneedtocheckeverysliceinavolumestackduringdewatermarkingandreducethesizeofthelocalizationregion(improvinglocalizationresolution),anewschemefortamper-detectionandlocalizationwasdeveloped.Theschemeconsidersthefirstimagesliceinavolumestackastheleadingimageandthesubsequentslicesthatfollowasthenon-leadingimages.Thetechniqueprocessestheleadingimageinadifferentmannerfromnon-leadingimages:Eachnon-leadingimageisembeddedonlywithhorizontalCRCwhereastheleadingimageisembeddedwithbothhorizontalCRCandverticalCRC.ThepurposeoftheverticalCRCistoenablerapidcheckingoftheoverallintegrityofthevolumestack(withouttheneedtocheckeverysingleslice)whilethehorizontalCRCistoallowcheckingoftheintegrityofisolatedregionswithineachimagewhenthecorrespondingverticalCRCcheckingfails.TheembeddingstructureisillustratedinFig.3.Thedetailsoftheembeddingprocessaredescribedinthefollowingsections.Fig.3TheintegrityembeddingstructureIntegrityEmbeddingforNon-leadingImageNon-leadingimageisembeddedonlywithhorizontalCRC.Atthesenderside,theimageisfirstlydividedintonon-overlappinglargeblocks(Inthispapertheterms“largeblock”and“smallblock”refertoablockofsize18by18andablockofsize9by9,respectively).AwatermarkingparameterQ2isthendefinedtoensuresufficientembeddingcapacityforthesecondwatermarklayer.OnenecessaryconditiontomaintainacomparableembeddingcapacitytothefirstwatermarklayeristhatQ22Q1,whereQ1isthewatermarkingparameterofthefirstwatermarklayerandalsothemaximumdifferencebetweenreferencepixelandalterablepixelthatcanbeextendedbytheembeddingalgorithminthefirstwatermarklayer.ThedefinitionsofbothQ1andQ2arealsorestrictedbytheunderflowandoverflowconditions,whichwilloccurrespectivelyinthecaseofDICOMimageswith16-bitdepthwhenthegraylevelofnisdecreasedbymorethannorthegraylevelof65535-nisincreasedbymorethann.Asaresult,true16-bitimageswouldnotbeappropriatefortheproposedalgorithm.However,currentmedicalimagingmodalitiesgenerallydonotproduceimagesthatutilizethefullrangeofgraylevels.Therefore,allthepixelvaluesoftheoriginalimageareincreasedbyQ,whereQ=Q1+Q2,toavoidoverflowandunderflowconditions,andallthepixelvalueswouldeventuallybedecreasedbyQtorestoretheoriginalimageatthereceiverdetectorside.AfterdefiningQ2,eachblockof18by18pixelsissubsequentlyprocessedinarasterscanmanneraccordingtothefollowingsteps.1.DeterminetheembeddingcapacityCofthecurrentblock.Here,theembeddingalgorithmofreversiblewatermarkingpresentedearlierwasmodifiedtodetermineCbyconsideringtheblockasinputimage,segmentingitintosquareblockswithN=3andselectingcentralpixelofthe3by3blockasreferencepixel.Pixelsarethenupdatedaccordingtothefollowingstep:If|R-A|Q,C=C+1.2.Decidetheappropriateerrordetectioncodetobeusedbasedontheavailableembeddingcapacitywhichcouldbeslightlylowerthan16insomeregionsoftheimage.1.IfC16,calculatetheCRC-16-CCITTcodeoftheblock.2,If12C16,calculatetheCRC-12codeoftheblock.HorizontalCRCcodecalculationisperformedinthisstep.Toreducethecomputationtimeandtomaintaintamperdetectioncapability,thealgorithmfirstgeneratesanewblockof9by9pixelsfromtheoriginal18by18pixelsblock.The9by9pixelsblockisobtainedbysummingthepixelvaluesofeverynon-overlapping2by2pixelsblockwithinthe18by18pixelsblock.Hence,eachpixelinthe9by9pixelsblockcorrespondstothesummationofa2by2pixelsblockinthe18by18pixelsblock.TheCRCcodeofthe9by9pixelsblockisthencalculated.3.Performthedataembeddingstageofthereversiblewatermarkingalgorithmpresentedinearliersectiononthecurrentblockbyconsideringtheblockasaninputimage(i.e.,W=H=18).SelectthesegmentedsquareblocksizeNtobe3.SelectthecentralpixelasreferencepixelforeveryblockanduseQ2asthewatermarkingparameterQ.4.Theembeddingprocesscontinuesuntil16bitshavebeenembeddedinthecaseofCRC-16-CCITToruntiltheCRCbitsarerepeatedlyembeddedtofullyoccupytheembeddingcapacityinthecaseofCRC-12.Oncestep4iscompleted,block-basederror-detectiondataaresuccessfullyembedded.IntegrityEmbeddingforLeadingImageWiththeaimtofurtherreducetheprocessingtimeatthereceiverdetectorside,thenewconceptofverticalwatermarkingwasintroduced.Throughtheuseoftheverticalwatermarks,thereceiverdetectorcanquicklydeterminewhethertamperinghasoccurredinaparticularnarrowvolume.Inthecasewhenthereisnotampering,nofurtherverificationisrequiredforeachsingleimageinthevolumestack;whileinthecasewhentamperinghasoccurred,verticalwatermarkingwouldhighlightthattamperinghasoccurredinthatnarrowvolumeandfurtherexaminationintothehorizontalCRCcodesthatcorrespondtotheverticalwatermarksforeachsingleimagecanlocatethetampermoreprecisely.Theleadingimageofthevolumestackisselectedtoaccommodatetheverticalwatermarks,inadditiontoitsownhorizontalCRCcodes.Usingsimilarembeddingtechniqueasthenon-leadingimage,theleadingimageisalsofirstlydividedintonon-overlappingblocksof18by18pixels.AthirdwatermarkingparameterQ3isthendefinedtobe8Q1toensuresufficientembeddingcapacityforthesecondwatermarklayer.Similarly,allthepixelvaluesofleadingimageareincreasedbyQ,whereQ=Q1+Q3,toavoidoverflowandunderflowconditions.AllthepixelvalueswilleventuallybedecreasedbyQtorestoretheoriginalimageatthereceiverdetectorside.AfterdefiningQ3,eachblockof18by18pixels(referredaslargeblock)issubsequentlyprocessedinarasterscanmanneraccordingtothefollowingsteps.1.CalculatetheCRC-16-CCITTcodeofthelargeblockusingthenewtechniqueasdescribedinthestep2forprocessingthenon-leadingimage.2.Dividethelargeblockintofournon-overlappingblocksof9by9pixels(referredassmallblock)andcalculatetheverticalCRC-8codeforeachofthefournarrowvolumeswherethefoursmallblocksreside,respectively.Eachnarrowvolumecomprisesa9by9pixelsblockfromtheleadingimageandaseriesof9by9pixelsblocksfromthesubsequentnon-leadingimagesatthesamehorizontalcoordinates(seeFig.3).TheverticalCRCcodeiscalculatedusingthevolumeblockofwhicheachpixelvalueisthesummationofallthepixelsatthesamehorizontalcoordinatesacrossthevolume.3.Obtainthedatabitstobeembeddedforeachsmallblock.Theentiredatabitsconsistoffourdatabitsofthe16-bithorizontalCRCcodeofthelargeblockaswellastheentireeight-bitverticalCRCcodeofthevolumewherethesmallblockresides.4.Performthedataembeddingstageofthereversiblewatermarkingalgorithmpresentedinearliersectiononeachsmallblockbyconsideringitasinputimage(i.e.,W=H=9).SelectthesegmentedsquareblocksizeNtobe3.SelectthecentralpixelasreferencepixelforeveryblockanduseQ3asthewatermarkingparameterQ.5.embeddingprocesscontinuesuntilall12bitshavebeenembeddedforeachsmallblock.IntegrityExtractionAtthereceiverdetectorside,thewatermarkedleadingimageisfirstlydividedintonon-overlappingblocksof18by18pixels.Eachblockissubsequentlyprocessedinarasterscanmanneraccordingtothefollowingsteps.1.Dividethelargeblockintofoursmallblocksinthesamewayasatthesenderside.2.Performthedataextractionstageofthereversiblewatermarkingalgorithmpresentedinearliersectiononeachsmallblockbyconsideringitasawatermarkedimage.Segmentitintoblocksofthesamesize(N=3)asatthesenderside.3.Theextractionprocesscontinuesuntil12bitshavebeenextractedforeachsmallblock.Basedontheextracteddatabits,obtainfourextractedeight-bitverticalCRCcodesforthefournarrowvolumeswherethesmallblocksbelongtoandoneextracted16-bithorizontalCRCcodefortheleadingimage.4.Foreachsmallblock,ifthenumberofbitsextractedislessthan12ortheextractedeight-bitverticalCRCcodedoesnotagreewiththenewlycalculatedverticalCRCcode,thenarrowvolumewherethatblockresidesisidentifiedasbeingtampered.Furtherexaminationonthelargeblockofeachnon-leadingimageinthesamevolumeasthelargeblockoftheleadingimageisrequiredtodeterminetheexactlocationofthetamperandiscarriedoutasfollows:Performthedataextractionstageofthereversiblewatermarkingalgorithmpresentedinearliersectiononthecurrentblockbyconsideringitasawatermarkedimage.Segmentitintoblocksofthesamesize(N=3)asatthesenderside.Theextractionprocesscontinuesuntil16bits,whicharedeemedtobetheextractedhorizontalCRCcode,oruntilalltheavailablebitshavebeenextracted.Ifthenumberofbitsextractedislargerthan11butsmallerthan16,thefirst12bitsaredeemedastheextractedhorizontalCRCcode.Ifthen