基于卷积神经网络模型的视频帧间伪造检测(IJIGSP-V12-N3-1)

I.J.Image,GraphicsandSignalProcessing,2020,3,1-12PublishedOnlineJune2020inMECS()DOI:10.5815/ijigsp.2020.03.01Copyright©2020MECSI.J.Image,GraphicsandSignalProcessing,2020,3,1-12DetectingVideoInter-FrameForgeriesBasedonConvolutionalNeuralNetworkModelXuanHauNguyen*1,2,YongjianHu,MuhmmadAhmadAminandKhanGoharHayat1ResearchCentreofMultimediaInformationSecurityDetectionandIntelligentProcessing,SchoolofElectronicsandInformationEngineering,SouthChinaUniversityofTechnology,Guangzhou510640,P.R.China.Email:[e-mail],[e-mail],[e-mail],[e-mail]2FacultyElectronicsofandInformaticsEngineeringMienTrungIndustrialandTradeCollege,PhuYen620000,VietnamEmail:[e-mail]-TuTruongNaturalLanguageProcessingandKnowledgeDiscoveryLaboratory,FacultyofInformationTechnology,TonDucThangUniversity,HoChiMinhCity,700000,VietnamEmail:[e-mail]:03December2019;Accepted:05February2020;Published:08June2020Abstract—Intheeraofinformationextensiontoday,videosareeasilycapturedandmadeviralinashorttime,andvideotamperinghasbecomemorecomfortableduetoeditingsoftware.So,theauthenticityofvideosbecomesmoreessential.Videointer-frameforgeriesarethemostcommontypeofvideoforgerymethods,whicharedifficulttodetectbythenakedeye.Untilnow,somealgorithmshavebeensuggestedfordetectinginter-frameforgeriesbasedonhandicraftfeatures,buttheaccuracyandprocessingspeedofthosealgorithmsarestillchallenging.Inthispaper,wearegoingtoputforwardavideoforgerydetectionmethodfordetectingvideointer-frameforgeriesbasedonconvolutionalneuralnetwork(CNN)modelsbyretrainingtheavailableCNNmodeltrainedonImageNetdataset.Theproposedmethodbasedonstate-the-artCNNmodels,whichareretrainedtoexploitspatial-temporalrelationshipsinavideotodetectinter-frameforgeriesrobustlyandwehavealsoproposedaconfidencescoreinsteadoftherawoutputscorebasedonthesenetworksforincreasingaccuracyoftheproposedmethod.Throughtheexperiments,thedetectionaccuracyoftheproposedmethodis99.17%.Thisresulthasshownthattheproposedmethodhassignificantlyhigherefficiencyandaccuracythanotherrecentmethods.IndexTerms—Videoforensic,videoforgerydetection,videointer-frameforgerydetection,convolutionalneuralnetwork,videoauthenticity,passiveforensic.I.INTRODUCTIONNowadays,smartphone,camcorder,andsecuritycamerasareusedextensivelyinmanyareasofdailylife.Especiallyintrafficlights,offices,houses,dormitoriesandmanyotherplaceswhicharemonitoredbycameras.Besidesthat,videoeditingsoftwarelikeVideoEditor,AdobePhotoshop,WindowMovieMaker,andAdobeAfterEffectareavailableandeasilyutilized.Thesetoolsprovidegreatsupportforeditingvideocontenteasily,andanyonecaneditvideocontentbytheirwill,eveneditedcontentcontrastwithoriginalcontent,whichleadstoseeingisnolongerbelieving.Inaddition,anauthenticvideogivesevidencestrongerthananauthenticimageincourt.Therefore,videoforensicprovesthatvideoauthenticitybecomesanurgentrequirementtoday.So,nowadays,videoforensichasbecomeahottopicofinterestamongstresearchersintheworld.Thevideoforensicmethodsaredividedintoactiveandpassivemethods.ActivemethodsusegiveninformationsuchasWatermarkingorSignaturewhichisinsertedintovideos,thenthatinformationischecked.Ifitdoesnotchange,thatvideoisauthenticotherwiseforged.Meanwhile,passivemethodsonlyanalyzevideocontenttofindtracesofforgeries.Now,mostofthevideosdonotusuallyinsertgiveninformation,sothepassivemethodshavebecomeanexcitingtopicthathasattractedmanyresearchers.2DetectingVideoInter-FrameForgeriesBasedonConvolutionalNeuralNetworkModelCopyright©2020MECSI.J.Image,GraphicsandSignalProcessing,2020,3,1-12Inreality,ManipulationsatframelevelsuchasFrameInsertion,FrameDeletion,FrameDuplication,andFrameShufflingeasilyconcealorimitatecontentinthevideo,thesemanipulationsaresimpleskillsineditingcontentofthevideo,buttheywouldcreateforgedvideoshardtodetectespeciallybynakedeyes.Inaddition,manipulationsoftampervideosattheframelevel,whichwerestronglysupportedbyvideocontenteditingapplicationssuchasAfterEffect,MovieMakerorPhotoshopvisually.Anyonecanperformdeletion,duplicationorinsertionofa-framessequenceefficientlybyonlyoneortwoactionsontheseapplications.AsshowninFig.1aframesequence(picturesfrom1’to4’)wascopiedandpastedtocreateaforgedvideo.Thisactionisintendedtofakethepresenceofthemaninthevideo.Andsimilarly,thesameoperationisshowninFig.2tohideababyinavideo.Throughthestate-of-the-art,thereweremanymethodssuggestedfordetectingvideointer-frameforgeries,mostofthembasedonhandicraftfeaturesanalysisofframesinsidevideo[1-7].ThosefeaturesareColorhistogram,Opticalflow,Motionenergy,texture,noise,singularvaluedecomposition(SVD),orcorrelationcoefficientsofgreyvalues.Becauseanalysisofthesehandicraftfeaturesisonalargenumberofframesinavideo.Allofthemhaveconsumedalotoftimeandtheaccuracyoftheabovemethodsisstilllow.So,Videointer-frameforgeriesdetectionisstillasignificantchallenge.Throughrecentresearches[8-15],deeplearninghasoutstandingresults.Particularly,theCNNshaveachievedexceptionalresultsinsolvingmanychallengingvisionproblemssuchaso