MUMBAISILICONVALLEYBENGALURUSINGAPORENEWDELHINEWYORKGIFTCITY

Research

UnmaskingDeepfakes

Legal,RegulatoryandEthicalConsiderations

October2024

?NishithDesaiAssociates2024

Research

UnmaskingDeepfakes

Legal,RegulatoryandEthicalConsiderations

October2024

DMSCode:100042.1

?NishithDesaiAssociates2024

Rankedasthe‘MostInnovativeIndianLawFirm’intheprestigiousFTInnovativeLawyersAsiaPacificAwardsformultipleyears.Alsorankedamongstthe‘MostInnovativeAsiaPacificLawFirm’intheseeliteFinancialTimesInnovationrankings.

llwwL

Aegermartet

回woruDX

BUSINESSTODAY

equalipi

?NishithDesaiAssociates2024

UnmaskingDeepfakes—Legal,RegulatoryandEthicalConsiderations

Disclaimer

ThisreportisacopyrightofNishithDesaiAssociates.Noreadershouldactonthebasisofanystatementcontainedhereinwithoutseekingprofessionaladvice.Theauthorsandthefirmexpresslydisclaimallandanyliabilitytoanypersonwhohasreadthisreport,orotherwise,inrespectofanything,andofconsequencesofanythingdone,oromittedtobedonebyanysuchpersoninrelianceuponthecontentsofthisreport.

Contact

Foranyhelporassistancepleaseemailuson

concierge@

orvisitusat

.

Acknowledgements

RhythmVijayvargiya

rhythm.vijayvargiya@

PurushothamKittane

purushotham.kittane@

VaibhavParikh

vaibhav.parikh@

WewouldliketothankSauravKumar,HiranyaBhandarkar,andNishkaKapoorfortheircontributiontotheresearchpaper.

?NishithDesaiAssociates2024Provideduponrequestonly

Contents

Introduction1

DimensionsofDeepfakes:Meaning,TypesandtheTechnologies3

A.Meaning3

B.TypesofDeepfakes4

C.TechnologiesbehindDeepfakes7

10

TheImpactofDeepfakes

A.UseCasesofDeepfakeTechnology10

B.MisusesofDeepfakeTechnology12

16

LegalandRegulatoryImplications

A.InternationalRegulatoryInterventions16

B.IndianLegalandRegulatoryImplications21

38

WayForward

?NishithDesaiAssociates2024Provideduponrequestonly

Introduction

Weareintheeraofartificialintelligence(“AI”),anditissafetosaythattoday,thespeedoftechnologicalbreakthroughsisdirectlyproportionaltothespeedoftransmissionofinformationaswellasmisinformation.Contentalterationormanipulationisanage-oldconcept,1buttheeasyaccessibilityofvarioustoolshascontributedtothegrowthrateofonlineorchestratedcontentincreasingby400%everyyear.2

Atpresent,deepfakesareoneofthemostadvancedformsofsyntheticallygeneratedmediaanditispredictedthattheycouldaccountforupto90%oftheonlineavailablecontentintheupcomingyears.3

Oneofthefirsttechnologiesthatproduceddeepfake-likeresultswastheVideoRewriteProgramin1997,4whichautomatedfacialreanimationinvideos.Basedonasimilarconcept,theGenerativeAdversarialNetwork5(“GAN”)wasintroducedin2014,whichwasfurtherimprovisedbyNvidia6in2017toproducegoodqualityforgedimages.

WiththeGANalgorithmsslowlycatchingtraction,laterin2017,theterm“deepfakes”wascoinedwhenanunidentifieduseronthesocialmediaplatformReddithaddevelopedanalgorithm,7thattheuserusedtotransposecelebrityfacesontopornographiccontent.8Thelikenessofthecelebritieswassuperimposedinthepornographiccontenttotheextentthatitappearedtobetrue.Owingtothenatureofthecontentbeingshared,itinstantlybecameviralandwidespread.Theunidentifieduserusedtogowiththeusernamedeepfakes,andhence,thetechnologycametobecommonlyreferredtoasdeepfaketechnology.

Essentially,deepfakesreferto“fake”contentthatiscreatedusing“deeplearning”technology.9Apartfromthisoversimplifiedmeaningofdeepfakes,ithasalsobeendefinedbytheOxfordUniversityPress10as:“avideoofapersoninwhichtheirfaceorbodyhasbeendigitallyalteredsothattheyappeartobesomeoneelse,typicallyusedmaliciouslyortospreadfalseinformation.”WiththeadvancesinAI-synthesizedtechniques,deepfakesarenowalsocapableofcreatinghighlyrealisticallysoundedvoices.11

Today,thetechnologyiswidelyknownforcreatingrealistic-lookingimagesandvideosofpeopleandobjectsthatmayormaynotexist.Aboutninety-fivepercentofthedeepfakecontentwasintheformofnon-consensualporntillDecember2018,andRanaAyyub’scase12wasoneofthebiggestexamplesinthedeepfakehistorytodepictthedepthofrevengepornplottingthroughthistechnology.

1Datamanipulationhasbeeninpracticesincethe1890s;Pleasesee:

/collections/spanish-american-war-in-motion-pictures/

articles-and-essays/the-motion-picture-camera-goes-to-war/remember-the-maine-the-beginnings-of-war/,

(lastaccessedOctober10,2024).

2Pleasesee:

/en/news-release/2022/10/27/2542944/0/en/Deepfake-content-on-the-internet-is-growing-at

-the-rate-of-a-whopping-400-year-on-year.html

,(lastaccessedOctober10,2024).

3Pleasesee:

/2023/01/22/business/media/deepfake-regulation-difficulty.html

,(lastaccessedOctober10,2024).

4Pleasesee:

/videorewrite/VideoRewrite.pdf,

(lastaccessedOctober10,2024).

5Pleasesee:

/pdf/1406.2661.pdf,

(lastaccessedOctober10,2024).

6Pleasesee:

/blog/generating-photorealistic-fake-celebrities-with-artificial-intelligence/,

(lastaccessedOctober10,2024).

7TheReddituser’ssoftwarewascalledFakeApp,andwasusedbyBuzzFeedin2018tocreateadeepfakevideoofBarackObamaaddressingthisissue.Pleasesee:

/craigsilverman/obama-jordan-peele-deepfake-video-debunk-buzzfeed

.

8Pleasesee:

/volumes/Vol97No22/7Vol97No22.pdf,

(lastaccessedOctober10,2024).

9Pleasesee:

https://iimk.ac.in/uploads/faculty/CAIS_20220810062848.pdf,

(lastaccessedOctober10,2024).

10Pleasesee:

/google-dictionary-en/,

(lastaccessedOctober10,2024).

11Pleasesee:

/pdf/2005.13770.pdf,

(lastaccessedOctober10,2024).

12Pleasesee:

/archive/in/entry/deepfake-porn_in_5c1201cfe4b0508b213746bd

,(lastaccessedOctober10,2024).

?NishithDesaiAssociates2024Provideduponrequestonly1

Introduction

However,overtime,inaround2019,theusecasesofdeepfaketechnologystartedtocomeintothelimelighttoo.13BigTechorganizationslikeMicrosoft,14Google,15andSamsung16adoptedGANforcontentgeneration.Currently,theusesofdeepfaketechnologycanbeseeninthemedical,17entertainment,18marketing,19andfashion,20industriesamongothers;settingsomenoteworthyexamplesanddepictingthatthetechnologycanhaveanarrayofbeneficialusestoo.

Irrespectiveoftheeventsthatbroughtdeepfaketechnologyintothefocusofattention,thepossibilitiesthatarisewithitsuseareendless.Additionally,thereasonforitssignificantbreakthroughishowconvincingthesemediaemployingdeepfaketechnologyaretotheperceptiblehumanmind,andastimeprogresses,thesealteredmediaarebecomingincreasinglyclosertoreality.21

Contentmanipulationhasnowbecomemainstreamandeasilyaccessible,withthequalityofforgedcontentbeingsohighthatitbecomesimpossibletofilteroutwithabareeye.However,itisimportanttoacknowledgethatwiththequalityofthisfabricatedcontentrising,theimplicationsandinturnliabilitieswouldrisetoo.

Inthispaper,wehavesystematicallyexaminedthetypesandunderlyingtechnologiesbehinddeepfakes,followedbytheuseandmisusecasesofsuchtechnology.Additionally,wehavediscussedthelegal,regulatory,andethicalimplicationsofdeepfakesinIndiaandotherjurisdictions.

13Pleasesee:

/impressive-applications-of-generative-adversarial-networks/,

(lastaccessedOctober10,2024).

14Pleasesee:

/microsoft-has-made-their-own-ai-powered-image-generator-and-its-pretty-meh/,

(lastaccessedOctober10,2024).

15Pleasesee:

/2020/11/using-gans-to-create-fantastical.html

,(lastaccessedOctober10,2024).

16Pleasesee:

/global/behind-the-snapshot-how-the-galaxy-s21s-ai-improves-your-photos-in-the-blink-of-an-eye-sin-

gle-take

,(lastaccessedOctober10,2024).

17Pleasesee:

/reader/sd/pii/S0300571222002676?token=5DA9F8AD14B7D8634EBC69B7FCBB9E0414B10BAEF3C33865

833CC51F36FD35144A6558CDA050505AE667AC4F6C17900C&originRegion=eu-west-1&originCreation=20230225021628

,

(lastaccessedOctober10,2024).

18Pleasesee:

/channel/UCi38HMIvRpGgMJ0Tlm1WYdw

,(lastaccessedOctober10,2024).

19Pleasesee:

/news/tech/cadburys-new-ai-tool-will-let-you-create-free-ad-with-shah-rukh-khans-face-and-

voice-4358408.html

,(lastaccessedOctober10,2024).

20Pleasesee:

/sites/katiebaron/2019/07/29/digital-doubles-the-deepfake-tech-nourishing-new-wave-retail/?sh=2d473f-

bc4cc7,

(lastaccessedOctober10,2024).

21Pleasesee:

/publication/338144721_Deepfakes_Trick_or_treat

,(lastaccessedOctober10,2024).

?NishithDesaiAssociates2024Provideduponrequestonly2

UnmaskingDeepfakes—Legal,RegulatoryandEthicalConsiderations

DimensionsofDeepfakes:Meaning,TypesandtheTechnologies

A.Meaning

Deepfakesaretypicallyunderstoodtobemanipulatedmedia.1However,tojustsaythatdeepfakesaresynthesizedmediacreatedbyusingidentity-swappingalgorithmswouldnotdojusticetothedepthofthisconcept.Inaworldwhereeveryimageorvideoontheinternetmaybealteredtosomedegreeusingdeeplearning,itbecomesdifficulttoagreeonacommondefinitionofdeepfakessincethelinewhereapictureorvideobecomes“manipulated”or“fake”isblurred.

Nevertheless,theusageofdeeplearningtechnologytoproducefakecontentcontributedtowardstheportmanteau–Deepfakes.2Eventhoughthereisnogloballyagreedupondefinitionfordeepfakes,theycametobe

understoodconventionallyas“fakeimagescreatedbyanadvancedimageediting,deeplearningsoftware.”3

Tounderstandthenatureofdeepfakes,itiscrucialtoestablishthatdeeplearningisamachinelearningtoolthatformsthebasisofpowerfulalgorithms.Thesealgorithmsaredesignedtouseenormousvolumesofdatatolearntoperformspecifictasks.Simplyput,deeplearningisatypeofAI,anditusesartificialneuralnetworkstomimicthelearningprocessofthehumanbrain.4Deeplearningisafundamentalelementofthedeepfakephenomenon,asthealgorithmsfeedontotheexistingandavailabledata5tocreatefakeimagesandvideosthathumanscannotdistinguishfromauthenticones.6

Thereisnodistinguishabledivisionthatseparates‘a(chǎn)cceptable’formsofaugmentationorspecialeffectstomediafromthosewhichmaytypicallybeseenasmanipulated.Although,themanipulatedcontentthatisnotgeneratedusingAIordeeplearningtechnologyusuallybearsalowerqualityandiseasiertopointoutfromthegenuinecontent.Theserelativelypoorlyengineeredphotosandvideosareusuallyreferredtoascheapfakesorshallowfakes.7

AJuly2021studyundertakenfortheEuropeanParliamenthasdefineddeepfakesas“manipulatedorsyntheticaudioorvisualmediathatseemauthentic,andwhichfeatureapersonthatappearstosayordosomethingtheyhaveneversaidordone,producedusingartificialintelligencetechniques,includingmachinelearninganddeeplearning.”8

1Pleasesee:

/news/2020/01/enforcing-against-manipulated-media/,

(lastaccessedOctober10,2024).

2Pleasesee:

/65354450/IRJET_V7I1265-libre.pdf?1609941112=&response-content-disposition=inline%

3B+filename%3DIRJET_A_Brief_Study_on_Deepfakes.pdf&Expires=1677852603&Signature=Ybf6GhM91ZkpLWPUmRAl6WrpJBbQXXFkHxuLIP

wsmhlwde99WwNaj-i64CUtQLrfiJ2aZLYjXMWAa75rcTHqXp3lClYx5yJAPx-q9od8qTUwTeV2VPlViuq7SeHMDzUzL40mWtmQW7VuPlt0MUkiq3D

LlBz0xhCaVqqoc5d2OiCStVaufRq5Jh9gCulojTMAzl15b9uLPTVeMcZFtadesEcjyxhJMSTOX0MiZStQosOp81sy91AFmeUTzBMOQbVk-r-QZiEBx

A2frryvmMHJrGzg6kpZ1ssHncdwam-P7j1SaeYXU-BLG0PEYlYGjryJgxIrcxUFIYWa6f5R9XdYcA__&Key-Pair-Id=APKAJLOHF5GGSLRBV4ZA

,

(lastaccessedOctober10,2024).

3Pleasesee:

/cgi/viewcontent.cgi?article=1742&context=honors-theses

,(lastaccessedOctober10,2024).

4Pleasesee:

/in-en/topics/deep-learning

,(lastaccessedOctober10,2024).

5Pleasesee:

/nl-en/blogs/insights/deepfakes-how-prepare-your-organization

,(lastaccessedOctober10,2024).

6Pleasesee:

/pdf/1909.11573.pdf,

(lastaccessedOctober10,2024).

7Pleasesee:

/wp-content/uploads/2019/09/DS_Deepfakes_Cheap_FakesFinal-1-1.pdf

;:

https://www.lexisnexis.co.uk/legal/

guidance/deepfakes#:~:text=A%20deepfake%20is%20a%20form,a%20realistic%20but%20fake%20video,(lastaccessedOctober10,2024).

8Pleasesee:

https://www.europarl.europa.eu/RegData/etudes/STUD/2021/690039/EPRS_STU(2021)690039_EN.pdf,

(lastaccessedOctober10,2024).

?NishithDesaiAssociates2024Provideduponrequestonly3

UnmaskingDeepfakes—Legal,RegulatoryandEthicalConsiderations

DimensionsofDeepfakes:Meaning,TypesandtheTechnologies

Therehasbeenaplethoraofattemptstodefinedeepfakes,anditisanarduoustasktofindanagreeabledefinitionforthem.Although,thefollowingelementsseemtobelargelyacceptedasbeingfundamentallypartofwhatdeepfakesare:

a.Alterationofmediausingdeeplearningtechnologicaltools;

b.Alteredmediathatinvolvestheassumptionofanidentityofaperson;and

c.Anuntruealteredmediadepictedastruetothecasualviewer.

B.TypesofDeepfakes

Mostdeepfakevideosinvolvefacialmanipulation,whereaface,faces,orpartsthereofaremanipulatedandthensuperimposedorinsertedonanotherfaceorpartthereof.9Themanipulationmaybestatic,asinthecaseofanimage,ordynamicasinavideo.Thedeepfaketechnologymaybeusedtocreatedoctoredcontentthroughreenactment,replacement,editing,andaudioorvisualsynthesis.10

Reenactment11

Reenactment,commonlyunderstoodasexpressionswap,isessentiallywheretheimagesarefedintothealgorithmtodrivetheexpressionsofapersonandcreateadeepfake.Areenactmentdeepfakeallowsthecreatortoimpersonateanindividual’sappearanceandinturn,controlwhattheyappeartosayordo.Thisisoneofthemostdangeroustypesofdeepfakes,astheycanbeusedtoperformactsofdefamation,spreadmisinformation,andtamperwithevidence.

FacialReenactment

GazeMouthExpressionPoseComplete

Source:Medium12

9Pleasesee:

https://iimk.ac.in/uploads/faculty/CAIS_20220810062848.pdf,

(lastaccessedOctober10,2024).

10Pleasesee:

/pdf/2004.11138.pdf,

(lastaccessedOctober10,2024).

11Pleasesee:

.au/cgi/viewcontent.cgi?article=2530&context=theses_hons

,(lastaccessedOctober10,2024).

12Pleasesee:

/voxel51/have-deepfakes-influenced-the-2020-election-c0fc890aca0f

.,(lastaccessedOctober10,2024).

?NishithDesaiAssociates2024Provideduponrequestonly4

DimensionsofDeepfakes:Meaning,TypesandtheTechnologies

Replacement13

Replacement,commonlyunderstoodasfaceoridentityswap,iswhentheimagesarefedintothealgorithmtoreplacethefaceofonepersoninanimageorvideowiththatofanotherperson.Thismethodposesdangertoo,asitcanbeusedfortheproductionofrevengepornographicvideos,cheating,andfinancialfraud.Someoftheusecasesofthismethodincludegeneratingmemesorsatiricalcontent,orfaceswappingforanonymiza-tionofone’sidentityinpubliccontent.

FaceReplacement

TransferSwap

Source:InfoQ14

Editing15

Througheditingorattributealteration,aportionofthefaceismanipulatedtoachieveadifferentresult.Mediaretouchingdonethroughthismethodcomprisesalteringfacialfeaturessuchasgender,age,ethnicity,etc.Thismethodismisusedwidely,forinstance,fortheremovalofavictim’sclothesforhumiliationorentertainment.Atthesametime,thismethodofcreatingdeepfakesisusedforentertainmentandeasyeditingpurposestoo.

FaceEditing

HairArticleAgeBeautyEthnicity

Source:InfoQ16

13Pleasesee:

/blog/how-easy-is-it-to-make-and-detect-a-deepfake/,

(lastaccessedOctober10,2024).

14Pleasesee:

/article/u7jtw13waskch2mpl918

.

15Pleasesee:

/2313-433X/9/1/18

,(lastaccessedOctober10,2024).

16Pleasesee:

/article/u7jtw13waskch2mpl918

,(lastaccessedOctober10,2024).

?NishithDesaiAssociates2024Provideduponrequestonly5

DimensionsofDeepfakes:Meaning,TypesandtheTechnologies

VisualSynthesis17

Visualsynthesisallowsthecreationofadeepfakewithnotargetimagesorvideosasabasis.Entirefacesynthesisisusuallybasedondatasetsthatareeasilyavailableonline.Thismethodallowsthecreationofunrealpersonasonlineandmaybeusefulfornon-personalcommunicationwhileatthesametimeholdingthepotentialforfraudorthespreadofmisinformation.Fore.g.,thismethodalsobeusedtogeneratenon-existingfaceimagesforthecharactersinmoviesandgames.

EntireFaceSynthesis

FakeReal

Source:Medium18

AudioSynthesis19

Incasesofaudiosynthesis,anaudioclipofapersonspeaking,oreventhespeechinwritingisobtained,andanaudioclipisproducedoutofthespeechortextinthetoneandstyleofachosentarget,therebygivingtheimpressionthatthetargetpersonisspeaking.Thismethodposesathreatasitmakesphone-callfraudsverystraightforwardforthescammers.Although,speechsynthesishasalsobeenrelieduponinthemedicalindustrytohelppatientswithimpairment.

DataCollection

SpeakerEncoding

ModelTrainingTestingand

Evaluation

>

DistributionandUse

AudioData

Preprocessing

SpeechSynthesisandoptimization

Source:BotTalk20

17Pleasesee:

/pdf/1912.04958.pdf,

(lastaccessedOctober10,2024).

18Pleasesee:

/deepfakes-production-detection-using-various-deep-learning-methodolo-

gies-3221e6002dd2

,(lastaccessedOctober10,2024).

19Pleasesee:

/ftp/arxiv/papers/2111/2111.14203.pdf,

(lastaccessedOctober10,2024).

20Pleasesee:

https://bottalk.io/learn-with-bottalk/everything-about-deepfake-voice/,

(lastaccessedOctober10,2024).

?NishithDesaiAssociates2024Provideduponrequestonly6

UnmaskingDeepfakes—Legal,RegulatoryandEthicalConsiderations

DimensionsofDeepfakes:Meaning,TypesandtheTechnologies

C.TechnologiesbehindDeepfakes

MethodsofCreation21

Thereexistsanarrayofmachinelearningtechniquesandalgorithmswhichcanbeusedtocreatedeepfakes,andthequalityoftheresultisdependentnotonlyonthequalityofthealgorithmbutalsoonthequalityandquantityofdatausedtotrainthealgorithm.Itisnocoincidencethatthemajorityofdeepfakevideosontheinternetinvolvefamouspeople,astheyhavelargeamountsofpublicimagesandvideosavailablethatcanbeusedtotrainthealgorithm.

Deepfakesareusuallycreatedusingvariationsorcombinationsofgenerativenetworksandencoder-decodernetworks.DifferentkindsofGenerativeNeuralNetwork(“GNN”)architecturesareusedforthegenerationofartificiallyorchestratedcontent.Aneuralnetworkisessentiallyanalgorithmicmodelwhichgeneratescontentbasedonaninput.BelowarethedifferentGNNsusedforthecreationofdeepfakes.

I.Encoder-DecoderNetworks(“EDN”)

AnEDNismadeupoftwoormorealgorithmsthatcompressanddecompresspictures.Deepfakesarecreatedusingautoencoders,whichrecreateasubjectbasedontheinformationrelativetothedataprovidedtoit.Thesubjectsusedtocreatethedeepfakemusthaveasmanysimilaritiesaspossibletotheintendedoutputsothatthesharedencodercanidentifymeaningfulfeaturesandtransferthemappropriately.

Thequantityandqualityofdatapresentedtothealgorithmduringtheprocessisdirectlyproportionatetotheresult’squality.Autoencoderstrainedonmillionsofphotosofafacefromvariousanglesandunderavarietyoflightingconditionswillperformfarbetterandproducemorerealisticresultsthanencoderstrainedonafewhundredimageswithlittlevariationacrossthem.

II.ConvolutionalNeuralNetwork(“CNN”)

ACNNlearnspatternhierarchiesindataandishenceaveryefficienttoolfordealingwithpictures.

CNNhastheuniquecapacitytoextractfeaturesfromimages,whichmaythenbeutilizedinavariety

ofapplications.ACNNextractsalayerofcharacteristicsfromadatasetandappliesittoaninputimage;bycombiningnumerouslayersofthesecharacteristics,itispossibletoconstructrealistic-lookingdeepfakes.

III.GenerativeAdversarialNetworks(“GAN”)

GANwasfirstintroducedin2014andhassincebecomeoneofthemostpopulartoolsforcreatingdeepfakes.GANsaremadeupoftwoneuralnetworksthatcompetewithoneanother.Theycan“l(fā)earnfromtheirblunders”anddetectpatternsinenormousvolumesofvisualdata.Severallicensedpictureandvideoeditingsoftware,augmentedandvirtualrealityapplications,andcutting-edgemedicalimagingtoolsusethem.22TherearetwopopularimagetranslationframeworksthatusethefundamentalprinciplesofGANsinthecreationofdeepfakes:(i)image-to-imagetranslationwhichisoftenrelieduponforgeneratinghigh-resolution

21Pleasesee:

/pdf/2004.11138.pdf,

(lastaccessedOctober10,2024).

22Pleasesee:

/impressive-applications-of-generative-adversarial-networks/,

(lastaccessedOctober10,2024).?NishithDesaiAssociates2024Provideduponrequestonly7

DimensionsofDeepfakes:Meaning,TypesandtheTechnologies

imagerywithbetterfidelity,and(ii)CycleGANwhichisacombinationoftwoGANsandcanbeusedforobjecttransfigurationandstyletransfer.

IV.RecurrentNeuralNetworks(“RNN”)

AnRNNisaformofneuralnetworkthatremembersitsinternalstateafterprocessingadatasetandcansubsequentlybeusedtoprocessfurtherdatasets.RNNsarefrequentlyemployedindeepfakegenerationtomodifyaudioand,insomecases,video.

Withtheusageoftheabove-discussedunderlyingtechnologies,multipleconsumer-gradewebsitesandapplicationshavealsobeencreatedthatallowuserstoproducedeepfakes.

SomeoftheseincludeFakeApp,FaceSwap,DeepFaceLab,DFaker,FaceSwap-GAN,DeepFake-tf,ZAO,Auto-FaceSwap,FSGAN,FewShotFace,andStarGAN.23

Further,videoeditingsoftwarelikeAdobeAfterEffectsorWondershareFilmoracanbeusedtocreatecheapfakes,ifnotdeepfakes.Mostoftheseapplicationsarepubliclyavailableandcanproducealarge

rangeofmanipulateddata–dependingonthequalityoftheimagesfedandtheplatformused.

MethodsofDetection

Deepfakesarearesultofsuperimpositionwhichbecomespossiblethroughtheextensivecombinationsofdatasetsandthetechnologiesdiscussedintheprevioussection.Sincethealgorithmsputtouseforthemanipulationareonly“editing”pre-existingdata,gapscontinuetoremainintheresults.Minutefeatureslikeeyemovements,lighting,andshadowsoftencreatediscrepanciesthatallowfordeepfakedetection.24Thereexistdifferentsetsoftoolsforthedetectionofdeepfakesinimages,andinvideos.25Forinstance,thedetectionofdeepfakeimagesmaybedonethroughthedeploymentofthefollowingmethods:

DetectionofGAN-generatedimagesmaybedonethroughanimagepreprocessingstepanddifferentiatingtheswappedimagesfromthegenuine.26

Atwo-phaseddeeplearningmethodmaybeimplemented,inwhichthefirstphaseusesafeature

extractorthroughthecommonfakefeaturenetworkandthesecondphasecategorizesthefakeandrealimagesbasedontheresultsofthefirst.27

ACNNmodelmaybeusedtoidentifytheimageswherethefacialexpressionsaremaliciouslytamperedwith.

Simila