Kā pētnieki mēra, atklāj un salīdzina AI manipulāciju

Autori : Enes Altuncu, ea483@kent.ac.uk (University of Kent, UK) Virginia N. L. Franqueira, V.Franqueira@kent.ac.uk (University of Kent, UK) Shujun Li, S.J.Li@kent.ac.uk (University of Kent, UK) Autori : Enes Altuncu, ea483@kent.ac.uk (Kenta Universitāte, Apvienotā Karaliste) Virginia N. L. Franqueira, V.Franqueira@kent.ac.uk (Kenta Universitāte, Apvienotā Karaliste) Shujun Li, S.J.Li@kent.ac.uk (Kenta Universitāte, Apvienotā Karaliste) Abstract Abstraktā Pēdējie uzlabojumi AI, jo īpaši dziļās mācīšanās, ir veicinājuši ievērojamu pieaugumu jaunu reālistiski izskatīgu sintētisko mediju (video, attēlu un audio) radīšanā un esošo mediju manipulēšanā, kas ir novedis pie jauna termina “deepfake” izveidošanas. Pamatojoties gan uz pētniecības literatūru un resursiem angļu, gan ķīniešu valodā, šis dokuments sniedz visaptverošu pārskatu par dziļo viltošanu, kas aptver vairākus svarīgus šī jaunā jēdziena aspektus, tostarp 1) dažādus definīcijas, 2) plaši izmantotos veiktspējas rādītājus un standartus, un 3) ar dziļo viltošanu saistītos datu kopumus, izaicinājumus, sacensības un kritērijus. : Deepfake, aptauja, definīcija, datu kopumi, kritēriji, izaicinājumi, sacensības, standarti, darbības rādītāji. Atslēgvārdi 1 Introduction Pēdējie uzlabojumi AI un mašīntulkošanas jomā ir palielinājuši spēju ražot reālistiskākus plašsaziņas līdzekļus, piemēram, video, attēlus un audio. Īpaši mūsdienīgas dziļās mācīšanās metodes ļāva radīt “deepfakes”, manipulētus vai sintētiskus plašsaziņas līdzekļus, kuru realitāti cilvēka acs nevar viegli atpazīt. Lai gan dziļās viltības ir salīdzinoši jauna parādība (tā pirmoreiz parādījās 2017. gada beigās), tās izaugsme ir bijusi ievērojama. Saskaņā ar 2019 un 2020 Deeptrace ziņojumiem par dziļās viltības stāvokli [ ], Deepfake video skaits angļu valodā runājošajā internetā pieauga no 7,964 (decembris 2018) līdz 14,678 (jūlijs 2019) līdz 85,047 (decembris 2020), kas ir 968% pieaugums no 2018. līdz 2020. gadam. 2 Šajā darbā mēs pārskatām esošo pētniecības ekosistēmu, kas saistīta ar dziļo viltojumu, dažādos aspektos, tostarp darbības rādītājus un standartus, datu kopumus, izaicinājumus, sacensības un kritērijus. turklāt mēs sniedzam 12 atlasīto ar dziļo viltojumu saistīto aptaujas rakstu meta-pārskatu, kas sistemātiski aptver vairākus papildu aspektus, izņemot minētos, piemēram, veiktspējas salīdzinājumu, galvenos izaicinājumus un ieteikumus. Neskatoties uz to, ka tas ir ļoti populārs termins, ir trūkst vienprātības par definīciju “deepfake” un robeža starp deepfakes un ne-deepfakes nav skaidrs samazinājums. Šim apsekojumam mēs pieņemam salīdzinoši iekļaujošāku pieeju, lai segtu visas manipulēto vai sintētisko mediju formas, kas tiek uzskatītas par deepfakes plašākā nozīmē. Mēs aptveram arī cieši saistītas tēmas, tostarp biometriku un multimediju forensiku, jo deepfakes bieži tiek izmantotas, lai uzsāktu prezentācijas uzbrukumus pret biometriku balstītām autentifikācijas sistēmām, un deepfakes atklāšanu var uzskatīt par daļu no multimediju forensikas. 1.1 Termina Deepfake definīcijas Kā norāda nosaukums, termins “deepfake” ir atvasināts no kombinācijas “deep” (attiecas uz (DL)) un "false". To parasti izmanto, lai atsauktu uz esošo plašsaziņas līdzekļu (attēlu, video un / vai audio) manipulāciju vai jaunu (sintētisko) plašsaziņas līdzekļu radīšanu, izmantojot DL balstītas pieejas. Visbiežāk apspriesti dziļi falsifikācijas dati ir viltus sejas attēli, viltus runas falsifikācijas un viltus videoklipi, kas apvieno gan viltus attēlus, gan viltus runas falsifikācijas. Kamēr vārdā "false" norāda uz manipulētiem vai sintezētiem plašsaziņas līdzekļiem, ir daudz labdabīgu deepfake tehnoloģijas pielietojumu, piemēram, izklaidei un radošajām mākslām. Šajā sakarā vēl viens termins "dziļa sintēze" ir ierosināts Tomēr šis jaunais termins nav plaši pieņemts. dziļa mācīšanās 60 Papildus vispārējas definīcijas trūkumam, kā jau minēts, robeža starp dziļajiem viltojumiem un ne-dziļajiem viltojumiem faktiski nav skaidra. Pirmkārt, dziļo failu atklāšana bieži seko ļoti līdzīgām pieejām tradicionālo viltojumu atklāšanai, kas ģenerēti, neizmantojot DL paņēmienus. Advanced detekcijas metodes ir arī sākušas izmantot DL, lai uzlabotu savu veiktspēju, bet viņiem nav obligāti jāzina, kā mērķa plašsaziņas līdzeklis tiek radīts (dziļi vai nē). Dažos aspektos var apgalvot, ka dziļo failu atklāšana neattiecas uz dziļo failu specifisko metožu izstrādi (pat ja daži pētnieki izvēlas to darīt), bet spēcīgāks un universālāks detektors, kas var tikt galā ar jebkādiem (dziļi vai nē) viltotiem plašsaziņas līdzekļiem. To var redzēt divās cieši saistītās tēmās: biometr Otrkārt, var apgalvot, ka vārds “dziļš” vārdā “dziļš” ne vienmēr attiecas uz “dziļās mācīšanās” izmantošanu, bet uz jebkuru “dziļo” (t.i., izsmalcinātu) tehnoloģiju, kas rada ļoti ticamu viltotu mediju. ] uzskatīja deepfake kā audio-vizuālo manipulāciju, izmantojot "tehniskās izsmalcinātības ... un tehnikas spektru". un , atsaucoties uz “zema līmeņa manipulāciju ar audiovizuālajiem plašsaziņas līdzekļiem, kas izveidoti ar (viegli) pieejamu programmatūru [vai bez programmatūras], lai paātrinātu, palēninātu, atjaunotu vai pārkontekstualizētu saturu”. 9 Lāčplēša Cheapfake 2018 Ļaunprātīga dziļa viltus aizlieguma likums1 Saskaņā ar 101. panta b) punktu: ” jēdziens “dziļš viltojums” nozīmē audiovizuālo ierakstu, kas izveidots vai izmainīts tādā veidā, ka ieraksts saprātīgam novērotājam kļūdaini šķiet autentisks ieraksts par personas faktisko runu vai rīcību. 2019 DEEP FAKES Atbildības likums2: Saskaņā ar 104. panta 3. punktu: “ Termins “dziļš viltojums” nozīmē jebkuru video ierakstu, filmu, skaņas ierakstu, elektronisko attēlu vai fotogrāfiju vai jebkuru tehnoloģisku runas vai uzvedības attēlu, kas būtiski izriet no tā; (a) kas, šķiet, autentiski attēlo jebkādu runu vai rīcību, ko veikusi persona, kura faktiski nav iesaistījusies šādā runā vai rīcībā; un ” (b) kuru ražošana lielā mērā bija atkarīga no tehniskajiem līdzekļiem, nevis no citas personas spējas fiziski vai mutiski pielīdzināt šādu personu. Kā mēs varam redzēt no iepriekš minētajām juridiskajām definīcijām par “dziļo viltošanu”, DL kā tehnoloģiju izmantošana vispār netiek minēta. 1.2 Apjoms un ieguldījums Pamatojoties uz iepriekš minēto diskusiju par dziļo viltojumu definīcijām, mēs varam redzēt, ka ne vienmēr ir vienkārši vai jēgpilni atšķirt dziļo viltojumu no ne-dziļo viltojumu. Turklāt mūsu uzmanības centrā ir veiktspējas novērtējums un salīdzinājums, robeža starp dziļo viltojumu un ne-dziļo viltojumu ir vēl neskaidrāka. Tas ir tāpēc, ka DL ir tikai īpaša (dziļāka) mašīnu mācīšanās (ML) forma, un kā rezultātā DL un ne-dziļo ML metodes kopīgi izmanto daudzas kopīgas koncepcijas, metrikas un procedūras. Neskatoties uz to, ka dziļo viltošanu var saprast daudz plašākā nozīmē, šajā darbā mums ir pietiekami šaurs fokuss, lai izvairītos no pārāk daudzām tēmām. Metrikām un standartiem mēs izvēlējāmies iekļaut visas parasti izmantojamās metodes, lai novērtētu vispārējās ML metodes, un tās, kas īpaši definētas, lai novērtētu dziļo viltojumu veidošanas vai atklāšanas metodes. Attiecībā uz datu kopām, izaicinājumiem, konkursiem un kritērijiem mēs izskatījām tos, kas saistīti ar viltus plašsaziņas līdzekļiem, uz kuriem attiecas aptaujas materiāli, kas saistīti ar dziļo viltošanu, un tos, kuriem ir skaidra atsauce uz terminu "dziļo viltošanu" vai līdzīgu terminu. Meta-pārskatīšanai mēs uzskatījām tikai aptaujas rakstus, kuru autori skaidri atsaucās uz terminu “deepfakes” meta datiem (titulu, abstraktu un atslēgvārdus). 2 Metodoloģija Šajā aptaujā aplūkotie pētījumu raksti (t.i., ar dziļo viltošanu saistītie pētījumu raksti) tika identificēti, izmantojot sistemātisku meklēšanu zinātniskajās datubāzēs, Scopus un China Online Journals (COJ) Sekojošie meklēšanas vaicājumi tika izmantoti, lai veiktu meklējumus attiecīgi Scopus un COJ: 3 (deepfake* OR deep-fake* OR “deep fake*”) UN (pārskats vai aptauja vai pārskats vai sistemati* OR SoK) (deepfake OR 深度偽造) UN (综述 OR 进展) Meklējumos tika atgriezti 41 aptaujas teksts angļu valodā un 15 aptaujas teksts ķīniešu valodā, no kuriem astoņi publicēti angļu valodā un četri publicēti ķīniešu valodā. Ar Deepfake saistītie izaicinājumi, konkursi un kritēriji tika identificēti, izmantojot vairākus avotus: atlasītos aptaujas dokumentus, pētniecības darbus no līdzautoru personīgajām kolekcijām, Google tīmekļa meklējumus un rokas pārbaudi tīmekļa vietnēs lielo AI saistīto konferenci, kas notika 2020. un 2021. gadā (kur šādi izaicinājumi un konkursi tiek regulāri organizēti). , ICCV, CVPR, AAAI, ICML, ICLR, KDD, SIGIR, WWW un daudzi citi. Turklāt tika sagatavots visaptverošs datu kopu saraksts, pamatojoties uz atlasītajiem aptaujas rakstiem un identificētajām problēmām, konkursiem un kritērijiem. Attiecīgie standarti tika identificēti galvenokārt ar pētījumu rakstiem, kas ietverti šajā aptaujā, līdzautoru personīgajām zināšanām un Google tīmekļa meklējumiem. Par darbības rādītājiem mēs aptvēruši tos, kas parasti tiek izmantoti, pamatojoties uz attiecīgajiem standartiem, aptaujas rakstiem un identificētajām problēmām, konkursiem un kritērijiem. 4 3 Deepfake saistītie veiktspējas rādītāji un standarti Šajā aptaujā mēs koncentrējamies uz veiktspējas novērtēšanu un dziļo viltus ģenerēšanas un atklāšanas metožu salīdzināšanu. Šādām veiktspējas novērtējumiem izmantotie rādītāji ir mūsu diskusiju centrā. Šajā sadaļā mēs pārskatām veiktspējas rādītājus, kurus parasti izmanto, lai novērtētu dziļo viltus ģenerēšanas un atklāšanas algoritmus. Ņemiet vērā, ka visas šajā sadaļā aprakstītās rādītājus parasti izmanto arī līdzīgu sistēmu veiktspējas novērtēšanai, kas nav paredzētas dziļo viltus ģenerēšanai vai atklāšanai. Pēdējā apakšiedaļā mēs arī īsumā apspriedīsim, kā saistītie veiktspējas rādītāji ir iekļauti formālos standartos. „Formālie standarti” attiecas uz standartiem, kurus definē pēc formālas procedūras, bieži vien vienu vai vairākas izveidotas standartizācijas struktūras, piemēram, Starptautiskā Standartizācijas organizācija (ISO). Starptautiskā elektrotehniskā komisija (IEC) Ņemiet vērā, ka mēs apsveram plašu dokumentu klāstu, ko standartizācijas struktūras, piemēram, Starptautiskā Telekomunikāciju savienība (ITU) ir definējušas kā standartus. ISO ieteikumi un tehniskie ziņojumi (TRs). 5 6 7 3.1 Apjukuma matrica Deepfake detekcija galvenokārt ir binārā klasifikācijas problēma. vai un izdala bināro vērtību, kas norāda, ka tā ir vai Piemēram, deepfake detekcijas sistēma ņems aizdomīgu attēlu kā ievadi, kas var būt vai un iznākums vai . Patiesībā pozitīvi Patiesībā negatīvi Pozitīvas prognozes Negatīva prognoze Patiesībā viltus Patiesībā reāli Prognozēts viltojums Reālā prognoze Būtisks instruments, ko izmanto, lai novērtētu bināro klasifikatoru, ir Tas apkopo klasifikācijas modeļa panākumus un neveiksmes. vērtības un otrā ass ir divi vērtības. klasifikācija ir (patiesas pozitīvās un patiesas negatīvās vērtības), ja faktiskās un paredzamās vērtības atbilst. (nepatiesi pozitīvi un nepatiesi negatīvi), ja faktiskās un prognozētās vērtības neatbilst. rāda bināro deepfake klasifikatoru (detektoru) neskaidrības matricu. divas šūnas zaļā krāsā, TP (skaits un TN (numurs no ), norāda pareizos prognozēšanas rezultātus, un divas šūnas sarkanā krāsā, FN (numurs un FP (numurs no ), norāda divu dažādu veidu kļūdas, veicot nepareizu prognozēšanas rezultātus. confusion matrix actual Prognozēts Patiesība / True / Correct Nepareizs / Nepareizs / Nepareizs 1 true positives true negatives false negatives false positives Tabula 1: Sajaukšanās matrica binārajam klasificētājam, lai atklātu dziļo viltojumu. fake (predicted) real (predicted) fake (actual) TP FN real (actual) FP TN Fake (patiesā) TP Apvienoto Nāciju Reālā (reālā) FP Mr 3.2 Precizitāte un atgādinājums Pamatojoties uz četrām pamatvērtībām, kas ieviestas sadaļā , t.i., TP, TN, FP un FN, mēs definējam divus svarīgus veiktspējas rādītājus binārajam klasificētājam – un . 3.1 precision recall Bināro klasifikatoru precizitāte ir definēta kā daļu no Sample starp visiem Saskaņā ar likuma “Par valsts un pašvaldību institūcijām” 1.panta pirmo daļu, tā var tikt uzskatīta par valsts vai pašvaldību institūciju. ( Tātad ) Patiesībā pozitīvi Pozitīvas prognozes 1 When the “natural” ratio between positive and negative samples is significantly different from the test set, it is often useful to adjust the weight of the false positives, which leads to the Saskaņā ar Līguma par Eiropas Savienību (turpmāk – Līgums) 2) kur 0 ir svars, ko nosaka attiecība starp negatīvo un pozitīvo paraugu. weighted precision 2 α > Bināro klasifikatoru atgādinājums ir daļa no Izmēģinājumi starp Sēklinieki, kā parādīts ekrānā ( Apjukuma matriksā tā ir īsto paraugu īpatsvars pirmajā rindā. Pozitīvas prognozes Patiesībā pozitīvi 3 Apsveriet piemēru ar bināro klasifikatoru, kas paredz, vai attēls no datu bāzes, kas satur gan deepfake, gan reālus (autentiskus) attēlus, ir viltots vai nē. klasifikācijas precizitāte ir pareizi klasificētu attēlu īpatsvars starp visiem attēliem, kas klasificēti kā deepfake. 3.3 Patiesas un nepatiesas pozitīvās likmes Koncentrējoties uz prognozētiem pozitīviem paraugiem, mēs varam arī definēt divas metrikas: (TPR) arī saukta par (CDR), kā prognozēto pozitīvo paraugu īpatsvars starp faktiski pozitīviem paraugiem un (FPR), arī saukta par (FAR), kā prognozēto pozitīvo paraugu īpatsvars starp faktiski negatīvajiem paraugiem, kā parādīts Eqs. ( ) and true positive rate correct detection rate false positive rate false alarm rate 4 ( TPR ir prognozēto pozitīvo paraugu īpatsvars pirmajā rindā un FPR ir prognozēto pozitīvo paraugu īpatsvars otrajā rindā. (Vienotības vārdā) Tāpat arī 5 recall 3 3.4 Patiesie un nepatiesie negatīvie rādītāji Līdzīgi kā patiesie un viltus pozitīvie rādītāji, mēs varam definēt divus citus rādītājus, kas koncentrējas uz negatīviem prognozētajiem rezultātiem: (TNR), kas norāda prognozēto negatīvo paraugu īpatsvaru faktisko negatīvo paraugu vidū, un (FNR) norāda prognozēto negatīvo paraugu īpatsvaru starp faktiski pozitīviem paraugiem, kā parādīts Eqs. ( un ( Tātad ) true negative rate false negative rate 6 7 3.5 Jūtīgums un specifika Dažos bināro klasifikatoru pielietojumos, jo īpaši bioloģijā un medicīnā, TPR un TNR tiek plašāk izmantoti, un tos bieži sauc par TPR un (TNR). The focus of these two terms is on the two types of correctness of the predicted results. These are less used in deepfake-related research, hence, we will not refer to them in the remainder of this paper. sensitivity specificity 3.6 Equal Error Rate Koncentrējoties uz kļūdu likmēm, mums ir jāapsver FPR un FNR. Šīs divas likmes parasti ir pretrunā viena ar otru, tāpēc, samazinot vienu likmi, parasti palielinās otra.Tāpēc, tā vietā, lai mēģinātu samazināt abus kļūdu likmes vienlaicīgi, kas parasti ir neiespējami, reālistiskāks uzdevums praktiskajās lietojumprogrammās ir atrast pareizo līdzsvaru, lai tie abi būtu zem pieņemama sliekšņa. Dažās lietojumprogrammās, piemēram, biometrikā, cilvēki ir īpaši ieinteresēti izveidot tā saukto (Eiropas Savienība) vai EER/CER dažām lietojumprogrammām ne vienmēr ir labs rādītājs, it īpaši, ja abiem kļūdu veidiem ir atšķirīgs nozīmes līmenis, piemēram, kritisko dziļo faktu (piemēram, viltus ziņu, kas var ietekmēt to, kā cilvēki balso) atklāšanai mēs bieži vien varam panest vairāk viltus pozitīvu (viltus trauksmes signālus) nekā viltus negatīvu (izlaistu trauksmes signālus). equal error rate crossover error rate 3.7 Precizitāte un F-score Papildus EER / CER, ir arī citi rādītāji, kas mēģina atspoguļot abu veidu kļūdas, lai sniegtu līdzsvarotāku norādi par bināro klasifikatoru kopējo veiktspēju. un (saukta arī par Abas metrikas var definēt, pamatojoties uz četrām pamatvērtībām (TP, TN, FP un FN). accuracy F-score F-measure Bināro klasifikatoru precizitāte tiek definēta kā daļu no Patiesie pozitīvie un patiesie negatīvie paraugi no kopējā klasificēto paraugu skaita, kā parādīts ekrānā (Eq. ). Pareizi prognozēts 8 Bināro klasifikatoru F-score faktiski ir metrikas ģimene. Saskaņā ar definīciju ( Tātad ) β 9 Visbiežāk izmantotais visu F-rādītāju izdevums ir tā sauktais , kas faktiski ir F-score ar = 1. More precisely, it is defined as shown in Eq. ( Tātad ) F1-score β 10 3.8 Saņēmēja darbības raksturīgās līknes un platība zem līknes (ROC) līknes parasti tiek izmantotas, lai novērtētu bināro klasifikatoru veiktspēju, kas rada prognozes rezultātu (vai varbūtību). Receiver operating characteristic Consider the following. Let be the set of all test samples and let the output scores ( ) (for all ∈ ) lie in the interval [ ] on the real line. Let Tātad [ ] be a prediction threshold for the model, and assume that the classifiers works as follows for all ∈ : S f s s S a, b t A, B s S To ir viegli saprast, jo = , all the samples will be classified as positive, leading to FN = TN = 0 so TPR = FPR = 1; while for = ir , all the samples will be classified as negative, leading to FP = TP = 0 so TPR = FPR = 0. For other threshold values between and , TPR un FPR vērtības parasti būs no 0 līdz 1. from Divi nepārtraukti, mēs parasti varam iegūt nepārtrauktu līkni, kas apraksta, kā TPR un FPR vērtības mainās no (0,0) līdz (1,1) 2D līmenī. t a t b a b t a b For a random classifier, assuming that (Tā kā ) vienmērīgi sadalīti uz [ ] for the test set, we can mathematically derive its ROC curve being the TPR = FPR line, whose area under the ROC curve (AUC) is 0.5. For a binary classifier that performs better than a random predictor, we can also mathematically prove that its AUC is always higher than 0.5, with 1 being the best possible value. Note that no binary classifier can have an AUC below 0.5, since one can simply flip the prediction result to get a better predictor with an AUC of 1 − AUC. The relationship between the ROC and the AUC is graphically illustrated in Figure . f s a, b 1 3.9 Log Loss Another widely used performance metric for binary classifiers that can return a probability score for the predicted label is Par bināro klasifikāciju ar patiesu etiķeti ∈ {0*,* 1} and an estimated probability = Pr( = 1), the log loss per sample is the negative log-likelihood of the classifier given the true label, defined as shown in Eq. ( ). log loss y p y 12 Given a testing set with samples, the log loss score of a binary classifier can be calculated using Eq. ( ), where ir 1 ja -th sample is true and 0 if false, and ˆ is the predicted probability of = 1. n 13 yi i y i Jē 3.10 Extension to Multi-class Classifiers All metrics that are defined based on the four basic values TP, TN, FP and FN can be easily extended to by considering the prediction to be true or false individually with respect to each class. For example, if the system is classifying animals (cats, dogs, horses, lions, tigers, etc.), then a true positive prediction of an image to be of a cat, would simultaneously be true negative predictions for the remaining classes (dogs, horses, lions, tigers, etc.). If an image of a cat is incorrectly predicted to be that of a dog, it would be a false negative with respect to a cat, a false positive with respect to a dog, and a true negative with respect to all other classes. multi-class classification 3.11 Perceptual Quality Assessment (PQA) Metrics By definition, the main goal of deepfakes is to make it hard or impossible for human consumers (listeners or viewers) to distinguish fake media from real media. Therefore, when evaluating the quality of deepfake media, the quality perceived by human consumers of the media is key. This calls for subjective assessment of the perceptual quality of the deepfake media as the “gold standard”. The most widely used subjective perceptual quality assessment (PQA) metric for audio-visual signals is (MOS), which has been widely used by the signal processing and multimedia communication communities, including digital TV and other multimedia-related consumer applications. As its name implies, MOS is calculated by averaging the subjective scores given by a number of human judges, normally following a numerical scale between 1 and 5 or between 0 and 100. MOS has been used in some deepfake-related challenges (see Section ) and also for evaluating and comparing the quality (realness/naturalness) of deepfake datasets (see Section ). mean opinion score 5.2 4.6 As a general subjective PQA metric, MOS has been standardised by the ITU . There are also ITU standards defining more specific subjective Video Quality Assessment (VQA) metrics and the standard procedures one should follow to conduct VQA user studies, e.g., ITU-T Recommendation P.910 “Subjective video quality assessment methods for multimedia applications” . Note that the ITU standards focus more on traditional perceptual quality, i.e., how good a signal looks or sounds, even if it looks or sounds not real (e.g., too smooth). On the other hand, for deepfakes, the focus is rather different because what matters is the realness and naturalness of the created media, i.e., how real and natural it looks or sounds, even if it is of low quality. To some extent, we can also consider realness and naturalness as a special aspect of perceptual quality. 8 9 One major problem of subjective PQA metrics like MOS is the need to recruit human judges and to have a well-controlled physical testing environment and protocol, which are not easy for many applications. To help reduce the efforts and costs of conducting PQA-related user studies, various objective PQA metrics have been proposed, where the term “objective” refers to the fact that such metrics are human-free, i.e., automatically calculated following a computational algorithm or process. Depending on whether a reference exists, such objective PQA metrics can be largely split into three categories: full-reference (FR) metrics (when the original “perfect-quality” signal is available as the reference), reduced-reference (RR) metrics (when some features of the original “perfect-quality” signal are available as the reference), and no-reference (NR) metrics (when the original signal is unavailable or such an original signal does not exist). For deepfakes, normally NR or RR metrics are more meaningful because the “fake” part of the word means that part of the whole data does not exist in the real world, hence a full reference cannot be obtained. RR metrics are still relevant because deepfakes are often produced for a target’s specific attributes (e.g., face and voice), where the reduced reference will be such attributes. NR metrics will be useful to estimate the realness and naturalness of a deepfake, simulating how a human judge would rate it in a controlled subjective PQA user study. PQA is a very active research area and many PQA metrics have been proposed, some of which have been widely used in real-world products and services, e.g., (MSE), (PSNR) and (14), (15), un (16), attiecīgi, kur X = {xi} n i ir atsauce (originālais signāls), Y = {yi} n i ir signāls, kura vizuālā kvalitāte tiek novērtēta, n ir pikseļu skaits X un Y , L ir maksimāli iespējamā pikseļu vērtība X un Y (piemēram, 255 8 bitu pelēka mēroga attēliem), c1 = (k1L) 2 un c2 = (k2L) 2 ) ir divi stabilizējošie parametri (k1 = 0,01 un k2 = 0,03 pēc noklusējuma). mean squared error peak signal-to-noise ratio structural similarity index measure 3.12 More about Standards Many of the basic performance metrics described in this section have been widely used by deepfake researchers as de facto standards, e.g., EER, log loss and MOS have been widely used in deepfake-related challenges (see Section Arī precizitātes, atsaukšanas un F1 rezultātu kombinācija ir plaši izmantota, lai novērtētu bināro klasifikatoru veiktspēju. Lai gan līdz šim ir bijuši vairāki ITU standarti par PQA, šķiet, ka nav daudz standartizācijas pūļu par bināro klasifikatoru veiktspējas rādītājiem. , a standardisation subcommittee (SC) focusing on AI under ISO/IEC JTC 1 , the joint technical committee for standardising “information technology”. 5 10 11 ISO/IEC JTC 1/SC 42 ir izstrādājis ISO/IEC TR 24029-1:2021 “Mākslīgais intelekts (AI) – Neironu tīklu izturības novērtējums – 1. daļa: Pārskats”. , a technical report (TR) that systematically covers many commonly used performance assessment concepts, methods and metrics. Although the technical report has “neural networks” in its title, most performance assessment concepts, methods and metrics included are common ones for all supervised machine learning models. 12 In terms of performance metrics, two other ongoing work items of the ISO/IEC JTC 1/SC 42 that deserve attention are as follows: ISO/IEC DTS (Draft Technical Specification) 4213 “Information technology – Artificial Intelligence – Assessment of machine learning classification performance” 13 ISO/IEC AWI (Approved Work Item) TS (Technical Specifications) 5471 “Artificial intelligence – Quality evaluation guidelines for AI systems” 14 While the ISO/IEC JTC 1/SC 42 was created very recently, another standardisation subcommittee under ISO/IEC JTC1 has a much longer history of nearly 20 years: the ISO/IEC JTC 1/SC 37 that focuses on biometrics-related technology. This standardisation subcommittee is highly relevant for deepfake since deepfake faces can be used to spoof biometrics-based user authentication systems. In this context, the following three standards are of particular relevance: 15 : This standard covers general metrics about evaluating biometric systems. Two major metrics in this context are (FAR) un (FRR), kas attiecas uz standarta FPR un FNR, attiecīgi. šis standarts arī novērš vienciparu metrikas izmantošanu, ieskaitot EER un AUC (kas agrāk tika plaši izmantotas ar biometriku saistītos pētījumos). ISO/IEC 19795-1:2021 “Information technology – Biometric performance testing and reporting – Part 1: Principles and framework” 16 false accept rate false reject rate : This standard defines a general framework about (PAD) mechanisms, where the term “ ” refers to the “ ”. It focuses on biometric recognition systems, where a PAD mechanism is a binary classifier trying to predict presentation attacks (also called attack presentations, e.g., fake faces) as positive and bona fide (real) presentations as negative. ISO/IEC 30107-1:2016 “Information technology – Biometric presentation attack detec-tion – Part 1: Framework” 17 presentation attack detection presentation attack presentation of an artefact or of human characteristics to a biometric capture subsystem in a fashion intended to in-terfere with system policy : This standard defines a number of special performance metrics for evaluating PAD mechanisms standardised in the ISO/IEC 30107-1:2016. Three such metrics look at error rates: (APCER) atsaucoties uz standarta FPR, (NPCER/BPCER) referring to the standard FNR, and (ACER) that is defined as the average of the APCER and the NPCER/BPCER. Such metrics have been used in biometrics-related challenges such as Face Anti-spoofing (Presentation Attack Detection) Challenges . When deepfake images or videos are used to spoof a biometric system, such standardised metrics will become relevant. ISO/IEC 30107-3:2017 “Information technology – Biometric presentation attack detection – Part 3: Testing and reporting” 18 attack presentation classification error rate normal/bona fide presentation classification error rate average classification error rate 19 3.13 Discussion: Performance Metrics & Standards This section provided a comprehensive summary of performance metrics used for evaluating and bench-marking binary classifiers. It is rare that all such metrics are used for a specific application. Instead, one or several are chosen based on specific needs. For a deepfake detection system as a binary classifier, many researchers have chosen to use overall metrics such as accuracy, AUC, EER and log loss, but the combination of precision, recall and F1-score is also common. Some deepfake-related challenges and competitions have introduced their own specific metrics, some of which will be described in Section . The use of different performance metrics can make comparison of different reported results more difficult, so we hope the expected new ISO/IEC standard particularly ISO/IEC 4213 will help. 5 It is worth mentioning that, in addition to evaluating performance of deepfake detectors, the introduced performance metrics for evaluating binary classifiers can also be used to evaluate performance of deepfake generation methods by considering how deepfake detectors fail. For instance, organisers of the Voice Conversion Challenge 2018 and 2020 used this approach to benchmark how well voice conversion (VC) systems can generate high-quality fake speech samples. Another point we would like to mention is that for deepfake videos there are two levels of performance metrics: those at the frame level (metrics of each frame), and those at the video level (metrics for the whole video). Generally speaking, the latter can be obtained by averaging the former for all frames, potentially following an adaptive weighting scheme, so that more important (key) frames will be counted more. 4 Deepfake-Related Datasets Šajā sadaļā mēs aptveram visus ar deepfake saistītos datu kopumus, kurus mēs identificējām no deepfake saistīto aptaujas rakstu meta-pārskata, ar deepfake saistītās problēmas, konkursi un kritēriji, kas aptverti, viena tiešsaistes datu kopa, kas saistīta ar deepfake GitHub , and the co-authors’ personal collections. Table shows basic information about these datasets. We explain them in four categories: deepfake image datasets, deepfake video datasets, deepfake audio/speech datasets, and hybrid deepfake datasets (mainly mixed image and video datasets). 20 2 Note that many datasets of real (authentic) media were also used by deepfake researchers for two purposes. First, any detectors would need both fake and real media to demonstrate their performance. Second, real media have also been used to train deepfake generators as the training set. In this section, we include only datasets containing deepfake media, some of which contain both deepfake and real media. Some datasets, especially those created for deepfake-related challenges and competitions, have separate subsets for training and evaluation (testing) purposes. The split is necessary for such challenges and competitions, but not very useful for people who just want to use such datasets. Therefore, in this section when introducing such datasets we will ignore that level of details and focus on the total number of data including the number of real and fake samples. 4.1 Deepfake attēlu datu kopas [Lasīt ]: This dataset contains 4,310 images, including 2,300 real images and 2,010 fake images created using FaceSwap and the SwapMe iOS app (now discontinued). SwapMe and FaceSwap dataset 78 21 [ Šī datu kopa satur 131 500 sejas attēlus, tostarp 78 500 attēlus, kas iegūti no 150 videoklipiem FaceForensics datu kopā, un 53 000 attēlus, kas iegūti no 150 viltus videoklipiem, kas savākti no YouTube. Fake Faces in the Wild (FFW) dataset 32 : This is a number of commercial datasets provided by the Generated Media, Inc., with up to nearly 2.7 million synthetic face images generated by StyleGAN. A free edition with 10,000 128x128 synthetic images is made available for academic research. The website also provides an interactive face generator un ugunsgrēks . The generated.photos datasets have a good diversity: five age groups (infants, children, youth, adults, middle-aged), two genders (male and female), four ethnicities (white, black, Latino, Asian), four eye colours (brown, grey, blue, green), four hair colours (brown, black, blond, gray), three hair length (short, medium, long), facial expressions, three head poses (front facing, left facing, right facing), two emotions (joy and neutral), two face styles (natural, beautified). (According to a number of research papers we read, an earlier 100K-Faces dataset was released by generated.photos for academic research in 2018, which was used by many researchers. This dataset is not currently available any longer.) generated.photos datasets 22 23 24 [ ]: Šī datu kopa ietver 19 457 sejas attēlus, tostarp 7 948 dziļi viltotus attēlus, kas ģenerēti no 175 viltotiem videoklipiem, kas savākti tiešsaistē, un 11 509 reālus sejas attēlus, kas savākti no dažādiem tiešsaistes avotiem. (papīra 2. tabula parāda datu kopas lielumu 19 509, bet datu kopā, kas lejupielādēts no pCloud, ir tikai 19 457 attēli.) MesoNet Deepfake Dataset 1 [ ]: This dataset includes 100,000 synthesised face, bedroom, car and cat images by a GAN generator trained based on real images in the FFHQ and LSUN datasets (three object types – bedrooms, cars and cats – for the latter). Note that the name “100K-Generated-Images” was not a proper one as the authors [ ] vienkārši to izmantoja, lai nosauktu savu Google Drive koplietošanas telpu apakšlodziņu, bet tas tika izmantots vienā no aptaujas rakstiem [ ]. 100K-Generated-Images 30 25 26 30 65 [ Šī datu kopa satur 420,053 slavenību attēlus, tostarp 156,930 reālus attēlus, kas lejupielādēti, izmantojot Google Image API, un 263,123 viltus sejas apmaiņas attēlus, kas izveidoti, izmantojot divas dažādas metodes (Nirkin metode un Auto-Encoder-GAN). Ding et al.’s swapped face dataset 17 [Lasīt ]: This dataset includes 87,000 224x224 face images, generated by processing some StyleGAN-generated synthetic images using the GAN-fingerprint Removal approach (GANprintR) proposed by Tā ir aizstājēja versija dataset, which contains 150,000 face images generated using an earlier version of GANprintR. iFakeFaceDB 48 Neves et al. FSRemovalDB [ ]: This dataset includes 40,000 images, half real and half deepfake. The images were collected from four sources: the CelebA-HQ dataset , the Flickr-Faces-HQ dataset , the 100K-Faces dataset (not available any longer, see the description of generated.photos datasets), and . Faces-HQ 21 27 28 29 thisperson-doesnotexist.com [ ]: Šī datu kopa ietver 625 537 sintezētus sejas attēlus no 10 177 slavenībām, ar 43 bagātajiem atribūtiem sejas, apgaismojuma, vides un spoof veidiem. 43 atribūti ietver 40 reāliem attēliem, kas aptver visas sejas sastāvdaļas un piederumus (piemēram, ādu, degunu, acis, skropstas, lūpas, matus, cepuri, brilles), un 3 viltus attēliem, kas aptver sūnu veidus, vidi un apgaismojuma apstākļus. CelebA-Spoof 75 30 [Lasīt ]: This dataset contains 299,039 images, including 58,703 real images sampled from three datasets (FFHQ , CelebA and FaceForensics++ ) and 240,336 fake ones in four main facial manipulation types (identity swap, expression swap, attribute manipulation, and entire synthesis). The images cover two genders (male and female), a wide age groups (the majority between 21 and 50 years old), and both low- and high-quality levels. Diverse Fake Face Dataset (DFFD) 11 31 32 33 4.2 Deepfake Video Datasets [Lasīt ]: This dataset contains 620 deepfake face videos, generated by face swapping without manipulation of audio, covering 32 subjects and two quality levels (high and low). DeepfakeTIMIT 35 (FF) [ ]: This dataset contains 1,004 face videos with over 500,000 frames, covering various quality levels and two types of facial manipulation. This dataset is now replaced by the larger FaceForensics++ dataset (see below). FaceForensics 55 (FF++) [ ]: This dataset contains 5,000 face videos with over 1.8 million manipulated frames, including 1,000 real videos (with 509,914 frames) downloaded from YouTube, and 4,000 fake videos created using four face manipulation methods (Deepfakes, Face2Face, FaceSwap and NeuralTextures). The videos cover two genders (male and female), and three quality levels (VGA/480p, HD/720p, and FHD/1080p). FaceForensics++ 56 [ ]: This dataset contains 98 face videos, half (49) are real ones downloaded from Youtube, and the other half are fake ones generated using the FakeApp mobile application (which is now discontinued). The video dataset was created to used to demonstrate a deepfake video detection method based on detection of eye blinking behaviours, so all videos contain at least one eye-blinking event. All fake videos were created by swapping the original face in each of the real videos with the face of the actor Nicolas Cage , thus, only one subject is represented. UADFV dataset 39 34 [ ]: This dataset contains 142 “in the wild” deepfake portrait videos, collected from a range of online sources including news articles, online forums, mobile apps, and research presentations. The videos are diverse, covering the source generative model, resolution, compression, illumination, aspect-ratio, frame rate, motion, pose, cosmetics, occlusion, content, and context. Deep Fakes Dataset 10 [ ]: This dataset contains 5,244 face videos of 66 subjects with both face and voice manipulation. It was released as a preview of the full dataset of the 2020 Deepfake Detection Challenge (DFDC, see below). DFDC (Deepfake Detection Challenge) preview dataset 18 : This dataset contains 1,203 face videos of celebrities, including 408 real videos collected from YouTube with subjects of different ages, ethic groups and genders, and 795 deepfake videos synthesised from these real videos. Celeb-DF v1 35 [ ]: This dataset contains 6,229 face videos of celebrities, including 590 real videos collected from YouTube with subjects of different ages, ethic groups and genders, and 5,639 deepfake videos synthesised from these real videos. Celeb-DF v2 40 [ ]: This dataset contains 3,363 face videos, covering 28 subjects, gender, and skin colour. It was created as a joint effort between two units of Google, Inc.: Google AI and JigSaw . DeepFake Detection (DFD) Dataset 20 36 37 [ ]: This dataset contains 60,000 indoor face videos (with 17.6 million frames) generated by face swapping, covering 100 subjects, four skin tones (white, black, yellow, brown), two gen-ders (male and female), different age groups (20-45), 26 nationalities, 7 different angles, 8 face expressions, and different head poses. DeeperForensics-1.0 27 [ ]: This dataset contains 128,154 face videos of 960 subjects, including 23,654 real videos from 3,426 paid actors and 104,500 deepfake videos created using eight different methods (DF-128, DF-256, MM/NN face swap, NTH, FSGAN, StyleGAN, refinement, and audio swap). DFDC (Deepfake Detection Challenge) full dataset 18 10 [ Šī datu kopa satur 10 000 augstas kvalitātes viltošanas video ar video un sejas līmeņa piezīmēm.Datu kopa koncentrējas uz sarežģītāku viltošanas atklāšanas gadījumu: katrā videoklipā ir iesaistīti viens līdz 15 indivīdi, bet tikai daži (mazākā daļa) sejas tiek manipulētas. FFIW K (Face Forensics in the Wild) dataset 79 FFIW [ ]: Šajā datu kopā ir 37,942 video no maksas subjektiem (395 korejieši un 8 Dienvidaustrumu Āzijas iedzīvotāji), tostarp 62,166 reāli video un 175,776 viltotas, kas izveidotas, izmantojot sešas metodes – FaceSwap, DeepFaceLab, FSGAN, First Order Motion Model (FOMM), Audio-driven Talking Face HeadPose (ATFHP) un Wav2Lip. Korean DeepFake Detection Dataset (KoDF) 36 [ ]: This dataset contains 1,737 videos with 1,666,816 frames, including 1,339,843 real frames and 326,973 fake frames generated using the Deep Video Portraits (DVP) [ ] method. The original videos were obtained from three sources: the dataset used in [ ], the Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) [ ], and YouTube. Most videos have a resolution of 1280×720. VideoForensicsHQ 23 34 33 42 [ ]: This dataset contains 7,314 face sequences extracted from 707 deepfake videos that were collected completely from the Internet. It covers diverse scenes, multiple persons in each scene and rich facial expressions. Different from other deepfake video datasets, WildDeepfake contains only face sequences not the full videos. This makes the dataset more like between an image dataset and a video one. We decided to keep it in the video category since the selection process was still more video-focused. WildDeepfake 81 4.3 Deepfake Audio/Speech Datasets Voice conversion (VC) is a technology that can be used to modify an audio and speech sample so that it appears as if spoken by a different (target) person than the original (source) speaker. Obviously, it can be used to generate deepfake audio/speech samples. The biennial Voice Conversion Challenge Tas, kas sākās 2016. gadā, ir liela izaicinājuma sērija par VC. No šīs izaicinājuma sērijas izlaisto datu kopas ļoti atšķiras no citām deepfake datu kopām: deepfake dati nav iekļauti katras izaicinājuma organizētāju radītajā sākotnējā datu kopā, bet gan dalībnieku iesniegumos (kas ir retargeted/fake utterances, ko ražo dalībnieku izveidotas VC sistēmas). izaicinājuma datu kopas ietver arī visu iesniegumu novērtējumu (uzklausīšanu balstītus) rezultātus. 38 [ ]: The original dataset created by the challenge organisers was derived from the DAPS (Device and Produced Speech) Dataset [ ]. It contains 216 utterances (162 for training and 54 for testing) per speaker from 10 speakers. Participating teams (17) developed their own VC systems for all 25 source-target speaker pairs, and then submitted generated utterances for evaluation. At least six participating teams used DL-related techniques (LSTM, DNN) in their VC systems (see Table 2 of the result analysis paper ), so the submitted utterances can certainly be considered deepfakes. Voice Conversion Challenge 2016 database 62 47 39 [ ]: The original dataset created by the challenge organisers was also based on the DAPS dataset. It contains 116 utterances (81 for training and 35 for testing) per speaker from 12 speakers in two different tasks (called Hub and Spoke). Participating teams (23 in total, all for Hub and 11 for Spoke) developed their own VC systems for all 16 source-target speaker pairs, and then submitted generated utterances for evaluation. Comparing with the 2016 challenge, more participating teams used DL-related techniques (e.g., WaveNet, LSTM, DNN, CycleGAN, DRM – deep relational models, and ARBM – adaptive restricted Boltzmann machines) in their VC systems. Voice Conversion Challenge 2018 database 44 [ ]: This dataset is based on the Effective Multilingual Interaction in Mobile Environments (EMIME) dataset , a bilingual (Finnish/English, German/English, and Mandarin/English) database. It contains 145 utterances (120 for training and 25 for testing) per speaker from 14 speakers for two different tasks (with 4 × 4 and 4 × 6 source-target speaker pairs, respectively). Participating teams (33 in total, out of which 31 for Task 1 and 28 for Task 2) developed their own VC systems for all source-target speaker pairs, and then submitted generated utterances for evaluation. Comparing with the 2018 challenge, DL-based VC systems were overwhelmingly used by almost all participating teams (WaveNet and WaveGAN among the most used DL-based building blocks). Voice Conversion Challenge 2020 database 70 40 A major set of deepfake speech datasets were created for the (Automatic Speaker Verification Spoofing and Countermeasures) Challenge (2015-2021, held biannually). The datasets for the 2019 and 2021 contain speech data that can be considered deepfakes. ASVspoof 41 [Lasīt ]: This dataset is based on the Voice Cloning Toolkit (VCTK) corpus , a multi-speaker English speech database captured from 107 speakers (46 males and 61 females). Two attack scenarios were considered: logical access (LA) involving spoofed (synthetic or converted) speech, and physical access (PA) involving replay attacks of previously recorded bona fide recordings). For our purpose in this survey, the LA scenario is more relevant. The LA part of the dataset includes 12,483 bona fide (real) utterances and 108,978 spoofed utterances. Some of the spoofed speech data for the LA scenario were produced using a generative model involving DL-based techniques such as long short-term memory (LSTM) , WaveNet [ ], WaveRNN [ ], WaveCycleGAN2 [ ]. Note that the challenge organisers did not use the term “deepfake” explicitly, despite the fact that the DL-generated spoofed speech data can be considered as deepfakes. ASVspoof 2019 Challenge database 67 42 43 50 28 58 [ ]: Šī datu kopa satur godprātīgus un viltus runas datus loģiskās piekļuves (LA) uzdevumam. izaicinājums joprojām turpinās, un mēs neesam atraduši detalizētu dokumentu par datu kopu, tāpēc nevar iekļaut citas detaļas, izņemot tā lielumu (7,8 GB pēc saspiešanas). Lai gan mēs neesam redzējuši detaļas par ģeneratīvo algoritmu, ko izmanto, lai ražotu viltus runas datus, mēs uzskatām, ka līdzīgi DL balstīti algoritmi tika izmantoti 2019. gada izaicinājumam. ASVspoof 2021 Challenge – Logical Access Database 14 [ ]: In 2021, the challenge included an explicitly defined track on deepfake, but the task description suggests that the organisers of the challenge considered a broader definition of the term “deepfake” by looking at spoofing human listeners rather than ASV (Automatic Speaker Verification) systems. The size of the dataset is 34.5 GB after compression. ASVspoof 2021 Challenge – Speech Deepfake Database 15 Possibly because of the long history and wide participation of the community in the ASVspoof challenges for creating the dedicated datasets, there are very few other deepfake audio/speech datasets. One such dataset was created by a group of researcher from Baidu Research [ ]. This dataset was created to demonstrate a proposed voice cloning method. It is relatively small, and contains 134 utterances, including 10 real ones, 120 cloned ones, and 4 manipulated ones. Another dataset was created by Google AI and Google News Initiative , but it was made part of the ASVspoof 2019 dataset. This dataset contains thousands of phrases spoken by 68 synthetic “voices” covering a variety of regional accents. 5 44 4.4 Hybrid Deepfake Datasets : These datasets were created by the DARPA Media Forensics (MediFor) Program for the 2020 OpenMFC . There are two GAN-generated deepfake datasets, one with more than 1,000 deepfake images and the other with over 100 deepfake videos. The datasets were made available to registered participants of the competition only. NIST OpenMFC (Open Media Forensics Challenge) Datasets 45 46 47 [ ]: This dataset is named as “a versatile benchmark for comprehensive forgery analysis”. It contains 2,896,062 images and 221,247 videos, including 1,457,861 fake images and 121,617 fake videos. The videos and images cover seven image-level and eight video-level manipulation approaches, 36 different types of perturbations and more mixed perturbations, and a large number of annotation labels (6.3 million classification labels, 2.9 million manipulated area annotations and 221,247 temporal forgery segment labels). The dataset is being used for supporting the Face Forgery Analysis Challenge 2021 SenseHuman 2021 (3. seminārs par cilvēku uztveri, izpratni un sintēzi) , co-located at the ICCV 2021 conference . ForgeryNet 25 48 49 50 4.5 A Deepfake Dataset Generator [ ]: This is not actually a dataset per se, but a system for producing large datasets more automatically, including generating deepfake datasets. One may argue the automatically generated datasets are fake since they are not produced from real-world scenes. DatasetGAN 74 4.6 Subjective Quality of Deepfakes in Different Databases As mentioned in Section , subjective quality evaluation is necessary to evaluate the realness, realisticness, and naturalness of deepfake media. While there has been very limited work on this topic, in 2020, Jiang et al. [ ] conducted a user study on realness of deepfake videos. They recruited 100 professional participants (most of whom are computer vision researchers), who were asked to evaluate the realness of 30 randomly selected videos from 7 deepfake video datasets (DeeperForensics-1.0, UADFV, DeepFake-TIMIT, Celeb-DF, FaceForensics++, Deep Fake Detection, and DFDC). Participants were asked to respond to the statement “The video clip looks real.” and gave scores following a five-point Likert scale (1 – clearly disagree, 2 – weakly disagree, 3 – borderline, 4 – weakly agree, 5 – clearly agree). 4.7 27 Table shows the results. Interestingly, we can see a huge difference between the realness levels of different datasets. What is probably quite surprising is that FaceForensics++, one of the most widely used deepfake datasets, has a very low MOS score and less than 9% of participants considered the 30 selected videos as real. 3 Table 3: Human-judged subjective quality (realness) of deepfake videos in 7 datasets. The MOS scores were not reported by , but calculated by us based on the raw data shown in Table 3 of [ ]. Jiang et al. 27 4.7 Discussion: Datasets Among all deepfake image and video datasets, a significant majority are about face images and videos. This is not surprising since face swapping, face attribution manipulation, and fully synthesised face images are among the hottest topics within deepfake research and real-world applications. We hope more non-face deepfake image and video datasets can be produced to support a broader range of research activities on deepfake. The subjective quality results shown in Table indicate that it is important to check realness of deep-fake media to support any performance evaluation or comparison. To ensure that the quality evaluation of datasets is fair, transparent and reliable, standard procedures need defining and a common pool of qualified human experts should be used. 3 Many authors of deepfake-related datasets attempted to classify such datasets into different generations. Chronologically speaking, we could broadly split such datasets into two generations: before 2019 and since 2019. Typically, datasets created before 2019 are relatively less advanced and smaller, while those created after 2019 tend to be larger, more diverse (i.e., covering more attributes), and of higher quality (i.e., produced by more advanced generative models). This can also be seen from the data in Table , in which the top two datasets (DeeperForensics-1 and Celeb-DF) fall within the new generation (2020), while others belong to the old generation. In addition to the two generations, a newer generation has also emerged in 2021: a number of very recent datasets started focusing on more realistic deepfakes (i.e., in the wild) or more specified areas of deepfakes (e.g., 10 focusing on multiple faces in the same video, and KoDF focusing on Korean faces). This trend shows that the deepfake research community has grown significantly in the past few years so that narrower topics have also started gaining attention and interest from some researchers. 3 FFIW K 5 Deepfake-Related Challenges, Competitions & Benchmarks Šajā sadaļā tiek pārskatītas iniciatīvas, kuru mērķis ir veicināt sintētisko vai manipulēto plašsaziņas līdzekļu (piemēram, video, attēlu un audio) atklāšanu un ģenerēšanu, izmantojot sabiedrībai atvērtas sacensības vai izaicinājumus, kā arī pastāvīgus kritērijus konkrētu problēmu risināšanai. 5.1 Detection of Manipulated Media Deepfake detekcijas izaicinājums (DFDC) was an initiative promoted by an AI and Media Steering Committee , including BBC, Facebook, Amazon, Microsoft and New York Times, and some universities around the world including the University of Oxford. The competition remained open from 5 September 2019 till 31 March 2020, and involved 3 stages. At first, the DFDC preview dataset was released. At a later stage, the DFDC full dataset was also made available to the 2,114 participants of the competition incorporating face and audio swap techniques for generation of deepfake content. At the final stage, the submitted models were evaluated using a test dataset (referred to as the “black box dataset”) of 10,000 videos which included deepfake videos. The best performance on the black box dataset had an accuracy of 65.18%, according to the released results [ ]. Submissions were ranked according to the overall log loss score, as defined in Eq. ( ). All top five ranked models (the winner had the lowest overall log loss) are available on GitHub. Results indicate how challenging the detection of deepfake is since the best accuracy was low and “ ”, according to Dolhansky et al. [ ]. Figure shows a screenshot of the leaderboard with the five finalists. The first top ranked model used MTCNN (Multi-tasked Cascaded Convolutional Network), the second used WS-DAN (Weakly Supervised Data Augumentation Network), and the third used the EfficientNetB7 architecture. Meta compiling the common themes observed in the winning models, they were: clever augmentations, architectures, and absence of forensics methods. Moving forward, they called for “ ”. 51 52 in-the-wild 22 53 13 many submissions were simply random 19 2 Risinājumi, kas pārsniedz attēlu un video analīzi, ņemot vērā kontekstu, izcelsmi un citus signālus, var būt veids, kā uzlabot dziļo viltojumu noteikšanas modeļus The Automatic Speaker Verification Spoofing And Countermeasures Challenge Workshop (ASVspoof) has been running biennially since 2015. This competition is organised by an international consortium that includes Inria and EURECOM (France), University of Eastern Finland, National Institute of Informatics (Japan), and Institute for Infocomm Research (Singapore). This year the ASVspoof challenge includes, for the first time, a sub-challenge focused on where the envisioned use case is an adversary trying to fool a human listener. The metric used for evaluating performance of submitted solutions (i.e., classifiers) is EER. Four baseline solutions (also called “countermeasures”), each using a different technique, were made available to participants with their corresponding EER metric values. The ASVspoof 2021 Speech Deepfake Database containing audio recordings with original and spoofed utterances has also been made available. The competition involves three phases : progresa fāze, novērtēšanas fāze un pēcvērtēšanas fāze; nav skaidrs, kā komandas pāriet no vienas fāzes uz nākamo. ]. The organisers of the competition noted that they opted for the EER as the performance evaluation met-ric for countermeasures submitted to the speech deepfake task for legacy reasons. They acknowledged, however, that “ ” by the ISO/IEC 19795-1:2021 Neskatoties uz to, ka tikai 2021. gada ASVspoof konkursā bija saturs, kas skaidri saistīts ar deepfake, daži dati ASVspoof 2019 datu kopā (Logical Access uzdevums), kas izmantots 2019. gada konkursam, tika ģenerēti, izmantojot DL algoritmus, kā minēts sadaļā . We expect that this also holds for the ASVspoof 2021 dataset (Logical Access task). The ASVspoof 2019 competition used the EER as secondary metric; the primary performance metric used was the tandem detection cost function (t-DCF) [ ]. According to its evaluation plan [ ], t-DCF assesses the performance of the whole tandem system whereby “ ”. It is calculated according to Eq. ( ), where cm ( ) and cm( ) are, respectively, “ ”. 54 Vārds DeepFake 55 56 13 EER ziņošana ir deprecēta 57 4 63 69 a CM [countermeasure] serves as a ‘gate’ to determine whether a given speech input originates from a bona fide (genuine) user, before passing it the main biometric verifier (the ASV system) 17 P s P s the miss rate and the false alarm rate of the CM system at threshold s For further information about Eq. ( ), including constants 1 un 2, lūdzu, skatiet ASVspoof 2019. gada novērtējuma plānu [ ]. 17 C C 69 An implementation of the t-DCF metric has been made available by the ASVspoof 2019’s organisers in Python and Matlab formats. 58 59 The Face Anti-spoofing (Presentation Attack Detection) Challenge started in 2019. Its first two editions were held at the 2019 and 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2020), respectively. Its third edition was moved to be co-located with the 2021 IEEE/CVF International Conference on Computer Vision (ICCV 2021). This competition series was organised by a group of researchers from academia and industry in China, Mexico, Spain, Finland and the US. The 2021 competition was focused on 3D high-fidelity mask attacks, and followed a 2-phased process. The first phase is the “development phase”; it started in April 2021 when the CASIA-SURF HiFiMask dataset was released to participants. The second phase is the “final ranking phase” (June 2021), when the competition ended. The competition adopted the following performance metrics for evaluation of the solutions submitted: attack presentation classification error rate (APCER), normal/bona fide presentation classification error rate (NPCER/BPCER), and average classification error rate (ACER), in accordance with the ISO/IEC 30107-3:2017 Standarta figūra provides the leaderboard for the top three solutions. 60 61 62 63 64 3 The FaceForensics Benchmark is an ongoing automated benchmark for detection of face manipulation. The organisers of the benchmark made the FaceForensics++ dataset available for training. Manipulated videos (4,000 in total) were created using four techniques, i.e., two computer graphics-based approaches (Face2Face and FaceSwap) and two learning-based approaches (DeepFakes and Neural Textures). The deepfakes videos were generated using a slightly modified version of FaceSwap , and the Neural Textures videos were created using the approach proposed by Thies et al. [ ]. The benchmark test dataset is created from the collection of 1,000 images randomly selected from either the manipulation methods or the original videos [ ]. Participants have to submit results to the benchmark, rather then code like other competitions; this is illustrated in Figure . The outcome of a submission is illustrated in Figure , where the scores are a measure of accuracy (Eq. ( )). 65 66 61 56 4a 4b 8 The Open Media Forensics Challenge (OpenMFC, formerly DARPA MFC) is an annual image and video forensics evaluation aiming to facilitate development of multimedia manipulation detection systems. It has been organised annually starting from 2017 under the name of DARPA MFC. In 2020, the National Institute of Standards and Technology (NIST) initiated the as a new evaluation platform, based on their previous experiences with the DARPA MFC series, to make the participation more convenient for all researchers. In OpenMFC 2020, two deepfake-related tasks were included for the first time: Image GAN Manipulation Detection (IGMD) and Video GAN Manipulation Detection (VGMD). The organisers provided an image evaluation dataset for the IGMD task, containing 1,000 images from over 200 image journals , and a video evaluation dataset for the VGMD task, including over 100 test videos. Furthermore, they provided the datasets used in the previous MFC challenges as development datasets. The challenge is composed of two main phases for development and evaluation, respectively, and a pre-challenge phase for quality control testing. For evaluation of submissions, AUC-ROC is used as the primary metric. Furthermore, CDR@FAR, where CDR refers to correct detection rate or TPR (Eq. ( )) and FAR refers to false alarm rate or FPR (Eq. ( )), is also used as a metric [ ]. The DeeperForensics Challenge 2020 is a deepfake face detection challenge held at the 2020 ECCV 67 68 OpenMFC 69 70 4 5 49 71 SenseHuman Workshop . The challenge used the DeeperForensics1.0 dataset. 72 The organisers provided a hidden test dataset to better simulate real-world scenarios. The challenge involved two phases: the “development phase” that started in August 2020 allowing 100 successful sub-missions, and the “final test phase” that started in October 2020 allowing 2 successful submissions until the end of the month. The submissions were evaluated using the binary cross-entropy loss (BCELoss) metric, calculated according to Eq. ( ), where is the number of videos in the hidden test set, is the ground truth label of video (fake:1, real:0), and ( ) is the predicted probability that video is fake. 18 N yi i p yi i Results of the competition were discussed by Jiang et al. [ ]. The top solution used three models, i.e., EfficientNet-B0, EfficientNet-B1 and EfficientNet-B2, for classification. The second top used EfficientNet-B5 for both an image-based model and a video-based model. The third ranked solution used a 3D convolutional neural network (3DCNN). 73 26 The Face Forgery Analysis Challenge 2021 is a competition hosted at the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2021). It is organised by researchers from a number of organisations in China including universities and SenseTime Research (the research arm of SenseTime , one of the major AI “unicorns” in China). The challenge aims to advance the state-of-the-art in detection of photo-realistic manipulation of images and videos. Participants are able to use a large annotated face dataset (i.e., the ForgeryNet dataset) that was obtained by applying a number of techniques for manipulation (15) and perturbation (36) to train their solutions. The phases comprise of Forgery Image Analysis, Forgery Video Analysis, Forgery Video Temporal Localization phases, and the final phase (i.e., “private test”) where participants’ models will be tested against an unseen dataset. The following metrics will be used [ ]: AUC, average precision (AP) at some “temporal Intersection over Union” (AP@tIoU) compared to a threshold ∈ [0*. ,* 0*.*95], and average recall (AR) at (AR@ ) where is the top marķējumi, kas atgriezušies vairāku klasifikatoru klasifikatoriem. 74 75 25 t 5 K K K K The 2020 CelebA-Spoof Face Anti-Spoofing Challenge was hosted at the 16 European Conference on Computer Vision (ECCV 2020). The challenge ran between August and October 2020, and aimed to advance the state-of-the-art in detecting “ ” [ ]. The organisers made the face CelebA-Spoof dataset available for the competition containing rich annotation across a range of attributes. The competition only had one phase where participants submitted their solutions to be evaluated against a test dataset; the spoof class was considered as “positive” and the live class as “negative”. Metric TPR@FPR was used and collected at three points where the TPR when FPR = 10 4 determined the final ranking. The top three finalists (see Figure ) used deep learning models ResNet, EfficientNet-B7, and a novel architecture combining Central Difference Convolutional Networks (CDCN) and Dual Attention Network (DAN). The two top ranked solutions used different strategies to boost their models’ performance: a heuristic voting scheme was used by the top-ranked solution, and a weight-after-sorting strategy was used by the second ranked solution. 76 th whether a presented face is live or spoof 76 − 5 The 2021 CSIG Challenge is the second edition of a challenge organised by the China Society of Image and Graphics 2021. gada izaicinājums ir Fake Media Forensic Challenge as its 6 track, co-organised by CSIG’s Digital Media Forensics and Security Technical Committee and Institute of Information Engineering, Chinese Academy of Sciences . This track has two tasks, one on deepfake video detection, and the other on deepfake audio/speech detection. For the deepfake video detection task, the dataset used contains a public training set with 10,000 sound-free face videos (including 4,000 fake videos), a public test set with 20,000 face videos (the percentage of deepfake videos is unknown to participants), and a private test set that will be determined and used at the final session for selecting the winners. All videos contain faces of Eastern Asian people, and cover a wide range of parameters such as multiple resolutions and encoding quality factors, the use of blurring or sharpening filters, and added noise. Deepfake videos were created using public tools including DeepFaceLab [ ], Faceswap , Faceswap-GAN, Recycle-GAN [ ] and ALAE (Adversarial Latent Autoencoders) [ ]. For the deepfake audio/speech detection task, the dataset used contains a public training set with 10,000 speech samples (including 6,000 fake ones), a public test set with 20,000 face videos (the percentage of deepfake videos is unknown to participants), and a private test set for the final session (the same as the deepfake video detection task). The tools used for generating the fake speech samples include TTS (text-to-speech) voice synthesis tools and VC (voice conversion) tools. The main TTS tools used include open-source tools such as DeepVoice, TensorFlowTTS and GAN-TTS [ ] and commercial software tools such as those from iFlytek and IBM. The main VC tools used include Adaptive-VC and CycleGAN-VC [ ]. For both deepfake detection tasks, the performance metric used is log loss. 77 78 79 th 80 81 53 82 6 54 83 8 84 29 2020 China Artificial Intelligence was the second edition of a Chinese AI competition open for the general public to participate, organised by the municipal government of the City of Xiamen in China. In 2020, it had two sub-competitions, Multimedia Information Recognition Technology Competition and Language and Knowledge Technology Competition . The Multimedia Information Recognition Technology Competition included two tasks on deepfakes: one on deepfake video detection and one on deepfake audio/speech detection . The deepfake video detection task used 3,000 videos, and log loss was used as the sole performance metric. The deepfake audio/speech detection task used 20,000 audio samples (mostly in Chinese, and the remaining in English), and EER was used as the sole performance metric. For both tasks, the ratio between real and deepfake samples was 1:1. We did not find where to download the datasets used for the tasks nor a more detailed technical description of the datasets. For the deepfake video detection tasks, the top two winning teams (with an A prize) were from Netease (Hangzhou) Network Co., Ltd. and Beijing RealAI Technology Co., Ltd., followed by three other teams winning a B prize: Xiamen Fuyun Information Technology Co., Ltd.; Institute of Computing Technology, Chinese Academy of Sciences; and Wuhan Daqian Information Technology Co., Ltd. For the deepfake audio/speech task, there was no team winning an A prize, but one team winning a B prize: SpeakIn Technologies Co., Ltd. The final results of some teams were published, but some teams were allowed to hide their results. We did not find a detailed technical report summarising the results and explaining the work of the winning teams. 85 86 87 88 89 One of the B-prize winning team is from Beijing RealAI Technology Co., Ltd., a Chinese company active in deepfake-related R&D. 5.2 Generation of Manipulated Media The Voice Conversion Challenge ir divu gadu konkurss, kas norisinās kopš 2016. gada. izaicinājums un attiecīgais seminārs, kas notika INTERSPEECH konferencē , is supported by the SynSig (Speech Synthesis Special Interest Group) of the International Speech Communication Association (ISCA) . Its aim is to promote progress in voice conversion (VC) technology that can be applied to a number of positive and negative use cases, such as spoofing voice biometric systems. The 2020 challenge focused on speaker conversion, a sub-problem of VC, and included two tasks. For the first task “intra-lingual semi-parallel voice conversion”, participants had to develop 16 VC systems (speaker-pair combinations) including male and female speakers and English sentences, using the provided Voice Conversion Challenge 2020 database v1.0 for training (refer to Section Otrajā uzdevumā “starpvalodu balss konvertēšana” dalībniekiem bija jāizstrādā 24 VC sistēmas, tostarp arī vīriešu un sieviešu runātāji, bet teikumu izrunāšana trīs valodās (finiešu, vācu un mandarīnu), pamatojoties uz sniegto apmācības datu kopumu. illustrates the process of training and generation of VC systems. 90 91 92 93 4 6 Submissions were evaluated for “ ” . As such, the organisers used [ ] and recruited both native and non-native English speakers (i.e., Japanese native speakers) via crowd-sourcing for the listening tests. Naturalness (answering the question “How natural does the converted voice sound? ”) was measured using the metric MOS (covered in Section 4.6), and similarity (answering the question “how similar the converted voice sound comparing source and target speakers? ”) was measured in terms of speaker recognition as “same” or “different”, as elaborated by Wester et al. [68]. Tests also focused on the effects of language differences on the performance of VC systems submitted to the competition. The most popular CNN/RNN/GANbased VC systems submitted used WaveNet, WaveRNN, and Parallel WaveGAN. Results indicated that, in terms of similarity, the best performing VC systems were as good as natural speech but none reached human-level naturalness for task 1; scores were lower for task 2 which was more complex [70]. The organisers of the 2020 competition also used objective evaluation [12]. The metrics used for evaluation of speaker similarity were: equal error rate (EER), false acceptance rate of target (P tar fa ), miss rate of source (P src miss), and cosine similarity of speaker embedding vectors (cos-sim) according to Eq. (19) where A is the speaker embedding vectors for the converter audio and B is the speaker embedding vectors for the original audio. The performance of the VC systems as a spoof countermeasure was also evaluated using EER, while to evaluate the quality of the subjective MOS obtained via listening tests, a DL-based model to predict MOS, called MOSNet [43], was used. Lastly, to evaluate intelligibility of the converted transcribed speech, in comparison with the original transcribed speech, the word error rate (WER) [4] was used. WER is calculated according to Eq. (20) where I refers to insertions, D refers to deletions, S refers to substitutions, and N refers to the total number of words in the original transcript. perceived naturalness and similarity through listening tests 94 subjective evaluation 70 The Deepfake Africa Challenge (2021) is a new initiative of the AI Africa Expo, in partnership with a film and media production company (Wesgro) and the African Data Science competition platform Zindi. Its aim is “ ”. Eligible participants were required to be citizens and residents of the African continent. Submissions, accepted up to end of July 2021, can be either video or audio. Evaluation of submissions is defined in terms of artistic creativity, relevance of challenge topic, and innovation in the process of generation as long as participants use tools and packages publicly available. The top three finalists will receive a prize, present their work at the Expo, and will have to grant copyrights to Zindi. Unlike the other competitions reviewed in this section, which were focused on advancing the state-of-the-art in detection of synthetic or manipulated media, this competition focused on the generation of deepfake which seems more humanities-centred. This is a trend observed in arts [ Kultūra un kultūra [ ].5.3 Manipulēto plašsaziņas līdzekļu radīšana un noteikšana 95 to create convincing deepfakes to highlight the power of this synthetic media, illustrating its creative potential for exploitation for both positive and negative outcomes and focusing debate about its ethical use / misuse in an African context 31 57 The DeepFake Game Competition (DFGC) is in its first edition, hosted at the 2021 International Joint Conference on Biometrics (IJCB 2021). Its organisers are mainly from the Institute of Automation Chinese Academy of Sciences (CASIA). The idea of the competition was to promote an adversarial game between agents pushing for advances in both deepfake creation and detection. In order to achieve this, a 6-stage protocol was designed interleaving three creation phase (C-phase) and detection phase (D-phase), typically one week apart; submissions closed in April 2021. Both C-phases and D-phases were bound to the Celeb-DF (v2) dataset [ ], containing 6,229 videos (590 real/original videos and 5,639 fake/manipulated videos), for training purposes. As such, submissions to a C-phase would consist of datasets extracted from Celeb-DF (v2) which included novel face-swap approaches to obtain evaluation results. Submissions to a D-phase would consist of detection models/codes to obtain evaluation results. The models submitted for a D-phase were evaluated against the datasets submitted for the previous C-phase [ ]. The metrics used for evaluation were: a detection score, used for evaluation of a D-phase, and a creation score, used for evaluation of a C-phase. The top three finalists for the detection phase employed CNN-based classifiers EfficientNet-B3, Efficientnet-B0 and EfficientNetV2. 96 40 52 97 The Detection Score ( ) metric captures the models’ ability to correctly classify fake images submitted to the previous C-phase against a set of real images in the CelebDF test dataset. It is calculated using Eq. ( ), where is the number of valid submissions of created synthesis test sets in the last C-phase. DS 21 NC The Creation Score ( ) metric used to evaluate creation models submitted to this challenge is calculated by Eq. ( 2) kur ir spēkā esošo detekcijas metožu iesniegumu skaits pēdējā D-fāzē, trokšņa rezultāts ( noise) penalises noisy images, the other three parts of the equation relate to the following : “ ”. CS 22 ND S 98 ID level similarity to the donor ID, image level similarity to the target frame, and the deception ability against detection models. ID level similarity is scored by a face recognition model using dot product of two ID features (fake face ID and donor ID). The image level similarity is scored by SSIM [Structural Similarity Index] to make sure the face-swapped image is similar to the corresponding target image in content and quality Peng et al. [ ] observed a commonality between the three winning teams for the creation task, i.e., the use of the FaceShifter [ ] framework for face swapping. They highlighted two overall reflections about the competition: (1) the limited diversity of the deepfake datasets submitted and the use of repetitive methods to generate them, and (2) the limited size of the Celeb-DF (v2) dataset itself flagging the need for a larger dataset for next year’s competition. The organisers of the competition also applied the top two detection models to unseen datasets (DFDC and FaceForensics++) and noticed that they do not generalise well. 52 37 6 A Meta-Review of Deepfake-Related Surveys This section presents a meta-review of 12 selected deepfake-related survey papers, including eight published in English [ , , , – , , ] and four published in Chinese [ , , , ]. It covers the following aspects in a systematic manner: definitions and scope, performance metrics, datasets, challenges/competitions/benchmarks, performance comparison, key challenges and recommendations. 16 45 46 64 66 71 73 7 38 41 59 The meta-review aims at drawing some high-level insights for monitoring future development of deepfake-related technologies and their applications. 6.1 Definitions and Scope As we discussed in Section , among researchers, practitioners and law makers there is no universally accepted definition of “deepfake” as a term. This is also reflected in how the authors of the 12 survey papers considered this aspect. Most authors talked about the history of deepfakes and pointed out that the term reflects the combination of “deep learning” and “fake”, but some used a broader definition, e.g., Lyu [ ] defined deepfake as “ ”. Some authors also referred to deepfake-related legislations, but none of them pointed out that the definitions in some such legislations are completely different from the more technical definitions involving the use of deep learning. No authors discussed the blurred boundary between deepfakes and non-deepfakes, although some surveys actually cover both, e.g., Tao et al. [ ] focused on speech forgery and did not explicitly highlight “deepfake”. 1.1 45 high quality fake videos and audios generated by AI algorithms 59 In terms of the scope, while some authors (correctly) considered all types of media that can be produced by deepfake-related techniques [ , , , ], some considered only a narrow scope, e.g., authors of [ , , , ] considered only videos, and only authors of [ , ] have considered images and videos. Another phenomenon we observed is that many authors focused more on face images and videos, and authors of three surveys [ , , ] even limited the definition of “deepfake” to such a narrow scope: 38 41 45 65 7 64 71 73 16 66 16 64 71 Deshmukh and Wankhade [ ] defined it as “ ”; 16 a technology which creates fake images or videos of targeted humans by swapping their faces [by] another character saying or doing things that are not absolutely done by them and humans start believing in such fake as it is not always recognisable with the everyday human eye Younus and Hasan [ ] considered deepfake as a technique allowing “ ”; and 71 any computer user to exchange the face of one person with another digitally in any video Tolosana et al. [ ] defined it as “ ”. 64 a deep learning based technique able to create fake videos by swapping the face of a person by the face of another person Such unnecessarily narrow definitions and scopes can lead to confusion and do not help exchanges between researchers and practitioners working on different types of deepfakes. We call on more researchers to accept a broader definition of “deepfake” so that highly realistic/natural media of any kind generated by a sophisticated automated method (often AI-based) is considered deepfake. Here, we provide two examples of such a broader definition: the image2image (or pixel2pixel) technique [ ] that allows the production of deepfake images and videos of any objects (e.g., the “horse2zebra” deepfake image shown in Figure ), and the the so-called “deepfake geography [ ]”, where AI-based techniques are used to generate realistic-looking satellite images. 80 7 77 Another important fact missed or not sufficiently discussed by authors of all the 12 surveys is that deepfake techniques can be used for positive applications, e.g., creative arts, entertainment and protecting online users’ privacy. We call for more researchers and practitioners to follow the proposal in the 2020 Tencent AI White Paper [ ] to start using the more neutral-sounding term “deep synthesis”. Accordingly,we can use different words for different types of data generated using “deep synthesis” techniques, e.g., “deep art”, “deep animation”, “deep music”, and “deepfake”. While authors of the 12 survey papers did not recognise the positive applications of “deepfake” technologies, some other researchers did, e.g., organisers of the Voice Conversion Challenge 2020 kas teica VC tehnoloģiju (par runas deepfake) “ ”. 60 99 is useful in many applications, such as customizing audio book and avatar voices, dubbing, movie industry, teleconferencing, singing voice modification, voice restoration after surgery, and cloning of voices of historical persons 6.2 Performance Metrics Surprisingly, none of the 12 surveys have covered performance metrics explicitly. Some directly used performance metrics to explain and compare performance of covered deepfake generation and detection methods. The most used performance metrics include accuracy, ERR, and AUC. This may be explained by the page constraints of such survey papers, which did not allow the authors to extend their coverage significantly to cover performance metrics systematically. The subjective quality of deepfakes is an area least covered by the surveys, which seems related to an unbalanced coverage on deepfake generation and deepfake detection in terms of performance evaluation and comparison (the former much less than the latter). 6.3 Datasets Many of the 12 survey papers list a number of deepfake-related datasets, but none of them have coverage as complete as ours shown in Section . For instance, none of the surveys have covered the Voice Conversion Challenge 2016/2018/2020 datasets and the ASVspoof 2019/2021 datasets are covered briefly only in two surveys [ , ]. In addition, more recent deepfake datasets especially those released in 2021 are also not covered by any of the surveys. We believe that our Section is the most comprehensive review of deepfake-related datasets so far. 4 38 59 4 Some survey papers include datasets that are likely deepfakes, e.g., Verdoliva [ ] aptvēra daudzas vispārīgas viltus attēlu datu kopas, kurās manipulētie attēli netika ģenerēti ar dziļo mācīšanos vai pat AI balstītām metodēm, un daži aptaujas (piemēram, [ ]) mentioned ASVspoof 2015 datasets but we did not see the use of deep learning for generating data used in the dataset. 66 38 6.4 Challenges, Competitions and Benchmarks Many surveys cover deepfake-related challenges, competitions and benchmarks. The coverage is, however, mostly limited, and some challenges (e.g., the Voice Conversion Challenge 2016/2018/2020 and the two Chinese challenges we covered in Section ) are not covered by any of the surveys. The level of detail of challenges, competitions and benchmarks is also normally limited, compared with what we chose to include in Section . Similar to the datasets we covered in Section , mēs uzskatām, ka mūsu segums par izaicinājumiem, konkursiem un kritērijiem, kas saistīti ar dziļo viltošanu is also the most comprehensive so far. 5 5 4 5 6.5 Performance Comparison Most surveys have a good coverage of related methods for deepfake generation and detection, but only some explicitly covered performance comparison between different methods [ , , ]. 38 46 64 Among all the survey papers, Li et al. [ ] conducted the most comprehensive study on performance of different deepfake detection methods. In addition to showing the performance metrics of a number of deepfake detection methods in Table 3 of [ ], they also looked at general characteristics and issues of different types of deepfake detection methods, as shown in Table . Furthermore, they also looked at research on robustness of deepfake detection methods against adversarial samples, referring to some work that showed a lack of such robustness. 38 38 4 Due to quality issues of many deepfake-related datasets (discussed in Section ), we need to treat any performance metrics and comparison of different detection methods with caution. Without testing all methods on a sufficiently large, diverse and high-quality deepfake dataset, the performance comparison results can be misleading. This highlights the importance of having more challenges, competitions and benchmarks to encourage performance comparison on standard datasets and using consistent performance metrics. 4.6 6.6 Challenges and Recommendations Dažu aptauju autori identificēja dažus galvenos izaicinājumus un nākotnes pētniecības virzienus dziļo viltojumu kopienai. Not surprisingly, how to develop more robust, scalable, generalisable and explainable deepfake detection methods is one of the most discussed key challenges and also a major future research direction [ , , , , , , , , ]. Considering the arms race between deepfake generation and detection, this research direction will likely remain the hottest topic in deepfake research. 7 16 38 41 45 59 65 66 71 A couple of surveys [ , ] mentioned fusion as a key future research direction, where “fusion” refers to combining different methods (e.g., combining multiple detectors of different types) and data sources (e.g., jointly considering audio-visual analysis) to achieve better performance for deepfake detection. Lyu [ ] ierosināja, ka, lai atklātu deepfake videoklipus, mums ir jāapsver vairāk video līmeņa atklāšana, ko var uzskatīt par visu video kadru atklāšanas rezultātu apvienošanu. 38 66 45 The authors of three surveys, Lyu [ ] , Deshmukh un Wankhade [ Jēkabs un Jēkabs ], apgalvoja, ka ir nepieciešami labākie (augstākas kvalitātes, vairāk atjaunināti un vairāk standarta) dziļo failu datu kopas, lai izstrādātu efektīvākas dziļo failu atklāšanas metodes. Tāpat viņš uzsvēra, ka jāapsver effects in training data and improve the evaluation of datasets. We agree with them on these points. 45 16 71 45 social media laundering Tao et al. [ ] suggested that low-cost deepfake generation/detection should be considered as a future research direction. This is a valid recommendation since lightweight methods will allow less powerful computing devices (e.g., IoT devices) to benefit from such technologies. 59 Two Chinese surveys [ , ] also mentioned the need to have new deepfake-related legislations combating malicious use of deepfakes and the need to train end users such as journalists. This is likely an area where interdisciplinary research can grow. 38 41 There are also other ad-hoc recommendations given by the authors of some surveys. For example, Lyu [ ] argued that deepfake detection should be considered a (more complicated) multi-class, multi-label and local detection problem. Tolosana et al. [ ] discussed specific research directions for different deep-fake generation methods (face synthesis, identity swap, attribute manipulation, and expression swap). Liang et al. [ ] and Li et al. [ ] recommended more active defence mechanisms such as using digital watermarking and blockchain technologies to build trustworthy media frameworks against deepfakes. 45 64 41 38 7 Conclusion The rapid growth in the capability to manipulate media or create synthetic media which look realistic and natural paved the way for deepfakes. At first, this paper adopted a critical approach to look at different definitions of the term “deepfake”. In that regard, we point out the different contradicting definitions and call for the wider community to consider how to define a new term that has a more consistent scope and meaning. For instance, replacing “deepfake” by “deep synthesis” can be more inclusive by embracing positive applications of deepfake techniques, e.g., in entertainment and for simulation purposes. This paper provided a comprehensive overview of multiple aspects of the deepfake ecosystem drawing from the research literature and other online sources published in two languages: English and Chinese. It covers commonly used performance metrics and standards, related datasets, challenges, competitions and benchmarks. It also presents a meta-review of 12 selected deepfake-related survey papers published in 2020 and 2021, covering not only the above mentioned aspects, but also highlighting key challenges and recommendations. References [1] Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. 2018. MesoNet: A Compact Facial Video Forgery Detection Network. In . IEEE, 1–7. Proceedings of the 2018 IEEE International Workshop on Information Forensics and Security https://doi.org/10.1109/WIFS.2018.8630 761 [2] Henry Ajder, Giorgio Patrini, Francesco Cavalli, and Laurence Cullen. 2019. The State of Deepfakes: Landscape, Threats, and Impact. Deeptrace. , 27 pages. https://sensity.ai/reports/ [3] Zahid Akhtar and Tiago H. Falk. 2017. Audio-Visual Multimedia Quality Assessment: A Comprehensive Survey. 5 (2017), 21090–21117. IEEE Access https://doi.org/10.1109/ACCESS.2017.2750918 [4] Ahmed Ali and Steve Renals. 2018. Word Error Rate Estimation for Speech Recognition: e-WER. In Informācijas tehnoloģiju asociācija, 20 – 24. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics https://doi.org/10.18653/v1/P18-2004 [5] Sercan O¨ . Arık, Jitong Chen, Kainan Peng, Wei Ping, and Yanqi Zhou. 2018. Neural Voice Cloning with a Few Samples. In . Curran Associates Inc., 10040–10050. Proceedings of the 32nd International Conference on Neural Information Processing Systems https://papers.nips.cc/paper/201 8/hash/4559912e7a94a9c32b09d894f2bc3c82-Abstract.html [6] Aayush Bansal, Shugao Ma, Deva Ramanan, and Yaser Sheikh. 2018. Recycle-GAN: Unsupervised Video Retargeting. In . Springer, 17 pages. Proceedings of the 2018 European Conference on Computer Vision https://doi.org/10.1007/978-3-030-01228-1 8 [7] Yu-xuan Bao, Tian-liang Lu, and Yan-hui Du. 2020. Overview of Deepfake Video Detection Technology. 47, 9 (2020), 283–292. Datorzinātne https://doi.org/10.11896/jsjkx.200400130 [8] Mikol-aj Bin´kowski, Jeff Donahue, Sander Dieleman, Aidan Clark, Erich Elsen, Norman Casagrande, Luis C. Cobo, and Karen Simonyan. 2019. High Fidelity Speech Synthesis with Adversarial Networks. https://doi.org/10.48550/ARXIV.1909.11646 [9] ![](file:///C:/Users/user/AppData/Local/Temp/msohtmlclip1/01/clip_image050.gif)Madeline Brady. 2020. Deepfakes: A New Desinformation Threat? Report by the Democracy Reporting International. , 9 pages. https://democracy-reporting.org/dri publications/de epfakes-a-new-disinformation-threat/ [10] Umur Aybars Ciftci, Ilke Demir, and Lijun Yin. 2020. FakeCatcher: Detection of Synthetic Portrait Videos using Biological Signals. (2020), 17 pages. IEEE Transactions on Pattern Analysis and Machine Intelligence https://doi.org/10.1109/TPAMI.2020.3009287 [11] Hao Dang, Feng Liu, Joel Stehouwer, Xiaoming Liu, and Anil K. Jain. 2020. On the Detection of Digital Face Manipulation. In . IEEE, 10 pages. Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition https://doi.org/10.1109/CVPR42600.2020.00582 [12] Rohan Kumar Das, Tomi Kinnunen, Wen-Chin Huang, Zhen-Hua Ling, Junichi Yamagishi, Zhao Yi, Xiaohai Tian, and Tomoki Toda. 2020. Predictions of Subjective Ratings and Spoofing Assessments of Voice Conversion Challenge 2020 Submissions. In . International Speech Communication Association, 99–120. Proceedings of the Joint Workshop for the Blizzard Challenge and Voice Conversion Challenge 2020 https://doi.org/10.21437/VCC BC.2020-15 [13] H´ector Delgado, Nicholas Evans, Tomi Kinnunen, Kong Aik Lee, Xuechen Liu, Andreas Nautsch, Jose Patino, Md Sahidullah, Massimiliano Todisco, Xin Wang, and Junichi Yamagishi. 2021. ASVspoof 2021: Automatic Speaker Verification Spoofing and Countermeasures Challenge Evaluation Plan. https://www.asvspoof.org/asvspoof2021/asvspoof2021 evaluation plan.pdf [14] H´ector Delgado, Nicholas Evans, Tomi Kinnunen, Kong Aik Lee, Xuechen Liu, Andreas Nautsch, Jose Patino, Md Sahidullah, Massimiliano Todisco, Xin Wang, and Junichi Yamagishi. 2021. ASVspoof 2021 Challenge - Logical Access Database. https://doi.org/10.5281/zenodo.4837263 [15] H´ector Delgado, Nicholas Evans, Tomi Kinnunen, Kong Aik Lee, Xuechen Liu, Andreas Nautsch, Jose Patino, Md Sahidullah, Massimiliano Todisco, Xin Wang, and Junichi Yamagishi. 2021. ASVspoof 2021 Challenge - Speech Deepfake Database. https://doi.org/10.5281/zenodo.4835108 [16] ![](file:///C:/Users/user/AppData/Local/Temp/msohtmlclip1/01/clip_image051.gif)Anushree Deshmukh and Sunil B. Wankhade. 2021. Deepfake Detection Approaches Using Deep Learning: A Systematic Review. In . Springer, 293–302. Intelligent Computing and Networking: Proceedings of IC-ICN 2020 (Lecture Notes in Networks and Systems, Vol. 146) https://doi.org/ 10.1007/978-981-15-7421-4 27 [17] Xinyi Ding, Zohreh Raziei, Eric C. Larson, Eli V. Olinick, Paul Krueger, and Michael Hahsler. 2020. Swapped Face Detection using Deep Learning and Subjective Assessment. 2020, 1 (2020), 1–12. EURASIP Journal on Information Security https://doi.org/10.1186/s13635-020-00109-8 [18] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang un Cristian Canton Ferrer. 2020. https://doi.org/10.48550/ARXIV.2006.07397 [19] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang un Cristian Canton Ferrer. 2020. The DeepFake Detection Challenge (DFDC) Dataset. arXiv:2006.07397. https://arxiv.org/abs/2006.07397 [20] Nick Dufour un Andrew Gully. 2019. gads Datu ieguldījums dziļo viltus atklāšanas pētniecībā. Google AI Blog. https://ai.googleblog.com/2019/09/contributing-data-to-deepfake-detectio n.html [21] Ricard Durall, Margret Keuper, Franz-Josef Pfreundt un Janis Keuper. 2019. https://doi.org/10.48550/ARXIV.1911.00686 [22] Cristian Canton Ferrer, Brian Dolhansky, Ben Pflaum, Joanna Bitton, Jacqueline Pan, and Jikuo Lu. 2020. Deepfake Detection Challenge Results: An Open Initiative to Advance AI. Meta AI Blog. Tātad https://ai.facebook.com/blog/deepfake-detection-challenge-results-an-open-initia tive-to-advance-ai/ [23] Gereon Fox, Wentao Liu, Hyeongwoo Kim, Hans-Peter Seidel, Mohamed Elgharib, and Christian Theobalt. 2021. Videoforensicshq: Detecting High-Quality Manipulated Face Videos. In . IEEE, 1–6. IEEE 2021. gada starptautiskā multivides konference un izstāde https://doi.or g/10.1109/ICME51207.2021.9428101 [24] Haiying Guan, Andrew Delgado, Yooyoung Lee, Amy N. Yates, Daniel Zhou, Timothee Kheyrkhah, and Jon Fiscus. 2021. User Guide for NIST Media Forensic Challenge (MFC) Datasets. https://doi.org/10.6028/NIST.IR.8377 [25] Yinan He, Bei Gan, Siyu Chen, Yichun Zhou, Guojun Yin, Luchuan Song, Lu Sheng, Jing Shao, and Ziwei Liu. 2021. ForgeryNet: A Versatile Benchmark for Comprehensive Forgery Analysis. In IEEE, 4360 līdz 4369. 2021. gada IEEE/CVF konference par datora redzi un modeļu atpazīšanu https://doi.org/10.1109/CVPR46437.2021.00434 [26] Liming Jiang, Zhengkui Guo, Wayne Wu, Zhaoyang Liu, Ziwei Liu, Chen Change Loy, Shuo Yang, Yuanjun Xiong, Wei Xia, Baoying Chen, Peiyu Zhuang, Sili Li, Shen Chen, Taiping Yao, Shouhong Ding, Jilin Li, Feiyue Huang, Liujuan Cao, Rongrong Ji, Changlei Lu, and Ganchao Tan. 2021. DeeperForensics Challenge 2020 on Real-World Face Forgery Detection: Methods and Results. arXiv:2102.09471. https://arxiv.org/pdf/2102.09471.pdf [27] Liming Jiang, Ren Li, Wayne Wu, Chen Qian, and Chen Change Loy. 2020. DeeperForensics-1.0: A Large-Scale Dataset for Real-World Face Forgery Detection. In IEEE, 2886 līdz 2895 Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition https://doi.org/ 10.1109/CVPR42600.2020.00296 [28] Nal Kalchbrenner, Erich Elsen, Karen Simonyan, Seb Noury, Norman Casagrande, Edward Lockhart, Florian Stimberg, Aaron van den Oord, Sander Dieleman, and Koray Kavukcuoglu. 2018. Efficient Neural Audio Synthesis. https://doi.org/10.48550/ARXIV.1802.08435 [29] Takuhiro Kaneko and Hirokazu Kameoka. 2017. Parallel-Data-Free Voice Conversion Using Cycle-Consistent Adversarial Networks. https://doi.org/10.48550/ARXIV.1711.11293 [30] Tero Karras, Samuli Laine, and Timo Aila. 2019. A Style-based Generator Architecture for Generative Adversarial Networks. In . IEEE, 4401–4410. Proceedings of the 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition https://doi.org/10.1109/CVPR.2019.00453 [31] Ross Kelly. 2021. gada digitālās mākslas izstāde, lai parādītu AI radošo potenciālu. https://www.digit.fyi/digital-art-exhibition-to-showcase-creative-potential-of-ai/ [32] Ali Khodabakhsh, Raghavendra Ramachandra, Kiran Raja, Pankaj Wasnik, and Christoph Busch. 2018. Fake Face Detection Methods: Can They Be Generalized?. In . IEEE, 1–6. Proceedings of the 2018 International Conference of the Biometrics Special Interest Group https://doi.org/10.2 3919/BIOSIG.2018.8553251 [33] Hyeongwoo Kim, Mohamed Elgharib, Hans-Peter Zoll¨ofer, Michael Seidel, Thabo Beeler, Christian Richardt, and Christian Theobalt. 2019. Neural Style-Preserving Visual Dubbing. 38, 6, Article 178 (2019), 13 pages. ACM Transactions on Graphics https://doi.org/10.1145/3355089.3356500 [34] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick P ́erez, Christian Richardt, Michael Zollh ̈ofer un Christian Theobalt. 2018. 37, 4, Article 163 (2018), 14 pages. ACM Transactions on Graphics https://doi.org/ 10.1145/3197517.3201283 [35] Pavel Korshunov un S'ebastien Marcel. 2019. ievainojamības novērtējums un dziļo viltus video atklāšana. Ieee, 1 līdz 6. Proceedings of the 2019 International Conference on Biometrics https://doi.org/10.1109/ICB45273.2019.8987375 [36] Patrick Kwon, Jaeseong You, Gyuhyeon Nam, Sungwoo Park, and Gyeongsu Chae. 2021. KoDF: A Large-scale Korean DeepFake Detection Dataset. In . IEEE, 10724–10733. Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision https://doi.org/10.1109/ICCV 48922.2021.01057 [37] Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. 2020. FaceShifter: Towards High Fidelity And Occlusion Aware Face Swapping. arXiv:1912.13457. https://arxiv.org/abs/1912.13457 [38] Xurong Li, Shouling Ji, Chunming Wu, Zhenguang Liu, Shuiguang Deng, Peng Cheng, Min Yang un Xiangwei Kong. 2021. 32, 2 (2021), 496–518. Journal of Software http://www.jos.org.cn/1000-9825/6140.htm [39] Yuezun Li, Ming-Ching Chang, and Siwei Lyu. 2018. In Ictu Oculi: Exposing AI Created Fake Videos by Detecting Eye Blinking. In . IEEE, 1–7. 2018. gada IEEE Starptautiskā seminārs par informācijas forenziku un drošību https://doi.org/10.1109/WIFS.2018.8630787 [40] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. 2020. Celeb-DF: A Large-Scale Challenging Dataset for DeepFake Forensics. In IEEE, 3204 līdz 3213. Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition https://doi.org/10.1109/CVPR42600. 2020.00327 [41] Ruigang Liang, Peizhuo Lv, Yue Zhao, Peng Chen, Hao Xing, Yingjun Zhang, Jizhong Han, Ran He, Xianfeng Zhao, Ming Li un Kai Chen. 2020. 5, 2 (2020), 1–17. Journal of Cyber Security http://jcs.iie.ac.cn/xxaqxb/ch/re ader/view abstract.aspx?file no=20200202&flag=1 [42] Steven R. Livingstone and Frank A. Russo. 2018. The Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS): A Dynamic, Multimodal Set of Facial and Vocal Expressions in North American English. 13, 5 (2018), 35 pages. Plāns 1 [43] Chen-Chou Lo, Szu-Wei Fu, Wen-Chin Huang, Xin Wang, Junichi Yamagishi, Yu Tsao, and Hsin-Min Wang. 2021. MOSNet: Deep Learning based Objective Assessment for Voice Conversion. arXiv:1904.08352. https://arxiv.org/pdf/1904.08352.pdf [44] Jaime Lorenzo-Trueba, Junichi Yamagishi, Tomoki Toda, Daisuke Saito, Fernando Villavicencio, Tomi Kinnunen, and Zhenhua Ling. 2018. The Voice Conversion Challenge 2018: Promoting Development of Parallel and Nonparallel Methods. In . International Speech Communication Association, 195–202. Proceedings of the Odyssey 2018 The Speaker and Language Recognition Workshop https://doi.org/10.21437/Odyssey.2018-28 [45] Siwei Lyu. 2020. Deepfake Detection: Current Challenges and Next Steps. In Ieee, 6 lappuses 2020 IEEE Starptautiskā konference par Multimediju izstāžu semināriem https://doi.org/10.1109/ICMEW46912.2020.9105991 [46] Yisroel Mirsky and Wenke Lee. 2021. The Creation and Detection of Deepfakes: A Survey. 54, 1, Article 7 (2021), 41 pages. ACM Computing Survey https://doi.org/10.1145/3425780 [47] Gautham J. Mysore. 2015. Can we Automatically Transform Speech Recorded on Common Consumer Devices in Real-World Environments into Professional Production Quality Speech?—A Dataset, Insights, and Challenges. 22, 8 (2015), 1006–1010. IEEE signālu apstrādes vēstules https://doi.org/10.1109/LSP.2014.2379648 [48] Jo ̃ao C. Neves, Ruben Tolosana, Ruben Vera-Rodriguez, Vasco Lopes, Hugo Proenca un Julian Fierrez. 2020. GANprintR: uzlabotas viltus un mākslas stāvokļa novērtēšana sejas manipulācijas noteikšanā. 14, 5 (2020), 1038–1048. IEEE Journal of Selected Topics in Signal Processing https://doi.org/10.1109/JSTSP.2020.3007250 [49] NIST Media Forensics Challenge Team. 2021. Open Media Forensics Challenge 2020-2021 Evaluation Plan. https://mig.nist.gov/MFC/Web/EvalPlan2020/OpenMFC2020EvaluationPlan.pdf [50] Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew Senior un Koray Kavukcuoglu. https://doi.org/10.48550/ARXIV.1609.03499 [51] Debajyoti Pal and Tuul Triyason. 2018. A Survey of Standardized Approaches towards the Quality of Experience Evaluation for Video Services: An ITU Perspective. 2018. gada 13. jūlijs, 131724 (2018), 25 lappuses Starptautiskais digitālo multimediju apraides žurnāls https://doi.org/10.1155/20 18/1391724 [52] Bo Peng, Hongxing Fan, Wei Wang, Jing Dong, Yuezun Li, Siwei Lyu, Qi Li, Zhenan Sun, Han Chen, Baoying Chen, Yanjie Hu, Shenghai Luo, Junrui Huang, Yutong Yao, Boyuan Liu, Hefei Ling, Guosheng Zhang, Zhiliang Xu, Changtao Miao, Changlei Lu, Shan He, Xiaoyan Wu, and Wanyi Zhuang. 2021. DFGC 2021: A DeepFake Game Competition. arXiv:2106.01217. https: //arxiv.org/abs/2106.01217 [53] Ivan Perov, Daiheng Gao, Nikolay Chervoniy, Kunlin Liu, Sugasa Marangonda, Chris Um ́e, Mr. Dpfks, Carl Shift Facenheim, Luis RP, Jian Jiang, Sheng Zhang, Pingyu Wu, Bo Zhou un Weiming Zhang. 2020. https://doi.org/10.48550/ARXIV.2005.05535 [54] Stanislav Pidhorskyi, Donald A. Adjeroh un Gianfranco Doretto. 2020. Ieee, 10 lappuses 2020. gada IEEE/CVF konference par datora redzējumu un modeļu atpazīšanu https://doi.org/10.1109/CVPR42600.2020.01411 [55] Andreas R¨ossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. 2018. FaceForensics: A Large-scale Video Dataset for Forgery Detection in Human Faces. https://doi.org/10.48550/ARXIV.1803.09179 [56] Andreas R ̈ossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies un Matthias Nießner. 2019. FaceForensics++: Mācīšanās atklāt manipulētus sejas attēlus. . IEEE, 1–11. 2019. gada starptautiskā konference par datora redzi https://doi.org/10.110 9/ICCV.2019.00009 [57] Anindya Sen. 2021. Art and Artificial Intelligence: How ‘Deepfakes’ Can Help Create Authentic Museum Experiences! Medium Blog. https://medium.com/art-world-zen/art-and-artificia l-intelligence-how-deepfakes-can-help-create-authentic-museum-experiences-65d8aa 7da29c [58] Kou Tanaka, Hirokazu Kameoka, Takuhiro Kaneko, and Nobukatsu Hojo. 2019. WaveCycleGAN2: Time-domain Neural Post-filter for Speech Waveform Generation. https://doi.org/10.48550/A RXIV.1904.02892 Izstāde [59] Jianhua Tao, Ruibo Fu, Jiangyan Yi, Chenglong Wang, and Tao Wang. 2020. Development and Challenge of Speech Forgery and Detection. 5, 2 (2020), 28 līdz 38. Kiberdrošības žurnāls http://jcs.iie.ac.cn/xxaqxb/ch/reader/view abstract.aspx?file no=20200204&flag=1 [60] Tencent. 2020. Artificial Intelligence White Paper. https://tech.sina.com.cn/roll/2020-07-14/doc-iivhvpwx5201226.shtml [61] Justus Thies, Michael Zollh, un Matthias Niessner. 2019. aizkavēta neironu rādīšana: attēlu sintēze, izmantojot neironu tekstūras. 38, Article 66 (2019), 12 pages. Issue 4. ACM darījumi ar grafiku https://doi.org/10.1145/3306346.3323035 [62] Tomoki Toda, Ling-Hui Chen, Daisuke Saito, Fernando Villavicencio, Mirjam Wester, Zhizheng Wu un Junichi Yamagishi. 2016. Starptautiskā runas komunikācijas asociācija, 1632–1636 Proceedings of Interspeech 2016 https://doi.org/10.21437/Interspeech.2016-1066 [63] Massimiliano Todisco, Xin Wang, Ville Vestman, Md Sahidullah, Hector Delgado, Andreas Nautsch, Junichi Yamagishi, Nicholas Evans, Tomi Kinnunen, and Kong Aik Lee. 2019. ASVspoof 2019: Future Horizons in Spoofed and Fake Audio Detection. arXiv:1904.05441. https://arxiv.org/ pdf/1904.05441.pdf [64] Ruben Tolosana, Ruben Vera-Rodriguez, Julian Fierrez, Aythami Morales un Javier Ortega-Garcia. 64 (2020), 131–148. Information Fusion https://doi.org/10.1016/j.inffus.2020.06.014 [65] Xin Tong, Luona Wang, Xiaoqin Pan un Jingya Wang. 2020. gada pārskats par Deepfake: Damokļa zobens AI. . IEEE, 265–273. 2020 Starptautiskā konference par datora redzi, attēlu un dziļo mācīšanos https://doi.org/10.1109/CVIDL51233.2020.00-88 [66] Luisa Verdoliva. 2020. gads Media Forensics un DeepFakes: pārskats. 14 5 (2020), 910 līdz 932. IEEE žurnāls par atlasītajām tēmām signālu apstrādē https://doi.org/10.1109/JSTSP.2020.3002101 [67] Xin Wang, Junichi Yamagishi, Massimiliano Todisco, H ́ector Delgado, Andreas Nautsch, Nicholas Evans, Md Sahidullah, Ville Vestman, Tomi Kinnunen, Kong Aik Lee, Lauri Juvela, Paavo Alku, Yu-Huai Peng, Hsin-Te Hwang, Yu Tsao, Hsin-Min Wang, S ́ebastien Le Maguer, Markus Becker, Fergus Henderson, Rob Clark, Yu Zhang, Quan Wang, Ye Jia, Kai Onuma, Koji Mushika, Takashi Kaneda, Yuan Jiang, Li-Juan Liu, Yi-Chiao Wu, Wen-Chin Huang, Tomoki Toda, Kou Tanaka, Hirazu Kameoka, Ingmar Steiner, Driss Matrouf, Jean-Francois Bonastre 64 (2020), 27 pages. Datoru valoda un runa https://doi.org/10.1016/j.csl.2020.101114 [68] Mirjam Wester, Zhizheng Wu un Junichi Yamagishi. 2016. gada balss konversijas izaicinājuma analīze 2016. gada novērtējuma rezultāti. Starptautiskā runas komunikācijas asociācija, 1637–1641 Interspeech 2016 konferences darba kārtība https://doi.org/10.21437/Interspeech.201 6-1331 [69] Junichi Yamagishi, Massimiliano Todisco, Md Sahidullah, H ́ector Delgado, Xin Wang, Nicholas Evans, Tomi Kinnunen, Kong Aik Lee, Ville Vestman un Andreas Nautsch. 2019. ASVspoof 2019: Automatic Speaker Verification Spoofing un Countermeasures Challenge Evaluation Plan. https://www.asvspoof.org/asvspoof2019/asvspoof2019 evaluation plan.pdf [70] Zhao Yi, Wen-Chin Huang, Xiaohai Tian, Junichi Yamagishi, Rohan Kumar Das, Tomi Kinnunen, Zhen-Hua Ling, and Tomoki Toda. 2020. Voice Conversion Challenge 2020 – Intra-lingual Semi-parallel and Cross-lingual Voice Conversion –. In . International Speech Communication Association, 80–98. Proceedings of the Joint Workshop for the Blizzard Challenge and Voice Conversion Challenge 2020 https://doi.org/10.21437/VCC BC.2020-14 [71] Mohammed A. Younus un Taha M. Hasan. 2020. saīsināts skatījums par Deepfake Video detekcijas paņēmieniem. IEEE, 115 līdz 120 2020. gada 6. starptautiskās inženierzinātņu konferences darba kārtība https://doi.org/10.1109/IEC49899.2020.9122916 [72] Guangtao Zhai un Xiongkuo 2020. gada min. uztveres attēla kvalitātes novērtējums: aptauja. 63, Article 211301 (2020), 52 pages. Zinātne Ķīna Informācijas zinātne https://doi.org/10.1007/s11432-019-2757-1 [73] Teng Zhang, Lirui Deng, Liang Zhang, and Xianglei Dang. 2020. Deep Learning in Face Synthesis: A Survey on Deepfakes. In Ieee, 67 līdz 70 IEEE III Starptautiskā datortehnikas un komunikāciju tehnoloģiju konference https://doi.org/10.1109/CCET5090 1.2020.9213159 [74] Yuxuan Zhang, Huan Ling, Jun Gao, Kangxue Yin, Jean-Francois Lafleche, Adela Barriuso, Antonio Torralba, and Sanja Fidler. 2021. DatasetGAN: Efficient Labeled Data Factory with Minimal Human Effort. In IEEE, 10140 līdz 10150 Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition https://doi.org/10.1109/CVPR46437.2021.01001 [75] ![](file:///C:/Users/user/AppData/Local/Temp/msohtmlclip1/01/clip_image051.gif)Yuanhan Zhang, ZhenFei Yin, Yidong Li, Guojun Yin, Junjie Yan, Jing Shao, un Ziwei Liu. 2020. CelebA-Spoof: Liela mēroga sejas anti-spoofing Dataset ar bagātām piezīmēm. Sprindžers 70–85 2020. gada Eiropas konference par datora redzi https://doi.org/10.1007/97 8-3-030-58610-2 5 [76] Yuanhan Zhang, Zhenfei Yin, Jing Shao, Ziwei Liu, Shuo Yang, Yuanjun Xiong, Wei Xia, Yan Xu, Man Luo, Jian Liu, Jianshu Li, Zhijun Chen, Mingyu Guo, Hui Li, Junfu Liu, Pengfei Gao, Tianqi Hong, Hao Han, Shijie Liu, Xinhua Chen, Di Qiu, Cheng Zhen, Dashuang Liang, Yufeng Jin, and Zhanlong Hao. 2021. CelebA-Spoof Challenge 2020 on Face Anti-Spoofing: Methods and Results. arXiv:2102.12642. https://arxiv.org/pdf/2102.12642.pdf [77] Bo Zhao, Shaozeng Zhang, Chunxue Xu, Yifan Sun, and Chengbin Deng. 2021. Deep Fake Ge-ography? When Geospatial Data Encounter Artificial Intelligence. 48, 4 (2021), 338 līdz 352 Cartography and Geographic Information Science https://doi.org/10.1080/15230406.2021.1910075 [78] Peng Zhou, Xintong Han, Vlad I. Morariu un Larry S. Davis. 2017. gada divu plūsmu neironu tīkli tampered sejas noteikšanai. IEEE, 1831 – 1839 2017. gada IEEE konferences par datora redzi un modeļu atpazīšanu darbnīcas https://doi.org/10.1109/CVPRW.2017.229 [79] Tianfei Zhou, Wenguan Wang, Zhiyuan Liang un Jianbing Shen. 2021. IEEE, 5774 līdz 5784 2021. gada IEEE/CVF konference par datora redzi un modeļu atpazīšanu https://doi.org/10.1109/CVPR46437.2021.00572 [80] Jun-Yan Zhu, Taesung Park, Phillip Isola un Alexei A. Efros. 2017. Nepareizs attēla un attēla tulkojums, izmantojot cikla konsekventu pretinieku tīklu. IEEE, 2242 līdz 2251 2017. gada IEEE starptautiskā konference par datora redzi https://doi.org/10.1109/ICCV.2 017.244 [81] Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma, and Yu-Gang Jiang. 2020. WildDeepfake: A Challenging Real-World Dataset for Deepfake Detection. In ACM, 2382 līdz 2390 Proceedings of the 2020 28th ACM International Conference on Multimedia https://doi.org/10.1145/339417 1.3413769 This paper is under CC by 4.0 Deed (Attribution 4.0 International) license. available on arxiv Šis dokuments ir Attiecības ir licencētas saskaņā ar CC by 4.0 Deed (Attribution 4.0 International) licenci. Pieejams arhīvā