Kako istraživači meraju, detektuju i usklađuju manipulaciju AI

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 (Universitet u Kentu, Velika Britanija) Virginia N. L. Franqueira, V.Franqueira@kent.ac.uk (University of Kent, UK) Shujun Li, S.J.Li@kent.ac.uk (Universitet u Kentu, Velika Britanija) Abstract Abstrakcija Nedavni napori u AI, posebno duboko učenje, pridonijeli su značajnom porastu stvaranja novih realističnih sintetičkih medija (video, slika i audio) i manipulacije postojećim medijima, što je dovelo do stvaranja novog termina „deepfake“. Na osnovu i istraživačke literature i resursa na engleskom i kineskom jeziku, ovaj članak daje sveobuhvatan pregled deepfake-a, pokrivajući više važnih aspekata ovog novog koncepta, uključujući 1) različite definicije, 2) uobičajeno korištene performanse i standarde, i 3) skupove podataka, izazove, takmičenja i referentne oznake povezane s deepfake-om. Pored toga, članak takođe izvještava o meta-reviziji 12 odabranih istraživačkih radova povezanih s deepfake : Deepfake, Anketa, Definicija, Datasets, Benchmarks, izazovi, takmičenja, standardi, performanse Metrika. Ključne reči 1 Uvod Posebno, state-of-the-art metode dubokog učenja omogućile su stvaranje „deepfakes“, manipuliranih ili sintetičkih medija čija stvarnost nije lako prepoznatljiva ljudskim okom. Iako je deepfake relativno novi fenomen (što se po prvi put pojavio krajem 2017. godine), njegov rast je bio izuzetan. Prema izvješćima 2019 i 2020 Deeptrace o stanju deepfake [ Broj deepfake videa na engleskom jeziku porastao je sa 7.964 (decembar 2018) na 14.678 (juli 2019) na 85.047 (decembar 2020), što predstavlja povećanje od 968% od 2018. do 2020. godine. 2 U ovom radu, preispitujemo postojeći istraživački ekosistem povezan sa deepfake-om u pogledu različitih aspekata, uključujući performanse i standarde, skupove podataka, izazove, natjecanja i referentne vrednosti. Nadalje, pružamo meta-pregled 12 odabranih istraživačkih radova povezanih sa deepfake-om koji sistematski pokrivaju nekoliko dodatnih aspekata osim navedenih, kao što su uspoređivanje performansi, ključni izazovi i preporuke. Uprkos tome što je izraz izuzetno popularan, postoji nedostatak konsenzusa o definiciji „deepfake“ i granica između deepfakes i non-deepfakes nije jasan rez. Za ovo istraživanje, mi usvojimo relativno inkluzivniji pristup za pokrivanje svih oblika manipuliranih ili sintetičkih medija koji se smatraju deepfakes u širem smislu. Mi takođe pokrivamo blisko povezane teme uključujući biometrijsku i multimedijsku forenziku, budući da se deepfakes često koriste za pokretanje prezentacijskih napada protiv biometrijskih sistema za autentifikaciju i otkrivanje deepfakes može se smatrati dijelom multimedijske forenzike. 1.1 Definicije pojma Deepfake Kao što ime sugerira, pojam "deepfake" potječe od kombinacije "deep" (koji se odnosi na (DL)) i "fake". To se obično koristi da se odnosi na manipulaciju postojećih medija (slika, video i / ili audio) ili generaciju novih (sintetičkih) medija koristeći DL-based pristupe. Najčešće raspravljani deepfake podaci su lažne slike lica, lažne govora krivotvorenja, i lažne video koji kombiniraju i lažne slike i lažne govora krivotvorenja. Dok ima "lažno" u reči ukazuje na manipulirane ili sintetizovane medije, postoji mnogo benignih aplikacija deepfake tehnologije, npr, za zabavu i kreativne umetnosti. U tom pogledu, još jedan termin "duboka sinteza" je predložena kao neutralniji zvuk alternativu [ ].Ta nova definicija, međutim, nije široko prihvaćena. Duboko učenje 60 Pored nedostatka univerzalne definicije, kao što je već spomenuto, granica između dubokih laži i ne-dubokih laži zapravo nije jasan rez. Prvo, detekcija dubokih faksa često prati vrlo slične pristupe detekciji tradicionalnih krivotvorina generisanih bez upotrebe DL tehnika. Napredne metode detekcije takođe su počele iskoristiti DL kako bi poboljšale svoje performanse, ali ne moraju nužno znati kako je ciljni medij stvoren (duboko ili ne). U određenoj mjeri, moglo bi se tvrditi da detekcija dubokih faksa ne uključuje razvoj deepfake-specifičnih metoda (iako neki istraživači to odluče), već robusniji i univerzalniji detektor koji može rukovati bilo kojim (dubokim ili ne) lažnim medijima. To se može vidjeti za dvije usko povezane teme: biometrija i multimedijalna forenzika. Za biometrijsku, postoji trend korišćenja tehnika Drugo, može se tvrditi da se riječ „deep“ u „deepfake“ ne odnosi nužno na upotrebu „deep learning“, već na bilo koju „deep“ (tj. sofisticiranu) tehnologiju koja stvara vrlo vjerodostojne lažne medije. ] smatrali su deepfake kao audio-vizualnu manipulaciju koristeći „spektar tehničke sofisticiranosti ... i tehnika“. i , koji se odnosi na „manipulaciju na niskom nivou audiovizualnih medija stvorenih pomoću (lakše) pristupačnog softvera [ili bez softvera] kako bi se ubrzao, usporio, obnovio ili ponovo kontekstualizovao sadržaj“. Ovo šire razumijevanje „deepfake“ takođe su usvojili zakonodavci za nova zakonodavstva koja se bore protiv zlonamjernih deepfakes. 9 lažna Krivotvorenje Zakon o zabrani lažnih lažnih lažnih lažnih lažnih lažnih lažnih lažnih lažnih lažnih lažnih lažnih lažnih lažnih lažnih laži 2018 U skladu sa člankom 104. stavkom 2. točkom (b): ” „deep fake” znači audiovizualni zapis koji je stvoren ili izmenjen na takav način da bi se razumnom posmatraču lažno činilo da je autentičan zapis stvarnog govora ili ponašanja pojedinca. 2019 DEEP FAKES Zakon o odgovornosti2: U skladu sa člankom 104. stavkom 3. točkom (a): Pojam „duboko krivotvorenje” znači bilo koji video snimak, film, zvučni snimak, elektronička slika ili fotografija, ili bilo koje tehnološko prikazivanje govora ili ponašanja koje je bitno rezultat tih snimaka. (A) which appears to authentically depict any speech or conduct of a person who did not in fact engage in such speech or conduct; and ” (b) čija je proizvodnja znatno zavisila od tehničkih sredstava, a ne od sposobnosti druge osobe da se fizički ili verbalno pretvara u takvu osobu. Kao što možemo vidjeti iz gore navedenih pravnih definicija „deepfake“, upotreba DL-a kao tehnologije uopšte nije spomenuta. 1.2 Opseg i doprinos Na osnovu gore navedenih rasprava o definicijama dubokog krivotvorenja, možemo vidjeti da nije uvek jednostavno ili smisleno razlikovati duboke krivotvorenja od ne-dubokih krivotvorenja. Pored toga, za naš fokus na evaluaciji performansi i usporedbi, granica između dubokih krivotvorenja i ne-dubokih krivotvorenja je još više zamagljena. To je zato što je DL samo poseban (dublji) oblik strojnog učenja (ML), i kao rezultat, DL i ne-duboki ML metode dijele mnoge zajedničke koncepte, metrike i postupke. Unatoč činjenici da se deepfake može shvatiti u mnogo širem smislu, u ovom radu imamo dovoljno uski fokus da izbegnemo pokrivanje previše tema. Za metrije i standarde, odlučili smo da uključimo sve one koji se obično koriste za evaluaciju općih metoda ML i one koji su posebno definisani za evaluaciju metoda stvaranja ili otkrivanja dubokih laži. Za skupove podataka, izazove, natjecanja i referentne oznake, razmotrili smo one koje se odnose na lažne medije obuhvaćene anketama koje se odnose na deepfake i one koje izričito spominju pojam “deepfake” ili sličan pojam. Za meta-pregled, razmotrili smo samo anketne radove čiji su autori eksplicitno upućivali na pojam “deepfakes” u meta podacima (naslov, abstrakt i ključne reči). 2 Metodologija Istraživački radovi obuhvaćeni ovom anketom (tj. istraživački radovi povezani s dubokom krivotvorenjem) identificirani su putem sistematskih pretraga u znanstvenim bazama podataka, Scopusu i China Online Journals (COJ) Sljedeće upite za pretraživanje korištene su za pretraživanje Scopusa i COJ-a, odnosno: 3 (deepfake* OR deep-fake* OR “deep fake*”) I (review OR survey OR overview OR systemati* OR SoK) (deepfake OR 深度偽造) I (综述 OR 进展) Pretraga je vratila 41 anketni članak na engleskom i 15 anketnih radova na kineskom jeziku. Od tih radova, osam objavljenih na engleskom i četiri objavljena na kineskom su odabrana za razmatranje. Izazovi povezani s Deepfake-om, natjecanjima i referentnim vrednostima identificirani su putem više izvora: odabranih anketa, istraživačkih radova iz ličnih zbirki suautora, Google Web pretraživanja i ručne inspekcije web stranica velikih konferencija vezanih za AI održanih 2020. i 2021. godine (gde se takvi izazovi i natjecanja rutinski organizuju). , ICCV, CVPR, AAAI, ICML, ICLR, KDD, SIGIR, WWW, i mnogi drugi. Pored toga, sastavljena je sveobuhvatna lista skupova podataka na osnovu odabranih anketa i utvrđenih izazova, natjecanja i referentnih standarda. Relevantni standardi identificirani su uglavnom putem istraživačkih radova obuhvaćenih ovom anketom, osobnog znanja suautora i Google Web pretraživanja. Za metriku performansi, pokrili smo one koji se obično koriste na osnovu relevantnih standarda, anketa i utvrđenih izazova, natjecanja i referentnih standarda. 4 3 Deepfake-povezane performanse Metrike i standardi U ovom istraživanju fokusiramo se na evaluaciju performansi i usporedbu metoda za generiranje i detekciju dubokih laži. Mjere koje se koriste za takve evaluacije performansi su u srcu naših rasprava. U ovom odjeljku pregledamo mjere performansi koje se obično koriste za evaluaciju algoritama za generiranje i detekciju dubokih laži. Imajte na umu da se sve metrike obuhvaćene ovim odjeljkom obično koriste i za evaluaciju performansi sličnih sistema koji nisu za generiranje ili detekciju dubokih laži. Stoga, ovaj odjeljak može se smatrati vrlo kratkim tutorialom o opštoj metrici performansi. U poslednjem pododjeljku, takođe ćemo ukratko raspravljati o tome kako su povezane parametre performansi obuhvaćeni formalnim standardima. „Formalni standardi” odnose se na standarde definisane formalnim postupkom, često od strane jednog ili više uspostavljenih tijela za normizaciju, kao što je Međunarodna organizacija za normizaciju (ISO). Međunarodna elektrotehnička komisija (IEC) Imajte na umu da razmatramo širok raspon dokumenata koji su definisani kao standardi od strane tijela za normizaciju, npr. Međunarodne telekomunikacijske unije (ITU). preporuke i ISO tehnička izvješća (TRs). 5 6 7 3.1 Matrix zbunjenosti Deepfake detekcija je prvenstveno problem binarne klasifikacije. ili i ispušta binarnu vrednost koja ga označava kao ili Na primer, sistem za detekciju dubokih laži će uzeti sumnjivu sliku kao unos koji bi mogao biti ili i proizvodnje ili . Stvarno pozitivno U stvari negativno Predviđeno pozitivno Predviđeno negativno U stvari, lažna actually real Predviđeno lažno Realni predviđanja Osnovni alat koji se koristi u procjeni binarnog klasifikatora je koji sažima uspeh i neuspeh klasifikacijskog modela. Na jednoj osovini su dva vrednosti i na drugoj osovini su dvije Vrednosti. klasifikacija je (istina pozitivna i istina negativna) kada se stvarne i predviđene vrijednosti podudaraju. (lažno pozitivno i lažno negativno) kada se stvarne i predviđene vrijednosti ne podudaraju. prikazuje matricu zbunjenosti za binarni deepfake klasifikator (detektor). Dve ćelije u zelenom, TP (broj ) i TN (broj ), ukazuju na točne rezultate predviđanja, a dvije ćelije u crvenom, FN (broj ) и ФП ( Број ), ukazuju na dvije različite vrste grešaka prilikom pravljenja pogrešnih rezultata predviđanja. confusion matrix actual Predviđanje successful/correct/true pogrešno / pogrešno / pogrešno 1 true positives true negatives false negatives false positives Tablica 1: Matrica zbunjenosti za binarni klasifikator za otkrivanje dubokog krivotvorenja. fake (predicted) real (predicted) fake (actual) TP FN real (actual) FP TN lažna (stvarna) Knjiga Ujedinjeni narodi Pravi (trenutni) FP za gnj 3.2 Preciznost i podsjetnik Na osnovu četiri osnovne vrednosti uvedene u odjeljku , to jest, TP, TN, FP i FN, definisali smo dvije važne performanse metrik za binarni klasifikator – i . 3.1 precision recall Preciznost binarnog klasifikatora definisana je kao frakcija od Uzorci među svim U matrici zbunjenosti, to je frakcija istinskih uzoraka u prvoj koloni. može se formalno definirati kao Eq. ( U svakom slučaju) Stvarno pozitivno Predviđeni pozitivni rezultati 1 Kada se "prirodni" omjer pozitivnih i negativnih uzoraka značajno razlikuje od set testova, često je korisno prilagoditi težinu lažnih pozitivnih, što dovodi do (npr) definisan u Eq. ( gde je 0 je težina određena odnosom između negativnih i pozitivnih uzoraka. weighted precision 2 α » Podsećanje na binarni klasifikator je frakcija od samples among the uzorke, kao što je prikazano u Eq. ( U matrici zbunjenosti, to je frakcija istinskih uzoraka u prvom redu. Predviđeno pozitivno Stvarno pozitivno 3 Let us consider an example binary classifier that predicts if an image from a database containing both deepfake and real (authentic) images is fake or not. Precision of the classifier is the fraction of correctly classified images among all images classified as deepfake. On the other hand, recall is the fraction of deepfake images identified by the classifier, among all deepfake images in the database. 3.3 Istinite i lažne pozitivne stope Usredotočujući se na predviđene pozitivne uzorke, možemo definisati i dvije metrike: (TPR), takođe nazvan (CDR), kao udio predviđenih pozitivnih uzoraka među stvarno pozitivnim uzorcima i (FPR), takođe nazvan (FAR), kao udio predviđenih pozitivnih uzoraka među stvarno negativnim uzorcima, kao što je prikazano u Eqs. ( ) i true positive rate correct detection rate false positive rate false alarm rate 4 ( U matrici zbunjenosti, TPR je udio predviđenih pozitivnih uzoraka u prvom redu, a FPR je udio predviđenih pozitivnih uzoraka u drugom redu. Uslovi poslovanja ( U svakom slučaju. ) 5 recall 3 3.4 Istinite i lažne negativne stope Slično istinskim i lažnim pozitivnim stopama, možemo definisati još dvije stope koje se fokusiraju na negativne predviđene rezultate: (TNR) označavajući udio predviđenih negativnih uzoraka među stvarno negativnim uzorcima, i (FNR) ukazuje na frakciju predviđenih negativnih uzoraka među stvarno pozitivnim uzorcima, kao što je prikazano u Eqs. ( i ) i ( U svakom slučaju) true negative rate false negative rate 6 7 3.5 Osetljivost i specifičnost U nekim primjenama binarnih klasifikatora, posebno u biologiji i medicini, TPR i TNR se češće koriste, a često se nazivaju (TPR) i (TNR) Fokus tih dva termina je na dva tipa ispravnosti predviđenih rezultata.Ovi su manje korišteni u istraživanjima vezanim za duboku laž, stoga se nećemo odnositi na njih u ostatku ovog papira. sensitivity specificity 3.6 Jednaka stopa greške Fokusiranje na stope grešaka znači da moramo razmotriti FPR i FNR. Te dvije stope obično sukobljuju jedna s drugom tako da smanjenje jedne stope obično dovodi do povećanja druge. stoga, umjesto da pokušamo smanjiti obje stope grešaka u isto vrijeme, što je obično nemoguće, realniji zadatak u praktičnim aplikacijama je pronaći pravu ravnotežu tako da su oboje ispod prihvatljivog praga. U nekim aplikacijama, kao što je biometrija, ljudi su posebno zainteresovani za uspostavljanje tzv. (EER) ili EER/CER nije nužno dobra metrika za neke aplikacije, pogotovo kada su dvije vrste grešaka različitih nivoa važnosti, npr. za otkrivanje kritičnih deepfakes (npr. lažne vijesti koje mogu uticati na to kako ljudi bacaju glasove) često možemo tolerirati više lažnih pozitivnih (lažne alarme) nego lažnih negativnih (promašene alarme). equal error rate crossover error rate 3.7 Tačnost i F-score In addition to the EER/CER, there are also other metrics that try to reflect both types of errors, in order to give a more balanced indication of the overall performance of a binary classifier. The two most commonly used are i (Takođe se zove ). Both metrics can be defined based on the four fundamental values (TP, TN, FP, and FN). accuracy F-score F-measure Tačnost binarnog klasifikatora definisana je kao frakcija od Uzorci (pravi pozitivni i pravi negativni) među ukupnim brojem uzoraka koji su klasificirani, kao što je prikazano u Eq. ( U svakom slučaju) Ispravno predviđeno 8 F-score binarnog klasifikatora je zapravo porodica metrika. Njegov opći oblik može se opisati na osnovu parametara kao što je definisano u Eq. ( U svakom slučaju) β 9 Najčešće korišćena verzija svih F-scora je tzv. , što je zapravo F-score sa = 1. točnije, to je definisano kao što je prikazano u Eq. ( U svakom slučaju) F1-score β 10 3.8 Prijemnik Operativna karakteristična krivina i područje ispod krivine (ROC) krivulje se obično koriste za mjerenje performansi binarnih klasifikatora koji daju ocjenu (ili verovatnoću) predviđanja. Receiver operating characteristic Consider the following. Let be the set of all test samples and let the output scores (Naravno ) (for all ∈ ) lie in the interval [ ] on the real line. Let U ovom slučaju [ ] be a prediction threshold for the model, and assume that the classifiers works as follows for all ∈ : S f s s S a, b i t a, b s S It is easy to see that, for = za , all the samples will be classified as positive, leading to FN = TN = 0 so TPR = FPR = 1; while for = za , all the samples will be classified as negative, leading to FP = TP = 0 so TPR = FPR = 0. For other threshold values between and , the values of TPR and FPR will normally be between 0 and 1. By changing from dva kontinuirano, obično možemo dobiti kontinuiranu krivulju koja opisuje kako se TPR i FPR vrednosti mijenjaju od (0,0) do (1,1) na 2D ravnini. t a t b a b t a b Za slučajnog klasifikatora, pretpostavljajući da (Naravno ) ravnomjerno raspodijeljen na [ ] 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 gubitak Još jedna široko korišćena metrika performansi za binarne klasifikatore koja može vratiti ocjenu vjerojatnosti za predviđenu oznaku je Za binarnu klasifikaciju sa pravim oznakom ∈ {0*,* 1} i procenjena verovatnoća = Pr( = 1), the log loss per sample is the negative log-likelihood of the classifier given the true label, defined as shown in Eq. ( U svakom slučaju) log loss y p y 12 Given a testing set with Ukoliko je potrebno, procenat gubitka dnevnog reda binarnog klasifikatora može se izračunati pomoću Eq. ( ), where 1 Ako je -th uzorak je istinit i 0 ako je lažan, i ˆ je predviđena vjerojatnost = 1. n 13 yi i y i Džej 3.10 Proširenje na višeklasne klasifikatore 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 Perceptualna procjena kvaliteta (PQA) Metrika Prema definiciji, glavni cilj deepfake-a je da se ljudskim potrošačima (slušnicima ili gledateljima) otežava ili onemogućuje razlikovanje lažnih medija od stvarnih medija. Stoga, prilikom procjene kvalitete deepfake medija, ključna je kvaliteta koju percipiraju ljudski potrošači medija. Ovo poziva na subjektivnu procjenu perceptualne kvalitete deepfake medija kao „zlatnog standarda“. Najčešće korišćena metrika subjektivne perceptualne kvalitete (PQA) za audio-vizualne signale je (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” Imajte na umu da se standardi ITU fokusiraju više na tradicionalnu perceptivnu kvalitetu, tj. koliko dobar signal izgleda ili zvuči, čak i ako izgleda ili zvuči ne stvarno (npr. Previše glatko). S druge strane, za deepfakes, fokus je prilično drugačiji jer ono što je važno je stvarnost i prirodnost stvorenog medija, tj. koliko stvarno i prirodno izgleda ili zvuči, čak i ako je niskog kvaliteta. 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 (SSIM) for FR PQA of digitalimages and videos defined as in Eqs. (14), (15), and (16), respectively, where X = {xi} n i is the reference (the original signal), Y = {yi} n i is the signal whose visual quality is assessed, n is the number of pixels in X and Y , L is the maximum possible pixel value of X and Y (e.g., 255 for 8-bit gray-scale images), c1 = (k1L) 2 and c2 = (k2L) 2 ) are two stabilising parameters (k1 = 0.01 and k2 = 0.03 by default). For more about PQA metrics for different types of multimedia signals, we refer readers to some relevant surveys [3, 51, 72]. mean squared error peak signal-to-noise ratio structural similarity index measure 3.12 More about Standards Mnoge od osnovnih pokazatelja performansi opisanih u ovom odeljku su široko korišteni od strane deepfake istraživača kao de facto standarde, npr, EER, log gubitak i MOS su široko korišteni u izazove povezane s deepfake (vidjeti odjeljak ). Also, the combination of precision, recall and F1-score has been widely used to assess performance of binary classifiers. While there have been a number of ITU standards on PQA to date, there does not seem to be many standardisation efforts on the performance metrics for evaluation of binary classifiers. This was the case until at least 2017, when ISO and IEC jointly set up the ISO/IEC JTC 1/SC 42 , a standardisation subcommittee (SC) focusing on AI under ISO/IEC JTC 1 Zajednički tehnički odbor za standardizaciju „informacione tehnologije“. 5 10 11 One recent effort that ISO/IEC JTC 1/SC 42 made is to produce the ISO/IEC TR 24029-1:2021 “Artificial Intelligence (AI) – Assessment of the robustness of neural networks – Part 1: Overview” , 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 Što se tiče pokazatelja performansi, još dva tekuća radna stavka ISO/IEC JTC 1/SC 42 koja zaslužuju pažnju su kako slijedi: ISO/IEC DTS (Draft Technical Specification) 4213 „Informacione tehnologije – Veštačka inteligencija – Procjena performansi klasifikacije strojnog učenja”13 ISO/IEC AWI (Approved Work Item) TS (Technical Specifications) 5471 “Artificial intelligence – Quality evaluation guidelines for AI systems” 14 Iako je ISO/IEC JTC 1/SC 42 stvoren vrlo nedavno, još jedan pododbor za standardizaciju u okviru ISO/IEC JTC1 ima mnogo dužu istoriju od gotovo 20 godina: 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) and (FRR), which refer to the standard FPR and FNR, respectively. This standard also deprecates the use of single-number metrics including the EER and AUC (which were widely used in biometrics-related research in the past). 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) referring to the standard FPR, (NPCER/BPCER) koji se odnosi na standardni FNR, i (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 Kada se duboko krivotvorene slike ili videozapisi koriste za krivotvorenje biometrijskog sistema, takve standardizovane metrike će postati relevantne. 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 In this section, we cover all deepfake-related datasets we identified from the meta-review of deepfake-related survey papers, deepfake-related challenges, competitions and benchmarks covered, one online collection of deepfake-related datasets on 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 datoteke za slike (Naravno ]: 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 (Naravno ]: This dataset contains 131,500 face images, including 78,500 images extracted from 150 videos in the FaceForensics dataset and 53,000 images extracted from 150 fake videos collected from 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 and an API Sastav podataka generated.photos ima dobru raznolikost: pet dobnih skupina (djeca, deca, mladi, odrasli, srednjih godina), dva spola (muški i ženski), četiri etničke skupine (bijeli, crni, latinski, azijski), četiri boje očiju (smeđe, sivo, plavo, zeleno), četiri boje kose (smeđe, crno, plavo, sivo), tri dužine kose (kratka, srednja, duga), izraze lica, tri pozice glave (pred, levo, desno), dvije emocije (veseli se i neutralno), dva stila lica (prirodno, ukrašeno). (Prema brojnim istraživačkim radovima koje smo pročitali, raniji set podataka 100K-Faces objavljen je od strane generated.photos za akademsko istraživanje generated.photos datasets 22 23 24 [ ]: This dataset includes 19,457 face images, including 7,948 deepfake images generated from on 175 forged videos collected online and 11,509 real face images collected from various online sources. (Table 2 of the paper shows the dataset size is 19,509, but the dataset downloaded from pCloud contains just 19,457 images.) 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 datasete (tri vrste objekata – spavaće sobe, automobili i mačke – za potonje). Imajte na umu da ime “100K-Generated-Images” nije bio odgovarajući kao autori [ ] samo je koristio ovo da nazove podfolder svog zajedničkog prostora na Google Drive-u, ali je korišćen u jednom od anketa [ ]. 100K-Generated-Images 30 25 26 30 65 [ ]: This dataset contains 420,053 images of celebrities, including 156,930 real ones downloaded using Google Image API and 263,123 fake face-swapped ones created using two different methods (Nirkin’s method and Auto-Encoder-GAN) Ding et al.’s swapped face dataset 17 [ ]: 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 . It is the replaced version of the dataset, koji sadrži 150.000 slika lica generisanih pomoću ranije verzije 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 Sastav podataka 100K-Faces (not available any longer, see the description of generated.photos datasets), and . Faces-HQ 21 27 28 29 thisperson-doesnotexist.com [ ]: This dataset includes 625,537 synthesised face images of 10,177 celebrities, with 43 rich attributes on face, illumination, environment and spoof types. The real images were selected from the CelebA dataset . The 43 attributes include 40 for real images, covering all facial components and accessories (e.g., skin, nose, eyes, eyebrows, lip, hair, hat, eyeglass), and 3 for fake images, covering spoof types, environments and illumination conditions. CelebA-Spoof 75 30 [ ]: This dataset contains 299,039 images, including 58,703 real images sampled from three datasets (FFHQ , CelebA Uslovi korišćenja + ) 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 [ ]: 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 (Naravno ]: 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 [ ]: This dataset contains 10,000 high-quality forgery videos, with video- and face-level annotations. The dataset focuses on a more challenging case for forgery detection: each video involves one to 15 individuals, but only some (a minority of) faces are manipulated. FFIW K (Face Forensics in the Wild) dataset 79 FFIW [ ]: This dataset contains 37,942 videos of paid subjects (395 Koreans and 8 Southeastern Asians), including 62,166 real videos and 175,776 fake ones created using six methods – FaceSwap, DeepFaceLab, FSGAN, First Order Motion Model (FOMM), Audio-driven Talking Face HeadPose (ATFHP) and Wav2Lip. The videos cover a balanced gender ratio and a wide range of age groups. 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 that started in 2016 is a major challenge series on VC. Datasets released from this challenge series are very different from other deepfake datasets: the deepfake data is not included in the original dataset created by the organisers of each challenge, but in the participant submissions (which are retargeted/fake utterances produced by VC systems built by participants). The challenge datasets also include the evaluation (listening-based) results of all submissions. Some fake utterances may be produced by DL-based VC systems, so we consider all datasets from this challenge series relevant for our purpose of this survey. 38 [ ]: The original dataset created by the challenge organisers was derived from the DAPS (Device and Produced Speech) Dataset [ ]. Sadrži 216 izjava (162 za obuku i 54 za testiranje) po govorniku od 10 govornika. Udruženja učesnika (17) razvila su vlastite VC sustave za svih 25 izvornog ciljanog govornika, a zatim su poslala generirane izjave za evaluaciju. Najmanje šest udruženja učesnika koristilo je DL-relaterane tehnike (LSTM, DNN) u svojim VC sistemima (vidjeti Tabelu 2. ), so the submitted utterances can certainly be considered deepfakes. Voice Conversion Challenge 2016 database 62 47 39 [ ]: Originalni skup podataka koji su stvorili organizatori izazova također je bio zasnovan na DAPS skupu podataka. Sadrži 116 izjava (81 za obuku i 35 za testiranje) po zvučniku od 12 zvučnika u dva različita zadatka (tzv. Hub i Spoke). Udruženje timova (23 ukupno, sve za Hub i 11 za Spoke) razvilo je vlastite VC sustave za svih 16 izvornih ciljanih parova zvučnika, a zatim je dostavilo generirane izjave za evaluaciju. U poređenju sa izazovom iz 2016. više udruženja je koristilo DL tehnike (npr. WaveNet, LSTM, DNN, CycleGAN, DRM – duboke relacijske modele, i ARBM – adaptivne ograničene Boltzmann strojeve) u svojim VC sistem 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 [ ]: 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 [ ]: This dataset contains bona fide and spoofed speech data for the logical access (LA) task. The challenge is still ongoing and we did not find a detailed paper on the dataset, so cannot include more details other than its size (7.8 GB after compression). Although we did not see details of the generative algorithms used to produce spoofed speech data, we believe similar DL-based algorithms were used like for the 2019 challenge. ASVspoof 2021 Challenge – Logical Access Database 14 [ ]: U 2021. godini izazov je uključivao eksplicitno definiranu stazu o deepfake-u, ali opis zadatka sugerira da su organizatori izazova razmotrili širu definiciju pojma "deepfake" gledajući na lažne ljudske slušalice umjesto na ASV (Automatic Speaker Verification) sisteme. 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 at the SenseHuman 2021 (3rd Workshop on Sensing, Understanding and Synthesizing Humans) , 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 , subjektivna procjena kvaliteta je neophodna za procjenu realnosti, realnosti i prirodnosti dubokih medija. Dok je na ovoj temi bilo vrlo ograničenog rada, 2020. godine, 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. Subjektivni kvalitativni rezultati prikazani u tabeli 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 fokusirajući se na više lica u istom videu, a KoDF fokusirajući se na korejske lica). Ovaj trend pokazuje da je duboko krivotvorena istraživačka zajednica značajno porasla u proteklih nekoliko godina tako da su uske teme počele da dobijaju pažnju i interes od nekih istraživača. 3 FFIW K 5 Deepfake-Related Challenges, Competitions & Benchmarks Ovaj odeljak pregledava inicijative usmjerene na unapređenje najsavremenije detekcije i proizvodnje sintetičkih ili manipuliranih medija (kao što su video, slika i audio) putem natjecanja ili izazova otvorenih javnosti, kao i tekuće referentne točke rješavanja specifičnih problema. 5.1 Detection of Manipulated Media The Deepfake Detection Challenge (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 u skladu sa ukupnim rezultatom log gubitka, kako je definisano u 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 solutions that go beyond analysing images and video. Considering context, provenance, and other signals may be the way to improve deepfake detection models 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 (također nazvan „kontramjere”), svaki koristeći različitu tehniku, stavljeni su na raspolaganje sudionicima sa svojim odgovarajućim EER metrikama. Baza podataka ASVspoof 2021 Speech Deepfake koja sadrži audio snimke s originalnim i lažnim izjavama je takođe dostupna. : faza napretka, faza evaluacije i faza posle evaluacije; nije jasno kako se timovi kreću iz jedne faze u drugu. ]. 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 standard. Despite the fact that only the 2021 ASVspoof competition contained a track explicitly related to deepfake, some data in the ASVspoof 2019 dataset (Logical Access task) used for the 2019 competition was generated using DL-based algorithms as mentioned in Section Očekujemo da se to odnosi i na ASVspoof 2021. dataset (zadatak logičkog pristupa). Natjecanje ASVspoof 2019. koristio je EER kao sekundarnu metriku; primarna metrika performansi koja se koristi bila je tandemna funkcija troškova otkrivanja (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 Uslovi poslovanja ( ) and cm( ) are, respectively, “ ”. 54 Speech DeepFake 55 56 13 EER reporting is deprecated 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 Za dodatne informacije o Eq. ( ), uključujući konstante 1 and 2, please refer to the ASVspoof 2019 evaluation plan [ ]. 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 standard. Figure 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 Kako bi se poboljšala efikasnost i efikasnost istraživanja, potrebno je utvrditi osnovne parametre koji se odnose na procjenu i procjenu rezultata istraživanja i procjenu rezultata istraživanja i procjenu rezultata istraživanja. )) 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 Izazov je uključivao dve faze: „razvojnu fazu” koja je počela u kolovozu 2020. godine omogućavajući 100 uspješnih podizvoda, i „zadnju fazu testiranja” koja je počela u oktobru 2020. godine omogućavajući 2 uspješna podnošenja do kraja meseca. ), 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 To je lažno. 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 , jedan od glavnih AI "jednoroga" u Kini). Cilj izazova je unapređenje najmodernijih tehnika u detekciji foto-realističke manipulacije slika i videa. Sudionici mogu koristiti veliki anotirani skup podataka o licu (tj. ForgeryNet dataset) koji je dobijen primenom niza tehnika za manipulaciju (15) i poremećaje (36) kako bi obučili svoja rješenja. Faze uključuju analizu krivotvorenih slika, analizu krivotvorenih videa, faze vremenske lokalizacije krivotvorenih videa i završnu fazu (tj. "privatni test") u kojoj će se učesnici testirati protiv nevidljivog skupa podataka. Koristiće se sledeće metrike. ]: 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 labels returned for multi-class classifiers. 74 75 25 t 5 K K K K 2020 CelebA-Spoof Face Anti-Spoofing izazov Bio je domaćin 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 . The 2021 challenge has the 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 [ ] i komercijalni softverski alati kao što su oni iz 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 is a biennial competition that has been running since 2016. The challenge and the corresponding workshop, hosted at the INTERSPEECH conference , podržava 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 ). For the second task “cross-lingual voice conversion”, participants had to develop 24 VC systems, also including male and female speakers, but uttering sentences in three languages (Finnish, German and Mandarin), based on the provided training dataset. Figure 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 [ ] and culture [ ].5.3 Generation and Detection of Manipulated Media 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 [ ], koji sadrži 6,229 video zapisa (590 stvarnih/originalnih video zapisa i 5,639 lažnih/manipuliranih video zapisa), u svrhe obuke. Kao takav, podnošenja u C-fazu sastoje se od skupova podataka izvučenih iz Celeb-DF (v2) koji uključuju nove pristupe za razmjenu lica kako bi se dobili rezultati evaluacije. Podnošenja u D-fazu sastoje se od modela detekcije/kodova kako bi se dobili rezultati evaluacije. Modeli podneseni za D-fazu su ocijenjeni protiv skupova podataka podnesenih za prethodnu C-fazu [ ]. metrike koje se koriste za evaluaciju 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. ( ), where is the number of valid submissions of detection methods in the last D-phase, the noise score ( 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 Kao što smo rekli u odjeljku , 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 [ ] definiran deepfake kao “ ”. 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 [ , U pitanju je , ], some considered only a narrow scope, e.g., authors of [ , , U pitanju je ] 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 [ , , ] čak je ograničio definiciju "deepfake" na tako uski opseg: 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 [ ]”, gde se tehnike zasnovane na AI-u koriste za stvaranje realističnih satelitskih slika. 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 koji je rekao VC tehnologiju (za govor dubokofake) “ ”. 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 Metode performansi 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 To je najopsežniji pregled skupova podataka povezanih sa deepfake-om do sada. 4 38 59 4 Some survey papers include datasets that are likely deepfakes, e.g., Verdoliva [ ] covered many general fake image datasets where the manipulated images were not generated by deep learning or even AI-based methods, and some surveys (e.g., [ ]) 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 , verujemo da je naša pokrivenost izazovima, natjecanjima i referentnim točkama povezanim sa deepfake-om u odjeljku 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 [ , U pitanju je ]. 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 Zbog problema sa kvalitetom mnogih skupova podataka koji se odnose na deepfake (diskutovano u odjeljku ), 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 The authors of some surveys identified some key challenges and future research directions for the deepfake community. 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 [ , U pitanju je , , , , , U pitanju je ]. 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 [ ] suggested that, for detection of deepfake videos, we need to consider video-level detection more, which can be considered fusion of detection results of all video frames. 38 66 45 Autori tri ankete, Lyu [ ] , Deshmukh and Wankhade [ ] and Younus and Hasan [ ], argued that better (higher-quality, more up-to-date, and more standard) deepfake datasets are needed to develop more effective deepfake detection methods. Lyu [ ] also suggested that we need to consider effects in training data and improve the evaluation of datasets. We agree with them on these points. 45 16 71 45 social media laundering Dao 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 Dva kineska istraživanja [ , ] 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 Postoje i druge ad hoc preporuke koje daju autori nekih anketa. Na primjer, 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 Brz rast sposobnosti manipulacije medijima ili stvaranja sintetičkih medija koji izgledaju realistično i prirodno otvorio je put za deepfake. U početku, ovaj članak je usvojio kritičan pristup za razmatranje različitih definicija pojma „deepfake“. U tom pogledu, ukazujemo na različite kontradiktorne definicije i pozivamo širu zajednicu da razmotri kako definisati novi termin koji ima dosljedniji opseg i značenje. Na primjer, zamjena „deepfake“ sa „deep synthesis“ može biti uključivija tako što će se prihvatiti pozitivne primjene tehnika deepfake, npr. u zabavi i za simulacijske svrhe. 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. Referencije [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 . Association for Computational Linguistics, 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 Computer Science 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 i 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, and Cristian Canton Ferrer. 2020. The DeepFake Detection Challenge (DFDC) Dataset. arXiv:2006.07397. https://arxiv.org/abs/2006.07397 [20] Nick Dufour and Andrew Gully. 2019. Contributing Data to Deepfake Detection Research. Google AI Blog. https://ai.googleblog.com/2019/09/contributing-data-to-deepfake-detectio Uslovi HTML [21] Ricard Durall, Margret Keuper, Franz-Josef Pfreundt, and Janis Keuper. 2019. Unmasking Deep-Fakes with Simple Features. 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. 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 i Christian Theobalt. 2021. Videoforensicshq: Detecting High-Quality Manipulated Face Videos. . IEEE, 1–6. Proceedings of the 2021 IEEE International Conference on Multimedia and Expo https://doi.or g/10.1109/ICME51207.2021.9428101 [24] Haiying Guan, Andrew Delgado, Yooyoung Lee, Amy N. Yates, Daniel Zhou, Timothee Kheyrkhah, i Jon Fiscus. 2021. 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, 4360–4369. Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition 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 do 2895 Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition https://doi.org/ 10.1109/CVPR42600.2020.00296 Uslovi korišćenja [28] Nal Kalchbrenner, Erich Elsen, Karen Simonyan, Seb Noury, Norman Casagrande, Edward Lockhart, Florian Stimberg, Aaron van den Oord, Sander Dieleman i Koray Kavukcuoglu. 2018. 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. Digital Art Exhibition to Showcase Creative Potential of AI. DIGIT News. 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, and Christian Theobalt. 2018. Deep Video Portraits. 37, 4, Članak 163 (2018), 14 stranica. ACM Transactions on Graphics https://doi.org/ 10.1145/3197517.3201283 [35] Pavel Koršunov i S'ebastien Marcel. 2019. procjena ranjivosti i otkrivanje Deepfake videa. u . IEEE, 1–6. Međunarodna konferencija o biometriji 2019 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 i 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, and Xiangwei Kong. 2021. Survey on Deepfakes and Detection Techniques. 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. Proceedings of the 2018 IEEE International Workshop on Information Forensics and Security 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, 3204–3213. Konferencija IEEE/CVF o računalnoj viziji i prepoznavanju uzoraka 2020. 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, and Kai Chen. 2020. A Survey of Audiovisual Deepfake Detection Techniques. 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. Plos jedan [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 pages. Međunarodna konferencija IEEE 2020 o radionicama Multimedijalne izložbe 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, Članak 7 (2021), 41 stranice. ACM Computing Survey https://doi.org/10.1145/3425780 [47] Gautham J. Mysore. 2015.Može li se automatski pretvoriti govor snimljen na uobičajenim potrošačkim uređajima u realnim okruženjima u profesionalni govor kvalitete proizvodnje? – Sastav podataka, uvid i izazovi. 22, 8 (2015), 1006–1010. IEEE Signal Processing Letters https://doi.org/10.1109/LSP.2014.2379648 [48] Jo ̃ao C. Neves, Ruben Tolosana, Ruben Vera-Rodriguez, Vasco Lopes, Hugo Proenca, i Julian Fierrez. 2020. GANprintR: Poboljšane krivotvorine i procjena statusa u detekciji manipulacije lica. 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 i Koray Kavukcuoglu. 2016. https://doi.org/10.48550/ARXIV.1609.03499 [51] Debajyoti Pal i Tuul Triyason. 2018. An Survey of Standardized Approaches to the Quality of Experience Evaluation for Video Services: An ITU Perspective. 2018, Article 1391724 (2018), 25 pages. Međunarodni časopis za digitalno multimedijalno emitovanje 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, and Weiming Zhang. 2020. DeepFaceLab: Integrated, Flexible and Extensible Face-swapping Framework. https://doi.org/10.48550/ARXIV.2005.05535 [54] Stanislav Pidhorskyi, Donald A. Adjeroh, and Gianfranco Doretto. 2020. Adversarial Latent Au-toencoders. In IEEE, 10 stranica Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition 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, and Matthias Nießner. 2019. FaceForensics++: Learning to Detect Manipulated Facial Images. In IEEE, 1 – 11. Međunarodna konferencija o računalnoj viziji 2019 https://doi.org/10.110 9/ICCV.2019.00009 Umetnost i veštačka inteligencija: Kako ‘deepfakes’ mogu pomoći u stvaranju autentičnih muzejskih iskustava! 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, i Nobukatsu Hojo. 2019. WaveCycleGAN2: Neuralni post-filter vremenskog domena za generaciju talasa govora. https://doi.org/10.48550/A RXIV.1904.02892 Izveštaj [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 do 38. Časopis za sajber sigurnost http://jcs.iie.ac.cn/xxaqxb/ch/reader/view abstract.aspx?file no=20200204&flag=1 [60] Tencent. 2020. veštačka inteligencija Bijela knjiga. https://tech.sina.com.cn/roll/2020-07-14/doc-iivhvpwx5201226.shtml [61] Justus Thies, Michael Zollh ̈ofe, i Matthias Niessner. 2019. odgođeno Neural rendering: Sinteza slike koristeći Neural teksture. 38 Članak 66 (2019), 12 stranica. ACM Transactions on Graphics https://doi.org/10.1145/3306346.3323035 [62] Tomoki Toda, Ling-Hui Chen, Daisuke Saito, Fernando Villavicencio, Mirjam Wester, Zhizheng Wu, and Junichi Yamagishi. 2016. The Voice Conversion Challenge 2016. In Međunarodno udruženje za komunikaciju govora, 1632–1636. Izveštaj 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/ Uslovi korišćenja – pdf [64] Ruben Tolosana, Ruben Vera-Rodriguez, Julian Fierrez, Aythami Morales, and Javier Ortega-Garcia. 2020. Deepfakes and beyond: A Survey of face manipulation and fake detection. 64 (2020), od 131 do 148 Information Fusion https://doi.org/10.1016/j.inffus.2020.06.014 [65] Xin Tong, Luona Wang, Xiaoqin Pan, and Jingya Wang. 2020. An Overview of Deepfake: The Sword of Damocles in AI. In IEEE, 265 – 273 Međunarodna konferencija o računalnom vidu, slikama i dubokom učenju 2020. https://doi.org/10.1109/CVIDL51233.2020.00-88 [66] Luisa Verdoliva. 2020. Media Forensics i DeepFakes: Pregled. 14, 5 (2020), 910–932. IEEE dnevnik odabranih tema u obradi signala 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 stranica Računalni govor & jezik https://doi.org/10.1016/j.csl.2020.101114 [68] Mirjam Wester, Zhizheng Wu i Junichi Yamagishi. 2016. Analiza rezultata evaluacije izazova za konverziju glasa 2016. Međunarodno udruženje za komunikaciju govora, 1637–1641. Konferencija Interspeech 2016 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 i Andreas Nautsch. 2019. ASVspoof 2019: Automatic Speaker Verification Spoofing and 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 i Tomoki Toda. 2020. Voice Conversion Challenge 2020 – Intra-lingual Semi-paralel i Cross-lingual Voice Conversion –. Međunarodno udruženje za komunikaciju govora, 80–98. Rezultati zajedničke radionice za izazov Blizzard Challenge i izazov za konverziju glasa 2020 https://doi.org/10.21437/VCC BC.2020-14 [71] Mohammed A. Younus and Taha M. Hasan. 2020. Abbreviated View of Deepfake Videos Detection Techniques. In IEEE, 115 – 120 Proceedings of the 2020 6th International Engineering Conference https://doi.org/10.1109/IEC49899.2020.9122916 [72] Guangtao Zhai and Xiongkuo Min. 2020. Perceptual Image Quality Assessment: A Survey. 63, Član 211301 (2020), 52 stranice. Znanost Kina Informacione nauke 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, 67–70. Proceedings of the 2020 IEEE 3rd International Conference on Computer and Communication Engineering Technology https://doi.org/10.1109/CCET5090 Uslovi korišćenja 1.2020.9213159 [74] Yuxuan Zhang, Huan Ling, Jun Gao, Kangxue Yin, Jean-Francois Lafleche, Adela Barriuso, Antonio Torralba i Sanja Fidler. 2021. DatasetGAN: Efikasna tvornica podataka s oznakama uz minimalni ljudski napor. . IEEE, 10140–10150. Konferencija IEEE/CVF 2021. o računalnoj viziji i prepoznavanju uzoraka 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, and Ziwei Liu. 2020. CelebA-Spoof: Large-Scale Face Anti-spoofing Dataset with Rich Annotations. In Škorpija 70-85 Proceedings of the 2020 European Conference on Computer Vision 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 i Zhanlong Hao. 2021. 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 – 352 Kartografija i geografske informacije nauka https://doi.org/10.1080/15230406.2021.1910075 [78] Peng Zhou, Xintong Han, Vlad I. Morariu, and Larry S. Davis. 2017. Two-Stream Neural Networks for Tampered Face Detection. In IEEE, od 1831. do 1839. Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops https://doi.org/10.1109/CVPRW.2017.229 [79] Tianfei Zhou, Wenguan Wang, Zhiyuan Liang, i Jianbing Shen. 2021. lice forenzički u divljini. u IEEE, 5774 do 5784 Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition https://doi.org/10.1109/CVPR46437.2021.00572 [80] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. 2017. Unpaired Image-to-Image Translation Using Cycle-Consistent Adversarial Networks. In IEEE, 2242 – 2251 Proceedings of the 2017 IEEE International Conference on Computer Vision https://doi.org/10.1109/ICCV.2 017.244 [81] Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma, i Yu-Gang Jiang. 2020. WildDeepfake: izazovni realni svjetski skup podataka za detekciju Deepfake. . ACM, 2382–2390. Proceedings of the 2020 28th ACM International Conference on Multimedia https://doi.org/10.1145/339417 1.3413769 Ovaj dokument je dostupan na arhivu pod licencom CC by 4.0 Deed (Attribution 4.0 International). Ovaj dokument je dostupan na arhivu pod licencom CC by 4.0 Deed (Attribution 4.0 International).