Com els investigadors mesuren, detecten i comparen la manipulació de la IA

Els autors: 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) Els autors: Enes Altuncu, ea483@kent.ac.uk (Universitat de Kent, Regne Unit) Virginia N. L. Franqueira, V.Franqueira@kent.ac.uk (Universitat de Kent, Regne Unit) Shujun Li, S.J.Li@kent.ac.uk (Universitat de Kent, Regne Unit) Abstract Abstracció Els recents avenços en la IA, especialment l'aprenentatge profund, han contribuït a un augment significatiu en la creació de nous mitjans sintètics realistes (vídeo, imatge i àudio) i la manipulació dels mitjans existents, el que ha portat a la creació del nou terme "deepfake". Basat tant en la literatura de recerca i els recursos en anglès com en xinès, aquest article proporciona una visió completa de deepfake, que cobreix múltiples aspectes importants d'aquest concepte emergent, incloent 1) diferents definicions, 2) mètriques de rendiment i estàndards comunament utilitzats, i 3) conjunts de dades relacionats amb deepfake, reptes, competicions i punts de referència. Deepfake, enquesta, definició, conjunts de dades, benchmarks, reptes, competicions, estàndards, mètriques de rendiment. Paraules clau 1 Introducció Els avanços recents en la IA i l'aprenentatge automàtic han augmentat la capacitat de produir mitjans més realistes, per exemple, vídeo, imatge i àudio. En particular, els mètodes d'aprenentatge profund d'avantguarda van permetre la generació de "deepfakes", mitjans manipulats o sintètics la realitat dels quals no són fàcilment reconeguts per l'ull humà. Tot i que el deepfake és un fenomen relativament nou (havia aparegut per primera vegada a finals de 2017), el seu creixement ha estat notable. El nombre de vídeos deepfake a la xarxa de parla anglesa va créixer de 7.964 (desembre de 2018) a 14.678 (juliol de 2019) a 85.047 (desembre de 2020), el que representa un augment del 968% de 2018 a 2020. 2 En aquest treball, revisem l'ecosistema de recerca relacionat amb la deepfake en diversos aspectes, incloent mètriques i estàndards de rendiment, conjunts de dades, reptes, competicions i punts de referència.A més, proporcionem una meta-revisió de 12 treballs d'enquesta relacionats amb la deepfake seleccionats que cobreixen diversos aspectes addicionals diferents dels esmentats de manera sistemàtica, com ara comparació de rendiment, reptes clau i recomanacions. Tot i ser un terme molt popular, hi ha una falta de consens sobre la definició de "deepfake" i la frontera entre deepfakes i no-deepfakes no és clara. Per a aquesta enquesta, adoptem un enfocament relativament més inclusiu per cobrir totes les formes de mitjans manipulats o sintètics que es consideren deepfakes en un sentit més ampli. També cobrim temes estretament relacionats, incloent biometria i forensia multimèdia, ja que els deepfakes sovint s'utilitzen per llançar atacs de presentació contra sistemes d'autenticació basats en biometria i la detecció de deepfakes es pot considerar part de la forensia multimèdia. Es dóna a continuació una discussió més detallada sobre les diferents definicions de "deepfake". 1.1 Definicions del terme Deepfake Com el seu nom indica, el terme “deepfake” deriva de la combinació de “deep” (referint-se a (DL)) i "fake". Normalment s'utilitza per referir-se a la manipulació dels mitjans existents (imatge, vídeo i / o àudio) o la generació de nous mitjans (sintètics) utilitzant enfocaments basats en DL. Les dades de deepfake més comunament discutides són les imatges facials falses, les falsificacions de paraules falses i els vídeos falsos que combinen les imatges falses i les falsificacions de paraules falses. Mentre que tenir "fake" en la paraula indica mitjans manipulats o sintetitzats, hi ha moltes aplicacions benignes de la tecnologia de deepfake, per exemple, per a l'entreteniment i les arts creatives. En aquest sentit, s'ha proposat un altre terme "síntesi profunda" com una alternativa de so més Aquest nou terme, però, no ha estat àmpliament adoptat. deep learning 60 A més de la manca d'una definició universal, com ja s'ha esmentat, la frontera entre els falsos profunds i els falsos no profunds no és realment un tall clar. En primer lloc, la detecció de deepfakes sovint segueix enfocaments molt similars a la detecció de falsos tradicionals generats sense utilitzar tècniques DL. Els mètodes avançats de detecció també han començat a aprofitar DL per millorar el seu rendiment, però no necessiten saber com es crea un mitjà objectiu (profund o no). En certa mesura, es podria argumentar que la detecció de deepfakes no implica el desenvolupament de mètodes específics de deepfakes (tot i que alguns investigadors opten per fer-ho), sinó un detector més robust i universal que pot manejar qualsevol (profund o no) mitjà fals. Això es pot veure en dos temes estretament relacionats: biometria i forensics multimèdia. Per a la biometria, hi ha una tendència a utilitzar tèc En segon lloc, es pot argumentar que la paraula "profund" en "deepfake" no es refereix necessàriament a l'ús de "aprenentatge profund", sinó a qualsevol tecnologia "profund" (és a dir, sofisticada) que crea un mitjà fals molt creïble. ] considerat deepfake com a manipulació audiovisual utilitzant "un espectre de sofisticació tècnica ... i tècniques". i , referint-se a “manipulació de baix nivell de mitjans audiovisuals creats amb programari (fàcilment) accessible [o sense programari] per accelerar, retardar, restablir o re-contextualitzar contingut”.Aquesta comprensió més àmplia de “deepfake” també ha estat adoptada pels legisladors per a noves legislacions que lluiten contra els deepfakes maliciosos. 9 Fal·làcia Cheapfake 2018 Malicious Deep Fake Prohibition Act: Prohibició de falsos profunds § 1041 (b) (2): « ” El terme "fals profund" significa un enregistrament audiovisual creat o alterat de manera que el enregistrament semblaria falsament a un observador raonable com un enregistrament autèntic del discurs o conducta real d'un individu. 2019 DEEP FAKES Act de responsabilitat 2: § 1041 (n).(3) « The term ‘deep fake’ means any video recording, motion-picture film, sound recording, electronic image, or photograph, or any technological representation of speech or conduct substantially derivative thereof— (A) que sembli representar autènticament qualsevol discurs o conducta d'una persona que no s'hagi implicat de fet en aquest discurs o conducta; i ” b) La producció de la qual depengués substancialment de mitjans tècnics, en comptes de la capacitat d'una altra persona per personalitzar-la físicament o verbalment. Com podem veure de les definicions legals anteriors de “deepfake”, l’ús de DL com a tecnologia no es menciona en absolut. 1.2 Escala i contribució Basant-nos en la discussió anterior sobre les definicions de deepfake, podem veure que no sempre és senzill o significatiu diferenciar deepfakes de no-deep fakes. A més, per al nostre enfocament en l'avaluació i comparació del rendiment, la frontera entre deepfakes i no-deep fakes és encara més borrosa. Això és perquè DL és només una forma especial (profunda) d'aprenentatge automàtic (ML), i com a resultat, els mètodes DL i no-deep ML comparteixen molts conceptes, mètriques i procediments comuns. Tot i que es pot entendre profundament en un sentit molt més ampli, en aquest treball tenim un enfocament prou estret per evitar cobrir massa temes. Per a les mètriques i estàndards, vam optar per incloure tots els mètodes comunament utilitzats per avaluar mètodes generals de ML i els específicament definits per avaluar mètodes de creació o detecció de falsos profundos. Per als conjunts de dades, reptes, concursos i punts de referència, vam considerar els relacionats amb els mitjans falsos que es cobreixen en els papers d'enquesta relacionats amb deepfake i els amb una menció explícita del terme "deepfake" o un terme comparable. Per a la meta-revisió, només vam considerar els treballs d'enquesta els autors dels quals es referien explícitament al terme "deepfakes" en les metadades (títol, abstracte i paraules clau). 2 Methodology Els treballs de recerca coberts en aquesta enquesta (és a dir, els treballs d'enquesta relacionats amb la falsificació profunda) van ser identificats a través de cerques sistemàtiques a les bases de dades científiques, Scopus i China Online Journals (COJ). Les següents consultes de cerca es van utilitzar per realitzar les cerques en Scopus i COJ, respectivament: 3 (deepfake* O deep-fake* O “deep fake*”) I (revisió O enquesta O revisió O sistemati* O SoK) (deepfake OR 深度偽造) I (综述 OR 进展) Les recerques van retornar 41 articles d'enquesta en anglès i 15 en xinès. D'aquests, vuit van ser publicats en anglès i quatre en xinès. Els reptes, els concursos i els criteris de referència relacionats amb Deepfake es van identificar a través de múltiples fonts: els treballs d'enquesta seleccionats, els treballs d'investigació de les col·leccions personals dels coautors, les cerques web de Google i la inspecció manual dels llocs web de les principals conferències relacionades amb la IA que es van celebrar el 2020 i el 2021 (on aquests reptes i concursos s'organitzen habitualment). , ICCV, CVPR, AAAI, ICML, ICLR, KDD, SIGIR, WWW, i molts altres. A més, es va compilar una llista exhaustiva de conjunts de dades basada en els treballs d'enquesta seleccionats i els reptes, concursos i punts de referència identificats. Es van identificar els estàndards rellevants principalment a través dels treballs d'investigació coberts en aquesta enquesta, el coneixement personal dels coautors i les cerques web de Google. Per a les mètriques de rendiment, es van cobrir els que s'utilitzen comunament sobre la base dels estàndards rellevants, els papers d'enquesta i els reptes, concursos i punts de referència identificats. 4 3 Mètriques i estàndards de rendiment relacionats amb Deepfake En aquesta enquesta, ens centrem en l'avaluació del rendiment i la comparació dels mètodes de generació i detecció de deepfake. Les mètriques utilitzades per a aquestes avaluacions de rendiment són el nucli de les nostres discussions. En aquesta secció, revisem les mètriques de rendiment que s'utilitzen comunament per avaluar els algoritmes de generació i detecció de deepfake. Tingueu en compte que totes les mètriques que es cobreixen en aquesta secció també s'utilitzen comunament per avaluar el rendiment de sistemes similars que no són per generar o detectar deepfakes. En l'últim paràgraf, també discutim breument com es cobreixen les mètriques de rendiment relacionades en normes formals.Per "normes formals", ens referim a normes definides seguint un procediment formal, sovint per un o més organismes de normalització establerts com l'Organització Internacional per a la Normalització (ISO). Comissió Electrotécnica Internacional (IEC) Tingueu en compte que considerem una àmplia gamma de documents definits com a estàndards per organismes de normalització, per exemple, la Unió Internacional de Telecomunicacions (UIT). Recomanacions i informes tècnics ISO (TRs). 5 6 7 3.1 La matriu de la confusió La detecció de deepfake és principalment un problema de classificació binària.Un classificador binari pren una entrada que és o i emet un valor binari que denota ser o . For example, a deepfake detection system will take a suspected image as the input that may be o i la producció o . De fet, positiu De fet, el negatiu Predicció positiva Predicció negativa Realment fals actually real Predicció falsa Predicció real Una eina fonamental utilitzada en l'avaluació d'un classificador binari és que resumeix l'èxit i el fracàs del model de classificació. Els valors i en l'altre eix són els dos La classificació és la (veritable positiu i veritable negatiu) quan els valors real i previst coincideixen. (fals positiu i fals negatiu) quan els valors reals i previstos no coincideixen. mostra la matriu de confusió per a un classificador de fals profund binari (detector). Les dues cèl·lules en verd, TP (el nombre de ) i TN (el nombre ), indiquen els resultats correctes de la predicció, i les dues cel·les en vermell, FN (el nombre de ) i FP (el nombre de ), indiquen dos tipus diferents d'errors en fer resultats de predicció incorrectes. confusion matrix Actualitat Predicció correcte / correcte / veritat fallida / incorrecta / falsa 1 true positives true negatives false negatives false positives Taula 1: Matriu de confusió per a un classificador binari per detectar deepfake. fake (predicted) real (predicted) fake (actual) TP FN real (actual) FP TN Fals (en realitat) El TP FN La realitat (actual) El FP TN 3.2 Precisió i recull Based on the four fundamental values introduced in Section , és a dir, TP, TN, FP i FN, definim dues mètriques de rendiment importants per a un classificador binari: i . 3.1 precision recall La precisió d'un classificador binari es defineix com la fracció de Les mostres entre tots els En la matriu de confusió, és la fracció de les mostres veritables en la primera columna. ) i De fet, positiu Prediccions positives 1 Quan la relació "natural" entre mostres positives i negatives és significativament diferent del conjunt de prova, sovint és útil ajustar el pes dels falsos positius, el que condueix a la (en anglès) que es defineix en Eq. ), on 0 és un pes determinat per la relació entre les mostres negatives i positives. weighted precision 2 «A» Reconeixement d'un classificador binari és la fracció de samples among the com es mostra a l'Eq. ( En la matriu de confusió, és la fracció de mostres veritables en la primera fila. Predicció positiva De fet, positiu 3 Considerem un exemple de classificador binari que prediu si una imatge d'una base de dades que conté imatges profundes i reals (autèntiques) és falsa o no. La precisió del classificador és la fracció d'imatges correctament classificades entre totes les imatges classificades com a profundes. 3.3 True and False Positive Rates Concentrant-nos en les mostres positives predites, també podem definir dues mètriques: (TPR), també anomenat (CDR), com la fracció de les mostres predites positives entre les mostres realment positives i (FPR), també anomenat (FAR), com la fracció de les mostres predites positives entre les mostres realment negatives, com es mostra a Eqs. ( ) i true positive rate correct detection rate false positive rate false alarm rate 4 ( En la matriu de confusió, TPR és la fracció de les mostres positives previstes en la primera fila i FPR és la fracció de les mostres positives previstes en la segona fila. (Equivalent a ) i el 5 recall 3 3.4 True and False Negative Rates Semblant a les taxes positives veritables i falses, podem definir dues altres taxes que se centren en els resultats predits negatius: (TNR) indicant la fracció de les mostres negatives previstes entre les mostres negatives efectives, i (FNR) indicant la fracció de les mostres predites negatives entre les mostres realment positives, com es mostra a Eqs. ( ) i ( ) i true negative rate false negative rate 6 7 3.5 Sensibilitat i especificitat En algunes aplicacions de classificadors binaris, especialment en biologia i medicina, el TPR i el TNR s'utilitzen més comunament, i sovint s'anomenen El TPR i (TNR) El focus d'aquests dos termes es troba en els dos tipus de correcció dels resultats predits. Aquests són menys utilitzats en la investigació relacionada amb la falsificació profunda, per tant, no ens referirem a ells en la resta d'aquest article. sensitivity specificity 3.6 Taxa d'error igual Concentrar-se en les taxes d'error significa que hem de considerar el FPR i el FNR. Aquestes dues taxes normalment conflicten entre si de manera que reduir una taxa normalment condueix a un augment en l'altra. Per tant, en comptes d'intentar reduir les dues taxes d'error al mateix temps, que normalment és impossible, la tasca més realista en aplicacions pràctiques és trobar el balanç correcte perquè tots dos estiguin per sota d'un llindar acceptable. En algunes aplicacions, com ara la biometria, les persones estan particularment interessades en establir l'anomenat (Sí) o La EER/CER no és necessàriament una bona mètrica per a algunes aplicacions, especialment quan els dos tipus d'errors són de diferents nivells d'importància, per exemple, per a la detecció de deepfakes crítics (per exemple, notícies falses que poden influir en com les persones fan votar) sovint podem tolerar més falsos positius (falses alarmes) que falsos negatius (alertes perdudes). equal error rate crossover error rate 3.7 Precisió i puntuació F A més del EER/CER, també hi ha altres mètriques que intenten reflectir ambdós tipus d'errors, per tal de donar una indicació més equilibrada del rendiment general d'un classificador binari. i (també anomenat Ambdues mètriques es poden definir basant-se en els quatre valors fonamentals (TP, TN, FP i FN). accuracy F-score F-measure La precisió d'un classificador binari es defineix com la fracció de les mostres (verdaderes positives i veritables negatives) entre el nombre total de mostres que han estat classificades, tal com es mostra a l'Eq. ) i Previsió correcta 8 La puntuació F d'un classificador binari és en realitat una família de mètriques.La seva forma general es pot descriure basant-se en un paràmetre as defined in Eq. ( ) i β 9 L'edició més àmpliament utilitzada de totes les puntuacions F és l'anomenada , que és efectivament la F-score amb = 1. més precisament, es defineix com es mostra a Eq. ( ) i F1-score β 10 3.8 Curva característica operativa del receptor i àrea sota curva Les corbes (ROC) s'utilitzen comunament per mesurar el rendiment dels classificadors binaris que produeixen una puntuació (o probabilitat) de predicció. Receiver operating characteristic Consider the following. Let ser el conjunt de totes les mostres de prova i deixar que les puntuacions de sortida ( ) (for all ∈ ) lie in the interval [ ] en la línia real. deixar ∈ [ ] ser un llindar de predicció per al model, i suposem que els classificadors funcionen de la següent manera per a tots ∈ : S f s s S a, b t a, b s S It is easy to see that, for = El , totes les mostres seran classificades com a positives, portant a FN = TN = 0 de manera que TPR = FPR = 1; mentre que per a = , 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 de to continuously, we can normally get a continuous curve that describes how the TPR and FPR values change from (0,0) to (1,1) on the 2D plane. This curve is the ROC curve of the binary classifier. t a t b a b t a b For a random classifier, assuming that (que ) es distribueix de manera uniforme en [ ] 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 Llista de pèrdues Una altra mètrica de rendiment àmpliament utilitzada per a classificadors binaris que pot retornar una puntuació de probabilitat per a l'etiqueta predictiva és . For a binary classification with a true label ∈ {0*,* 1} and an estimated probability = Pr( = 1), the log loss per sample is the negative log-likelihood of the classifier given the true label, defined as shown in Eq. ( ). log loss y p y 12 Given a testing set with samples, the log loss score of a binary classifier can be calculated using Eq. ( ), on is 1 if the -th sample is true and 0 if false, and ˆ is the predicted probability of = 1. n 13 yi i y i yi 3.10 Extension to Multi-class Classifiers All metrics that are defined based on the four basic values TP, TN, FP and FN can be easily extended to Per exemple, si el sistema està classificant animals (gats, gossos, cavalls, lleons, tigres, etc.), llavors una veritable predicció positiva d'una imatge a ser d'un gat, seria simultàniament veritable prediccions negatives per a les classes restants (gossos, cavalls, lleons, tigres, etc.). multi-class classification 3.11 Perceptual Quality Assessment (PQA) Metrics By definition, the main goal of deepfakes is to make it hard or impossible for human consumers (listeners or viewers) to distinguish fake media from real media. Therefore, when evaluating the quality of deepfake media, the quality perceived by human consumers of the media is key. This calls for subjective assessment of the perceptual quality of the deepfake media as the “gold standard”. The most widely used subjective perceptual quality assessment (PQA) metric for audio-visual signals is (MOS), which has been widely used by the signal processing and multimedia communication communities, including digital TV and other multimedia-related consumer applications. As its name implies, MOS is calculated by averaging the subjective scores given by a number of human judges, normally following a numerical scale between 1 and 5 or between 0 and 100. MOS has been used in some deepfake-related challenges (see Section ) i també per avaluar i comparar la qualitat (realitat/naturalitat) dels conjunts de dades profundament falsos (veure Secció ). mean opinion score 5.2 4.6 Com a mètrica general subjectiva de PQA, MOS ha estat estandarditzada per la ITU. . There are also ITU standards defining more specific subjective Video Quality Assessment (VQA) metrics and the standard procedures one should follow to conduct VQA user studies, e.g., ITU-T Recommendation P.910 “Subjective video quality assessment methods for multimedia applications” . Note that the ITU standards focus more on traditional perceptual quality, i.e., how good a signal looks or sounds, even if it looks or sounds not real (e.g., too smooth). On the other hand, for deepfakes, the focus is rather different because what matters is the realness and naturalness of the created media, i.e., how real and natural it looks or sounds, even if it is of low quality. To some extent, we can also consider realness and naturalness as a special aspect of perceptual quality. 8 9 One major problem of subjective PQA metrics like MOS is the need to recruit human judges and to have a well-controlled physical testing environment and protocol, which are not easy for many applications. To help reduce the efforts and costs of conducting PQA-related user studies, various objective PQA metrics have been proposed, where the term “objective” refers to the fact that such metrics are human-free, i.e., automatically calculated following a computational algorithm or process. Depending on whether a reference exists, such objective PQA metrics can be largely split into three categories: full-reference (FR) metrics (when the original “perfect-quality” signal is available as the reference), reduced-reference (RR) metrics (when some features of the original “perfect-quality” signal are available as the reference), and no-reference (NR) metrics (when the original signal is unavailable or such an original signal does not exist). For deepfakes, normally NR or RR metrics are more meaningful because the “fake” part of the word means that part of the whole data does not exist in the real world, hence a full reference cannot be obtained. RR metrics are still relevant because deepfakes are often produced for a target’s specific attributes (e.g., face and voice), where the reduced reference will be such attributes. NR metrics will be useful to estimate the realness and naturalness of a deepfake, simulating how a human judge would rate it in a controlled subjective PQA user study. PQA is a very active research area and many PQA metrics have been proposed, some of which have been widely used in real-world products and services, e.g., (MSE), (PSNR) and (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 Més sobre Normes Many of the basic performance metrics described in this section have been widely used by deepfake researchers as de facto standards, e.g., EER, log loss and MOS have been widely used in deepfake-related challenges (see Section ). 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 , the joint technical committee for standardising “information technology”. 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 In terms of performance metrics, two other ongoing work items of the ISO/IEC JTC 1/SC 42 that deserve attention are as follows: ISO/IEC DTS (Draft Technical Specification) 4213 “Information technology – Artificial Intelligence – Assessment of machine learning classification performance” 13 ISO/IEC AWI (Approved Work Item) TS (Technical Specifications) 5471 “Artificial intelligence – Quality evaluation guidelines for AI systems” 14 While the ISO/IEC JTC 1/SC 42 was created very recently, another standardisation subcommittee under ISO/IEC JTC1 has a much longer history of nearly 20 years: the ISO/IEC JTC 1/SC 37 that focuses on biometrics-related technology. This standardisation subcommittee is highly relevant for deepfake since deepfake faces can be used to spoof biometrics-based user authentication systems. In this context, the following three standards are of particular relevance: 15 : This standard covers general metrics about evaluating biometric systems. Two major metrics in this context are (FAR) 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) referring to the standard FNR, and (ACER) that is defined as the average of the APCER and the NPCER/BPCER. Such metrics have been used in biometrics-related challenges such as Face Anti-spoofing (Presentation Attack Detection) Challenges Quan s'utilitzen imatges o vídeos profundament falsos per falsificar un sistema biomètric, aquestes mètriques estandarditzades es tornaran rellevants. 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 Image Datasets [ ]: 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 [ ]: 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 . The generated.photos datasets have a good diversity: five age groups (infants, children, youth, adults, middle-aged), two genders (male and female), four ethnicities (white, black, Latino, Asian), four eye colours (brown, grey, blue, green), four hair colours (brown, black, blond, gray), three hair length (short, medium, long), facial expressions, three head poses (front facing, left facing, right facing), two emotions (joy and neutral), two face styles (natural, beautified). (According to a number of research papers we read, an earlier 100K-Faces dataset was released by generated.photos for academic research in 2018, which was used by many researchers. This dataset is not currently available any longer.) generated.photos datasets 22 23 24 [ ]: 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 i lleixes conjunt de dades (tres tipus d'objectes - dormitoris, cotxes i gats - per a aquest últim). Tingueu en compte que el nom "100K-Generated-Images" no era un correcte com els autors [ ] just used this to name a sub-folder of their Google Drive shared space, but it was used in one of the survey papers [ ]. 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, which contains 150,000 face images generated using an earlier version of GANprintR. iFakeFaceDB 48 Neus i al. FSRemovalDB [ ]: This dataset includes 40,000 images, half real and half deepfake. The images were collected from four sources: the CelebA-HQ dataset , the Flickr-Faces-HQ dataset , the 100K-Faces dataset (ja no està disponible, vegeu la descripció del conjunt de dades generated.photos), i . Faces-HQ 21 27 28 29 thisperson-doesnotexist.com [ ]: Aquest conjunt de dades inclou 625,537 imatges facials sintetitzades de 10.177 celebritats, amb 43 rics atributs sobre la cara, la il·luminació, l'entorn i els tipus d'espò. . 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 i FaceForensics++ ) and 240,336 fake ones in four main facial manipulation types (identity swap, expression swap, attribute manipulation, and entire synthesis). The images cover two genders (male and female), a wide age groups (the majority between 21 and 50 years old), and both low- and high-quality levels. Diverse Fake Face Dataset (DFFD) 11 31 32 33 4.2 Deepfake Video Datasets [ ]: 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 El [ ]: 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 El [ ]: 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 [ ]: Aquest conjunt de dades conté 10.000 vídeos de falsificació d'alta qualitat, amb anotacions a nivell de vídeo i de cara. El conjunt de dades se centra en un cas més desafiant per a la detecció de falsificació: cada vídeo involucra una a 15 persones, però només algunes (una minoria de) cares es manipulen. 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) [ La majoria de vídeos tenen una resolució de 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 que va començar el 2016 és una sèrie de reptes importants sobre VC. Els conjunts de dades llançats d'aquesta sèrie de reptes són molt diferents d'altres conjunts de dades de deepfake: les dades de deepfake no s'inclouen en el conjunt de dades original creat pels organitzadors de cada repte, sinó en les presentacions dels participants (que són retargeted/fake utterances produïdes pels sistemes de VC construïts pels participants). Els conjunts de dades del repte també inclouen els resultats d'avaluació (basats en l'escolta) de totes les presentacions. 38 [ ]: El conjunt de dades original creat pels organitzadors del repte va ser derivat del conjunt de dades DAPS (Device and Produced Speech) [ ]. It contains 216 utterances (162 for training and 54 for testing) per speaker from 10 speakers. Participating teams (17) developed their own VC systems for all 25 source-target speaker pairs, and then submitted generated utterances for evaluation. At least six participating teams used DL-related techniques (LSTM, DNN) in their VC systems (see Table 2 of the result analysis paper ), so the submitted utterances can certainly be considered deepfakes. Voice Conversion Challenge 2016 database 62 47 39 [ ]: The original dataset created by the challenge organisers was also based on the DAPS dataset. It contains 116 utterances (81 for training and 35 for testing) per speaker from 12 speakers in two different tasks (called Hub and Spoke). Participating teams (23 in total, all for Hub and 11 for Spoke) developed their own VC systems for all 16 source-target speaker pairs, and then submitted generated utterances for evaluation. Comparing with the 2016 challenge, more participating teams used DL-related techniques (e.g., WaveNet, LSTM, DNN, CycleGAN, DRM – deep relational models, and ARBM – adaptive restricted Boltzmann machines) in their VC systems. Voice Conversion Challenge 2018 database 44 [ ]: This dataset is based on the Effective Multilingual Interaction in Mobile Environments (EMIME) dataset , a bilingual (Finnish/English, German/English, and Mandarin/English) database. It contains 145 utterances (120 for training and 25 for testing) per speaker from 14 speakers for two different tasks (with 4 × 4 and 4 × 6 source-target speaker pairs, respectively). Participating teams (33 in total, out of which 31 for Task 1 and 28 for Task 2) developed their own VC systems for all source-target speaker pairs, and then submitted generated utterances for evaluation. Comparing with the 2018 challenge, DL-based VC systems were overwhelmingly used by almost all participating teams (WaveNet and WaveGAN among the most used DL-based building blocks). Voice Conversion Challenge 2020 database 70 40 A major set of deepfake speech datasets were created for the (Spoofing i contramesures automàtiques de verificació d'espectadors) (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 , una base de dades de parla anglesa multilingüe capturada de 107 parlants (46 homes i 61 dones). Es van considerar dos escenaris d'atac: accés lògic (LA) que involucra el discurs falsificat (sintètic o convertit) i accés físic (PA) que involucra atacs de replicació de gravacions de bona fe enregistrades anteriorment). Per al nostre propòsit en aquesta enquesta, l'escenari de LA és més rellevant. La part de LA del conjunt de dades inclou 12.483 pronunciaments de bona fe (real) i 108.978 pronunciaments falsificats. Algunes de les dades de parla falsificades per a l'escenari de LA es van produir utilitzant un model generatiu que involucra tècniques basades en DL com la mem , 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 [ ]: In 2021, the challenge included an explicitly defined track on deepfake, but the task description suggests that the organisers of the challenge considered a broader definition of the term “deepfake” by looking at spoofing human listeners rather than ASV (Automatic Speaker Verification) systems. The size of the dataset is 34.5 GB after compression. ASVspoof 2021 Challenge – Speech Deepfake Database 15 Possibly because of the long history and wide participation of the community in the ASVspoof challenges for creating the dedicated datasets, there are very few other deepfake audio/speech datasets. One such dataset was created by a group of researcher from Baidu Research [ ]. This dataset was created to demonstrate a proposed voice cloning method. It is relatively small, and contains 134 utterances, including 10 real ones, 120 cloned ones, and 4 manipulated ones. Another dataset was created by Google AI and Google News Initiative , but it was made part of the ASVspoof 2019 dataset. This dataset contains thousands of phrases spoken by 68 synthetic “voices” covering a variety of regional accents. 5 44 4.4 Hybrid Deepfake Datasets : These datasets were created by the DARPA Media Forensics (MediFor) Program for the 2020 OpenMFC . There are two GAN-generated deepfake datasets, one with more than 1,000 deepfake images and the other with over 100 deepfake videos. The datasets were made available to registered participants of the competition only. NIST OpenMFC (Open Media Forensics Challenge) Datasets 45 46 47 [ ]: This dataset is named as “a versatile benchmark for comprehensive forgery analysis”. It contains 2,896,062 images and 221,247 videos, including 1,457,861 fake images and 121,617 fake videos. The videos and images cover seven image-level and eight video-level manipulation approaches, 36 different types of perturbations and more mixed perturbations, and a large number of annotation labels (6.3 million classification labels, 2.9 million manipulated area annotations and 221,247 temporal forgery segment labels). The dataset is being used for supporting the Face Forgery Analysis Challenge 2021 at the SenseHuman 2021 (3rd Workshop on Sensing, Understanding and Synthesizing Humans) , co-localitzat a la conferència ICCV 2021 . 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 Com es va esmentar a la secció , subjective quality evaluation is necessary to evaluate the realness, realisticness, and naturalness of deepfake media. While there has been very limited work on this topic, in 2020, Jiang et al. [ ] conducted a user study on realness of deepfake videos. They recruited 100 professional participants (most of whom are computer vision researchers), who were asked to evaluate the realness of 30 randomly selected videos from 7 deepfake video datasets (DeeperForensics-1.0, UADFV, DeepFake-TIMIT, Celeb-DF, FaceForensics++, Deep Fake Detection, and DFDC). Participants were asked to respond to the statement “The video clip looks real.” and gave scores following a five-point Likert scale (1 – clearly disagree, 2 – weakly disagree, 3 – borderline, 4 – weakly agree, 5 – clearly agree). 4.7 27 Taula shows the results. Interestingly, we can see a huge difference between the realness levels of different datasets. What is probably quite surprising is that FaceForensics++, one of the most widely used deepfake datasets, has a very low MOS score and less than 9% of participants considered the 30 selected videos as real. 3 Table 3: Human-judged subjective quality (realness) of deepfake videos in 7 datasets. The MOS scores were not reported by , but calculated by us based on the raw data shown in Table 3 of [ ]. Jiang et al. 27 4.7 Discussion: Datasets Among all deepfake image and video datasets, a significant majority are about face images and videos. This is not surprising since face swapping, face attribution manipulation, and fully synthesised face images are among the hottest topics within deepfake research and real-world applications. We hope more non-face deepfake image and video datasets can be produced to support a broader range of research activities on deepfake. The subjective quality results shown in Table Indica que és important comprovar la realitat dels mitjans falsos profunds per donar suport a qualsevol avaluació de rendiment o comparació.Per assegurar que l'avaluació de la qualitat dels conjunts de dades sigui justa, transparent i fiable, cal definir els procediments estàndard i s'ha d'utilitzar un pool comú d'experts humans qualificats. 3 Many authors of deepfake-related datasets attempted to classify such datasets into different generations. Chronologically speaking, we could broadly split such datasets into two generations: before 2019 and since 2019. Typically, datasets created before 2019 are relatively less advanced and smaller, while those created after 2019 tend to be larger, more diverse (i.e., covering more attributes), and of higher quality (i.e., produced by more advanced generative models). This can also be seen from the data in Table , in which the top two datasets (DeeperForensics-1 and Celeb-DF) fall within the new generation (2020), while others belong to the old generation. In addition to the two generations, a newer generation has also emerged in 2021: a number of very recent datasets started focusing on more realistic deepfakes (i.e., in the wild) or more specified areas of deepfakes (e.g., 10 focusing on multiple faces in the same video, and KoDF focusing on Korean faces). This trend shows that the deepfake research community has grown significantly in the past few years so that narrower topics have also started gaining attention and interest from some researchers. 3 FFIU K 5 Deepfake-Related Challenges, Competitions & Benchmarks This section reviews initiatives aiming to advance the state-of-the-art of detection and generation of synthetic or manipulated media (such as video, image and audio) via competitions or challenges open to the public, and ongoing benchmarks tackling specific problems. 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 according to the overall log loss score, as defined in Eq. ( ). All top five ranked models (the winner had the lowest overall log loss) are available on GitHub. Results indicate how challenging the detection of deepfake is since the best accuracy was low and “ ”, according to Dolhansky et al. [ ]. Figure shows a screenshot of the leaderboard with the five finalists. The first top ranked model used MTCNN (Multi-tasked Cascaded Convolutional Network), the second used WS-DAN (Weakly Supervised Data Augumentation Network), and the third used the EfficientNetB7 architecture. Meta compiling the common themes observed in the winning models, they were: clever augmentations, architectures, and absence of forensics methods. Moving forward, they called for “ ”. 51 52 El salvatge 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 Taller Automatic Speaker Verification Spoofing and Countermeasures Challenge (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 on el cas d'ús previst és un adversari que intenta enganyar a un escoltant humà.La mètrica utilitzada per avaluar el rendiment de les solucions presentades (és a dir, els classificadors) és EER. (also called “countermeasures”), each using a different technique, were made available to participants with their corresponding EER metric values. The ASVspoof 2021 Speech Deepfake Database containing audio recordings with original and spoofed utterances has also been made available. The competition involves three phases : a progress phase, an evaluation phase and a post-evaluation phase; it is unclear how teams move from one phase to the next. More information about the 2021 competition is available in the published evaluation plan [ ]. 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 . We expect that this also holds for the ASVspoof 2021 dataset (Logical Access task). The ASVspoof 2019 competition used the EER as secondary metric; the primary performance metric used was the tandem detection cost function (t-DCF) [ ]. According to its evaluation plan [ ], t-DCF avalua el rendiment de tot el sistema tandem per la qual cosa " ”. It is calculated according to Eq. ( ), where cm ( ) and cm( ) are, respectively, “ ”. 54 Discurs 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 Per a més informació sobre EQ. ( ), including constants 1 i 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 a partir de 2017 sota el nom de DARPA MFC. El 2020, l'Institut Nacional d'Estàndards i Tecnologia (NIST) va com a nova plataforma d'avaluació, basada en les seves experiències anteriors amb la sèrie DARPA MFC, per fer que la participació sigui més convenient per a tots els investigadors. En OpenMFC 2020, es van incloure per primera vegada dues tasques relacionades amb el deepfake: Detecció de manipulació de imatges GAN (IGMD) i Detecció de manipulació de vídeo GAN (VGMD). Els organitzadors van proporcionar un conjunt de dades d'avaluació d'imatges per a la tasca IGMD, que contenia 1.000 imatges de més de 200 revistes d'imatges , i un conjunt de dades d'avaluació de vídeo per a la tasca VGMD, incloent-hi més de 100 vídeos de prova. used in the previous MFC challenges as development datasets. The challenge is composed of two main phases for development and evaluation, respectively, and a pre-challenge phase for quality control testing. For evaluation of submissions, AUC-ROC is used as the primary metric. Furthermore, CDR@FAR, where CDR refers to correct detection rate or TPR (Eq. ( )) and FAR refers to false alarm rate or FPR (Eq. ( )), també s'utilitza com a metàl·lic [ ]. The DeeperForensics Challenge 2020 is a deepfake face detection challenge held at the 2020 ECCV 67 68 OpenMFC 69 70 4 5 49 71 Taller SenseHuman . The challenge used the DeeperForensics1.0 dataset. 72 The organisers provided a hidden test dataset to better simulate real-world scenarios. The challenge involved two phases: the “development phase” that started in August 2020 allowing 100 successful sub-missions, and the “final test phase” that started in October 2020 allowing 2 successful submissions until the end of the month. The submissions were evaluated using the binary cross-entropy loss (BCELoss) metric, calculated according to Eq. ( ), where is the number of videos in the hidden test set, is the ground truth label of video (fake:1, real:0), and ( ) is the predicted probability that video is fake. 18 N yi i p yi i Results of the competition were discussed by Jiang et al. [ ]. The top solution used three models, i.e., EfficientNet-B0, EfficientNet-B1 and EfficientNet-B2, for classification. The second top used EfficientNet-B5 for both an image-based model and a video-based model. The third ranked solution used a 3D convolutional neural network (3DCNN). 73 26 The Face Forgery Analysis Challenge 2021 is a competition hosted at the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2021). It is organised by researchers from a number of organisations in China including universities and SenseTime Research (the research arm of SenseTime , one of the major AI “unicorns” in China). The challenge aims to advance the state-of-the-art in detection of photo-realistic manipulation of images and videos. Participants are able to use a large annotated face dataset (i.e., the ForgeryNet dataset) that was obtained by applying a number of techniques for manipulation (15) and perturbation (36) to train their solutions. The phases comprise of Forgery Image Analysis, Forgery Video Analysis, Forgery Video Temporal Localization phases, and the final phase (i.e., “private test”) where participants’ models will be tested against an unseen dataset. The following metrics will be used [ ]: AUC, average precision (AP) at some “temporal Intersection over Union” (AP@tIoU) compared to a threshold ∈ [0*. ,* 0*.*95) i el recull mitjà (AR) en (AR@ ) where is the top labels returned for multi-class classifiers. 74 75 25 t 5 K K K K The 2020 CelebA-Spoof Face Anti-Spoofing Challenge was hosted at the 16 European Conference on Computer Vision (ECCV 2020). The challenge ran between August and October 2020, and aimed to advance the state-of-the-art in detecting “ ” [ ]. The organisers made the face CelebA-Spoof dataset available for the competition containing rich annotation across a range of attributes. The competition only had one phase where participants submitted their solutions to be evaluated against a test dataset; the spoof class was considered as “positive” and the live class as “negative”. Metric TPR@FPR was used and collected at three points where the TPR when FPR = 10 4 determined the final ranking. The top three finalists (see Figure ) used deep learning models ResNet, EfficientNet-B7, and a novel architecture combining Central Difference Convolutional Networks (CDCN) and Dual Attention Network (DAN). The two top ranked solutions used different strategies to boost their models’ performance: a heuristic voting scheme was used by the top-ranked solution, and a weight-after-sorting strategy was used by the second ranked solution. 76 th whether a presented face is live or spoof 76 − 5 The 2021 CSIG Challenge is the second edition of a challenge organised by the China Society of Image and Graphics . 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 [ ] and commercial software tools such as those from iFlytek and IBM. The main VC tools used include Adaptive-VC and CycleGAN-VC [ ]. For both deepfake detection tasks, the performance metric used is log loss. 77 78 79 th 80 81 53 82 6 54 83 8 84 29 2020 China Artificial Intelligence was the second edition of a Chinese AI competition open for the general public to participate, organised by the municipal government of the City of Xiamen in China. In 2020, it had two sub-competitions, Multimedia Information Recognition Technology Competition and Language and Knowledge Technology Competition . The Multimedia Information Recognition Technology Competition included two tasks on deepfakes: one on deepfake video detection and one on deepfake audio/speech detection . The deepfake video detection task used 3,000 videos, and log loss was used as the sole performance metric. The deepfake audio/speech detection task used 20,000 audio samples (mostly in Chinese, and the remaining in English), and EER was used as the sole performance metric. For both tasks, the ratio between real and deepfake samples was 1:1. We did not find where to download the datasets used for the tasks nor a more detailed technical description of the datasets. For the deepfake video detection tasks, the top two winning teams (with an A prize) were from Netease (Hangzhou) Network Co., Ltd. and Beijing RealAI Technology Co., Ltd., followed by three other teams winning a B prize: Xiamen Fuyun Information Technology Co., Ltd.; Institute of Computing Technology, Chinese Academy of Sciences; and Wuhan Daqian Information Technology Co., Ltd. For the deepfake audio/speech task, there was no team winning an A prize, but one team winning a B prize: SpeakIn Technologies Co., Ltd. The final results of some teams were published, but some teams were allowed to hide their results. We did not find a detailed technical report summarising the results and explaining the work of the winning teams. 85 86 87 88 89 One of the B-prize winning team is from Beijing RealAI Technology Co., Ltd., a Chinese company active in deepfake-related R&D. 5.2 Generation of Manipulated Media El repte de la conversió de veu is a biennial competition that has been running since 2016. The challenge and the corresponding workshop, hosted at the INTERSPEECH conference , is supported by the SynSig (Speech Synthesis Special Interest Group) of the International Speech Communication Association (ISCA) . Its aim is to promote progress in voice conversion (VC) technology that can be applied to a number of positive and negative use cases, such as spoofing voice biometric systems. The 2020 challenge focused on speaker conversion, a sub-problem of VC, and included two tasks. For the first task “intra-lingual semi-parallel voice conversion”, participants had to develop 16 VC systems (speaker-pair combinations) including male and female speakers and English sentences, using the provided Voice Conversion Challenge 2020 database v1.0 for training (refer to Section ). 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) es troba en la seva primera edició, organitzada en la Conferència Conjunta Internacional sobre Biometria de 2021 (IJCB 2021). Els seus organitzadors són principalment de l'Institut d'Automatització de l'Acadèmia de Ciències de la Xina (CASIA). La idea de la competició era promoure un joc adversari entre agents que pressionaven per avançar tant en la creació com en la detecció de falsificacions profundes. Per aconseguir-ho, es va dissenyar un protocol de 6 etapes que intercalava tres fases de creació (fase C) i una fase de detecció (fase D), normalment una setmana a l'altre; les presentacions es van tancar a l'abril de 2021. ], containing 6,229 videos (590 real/original videos and 5,639 fake/manipulated videos), for training purposes. As such, submissions to a C-phase would consist of datasets extracted from Celeb-DF (v2) which included novel face-swap approaches to obtain evaluation results. Submissions to a D-phase would consist of detection models/codes to obtain evaluation results. The models submitted for a D-phase were evaluated against the datasets submitted for the previous C-phase [ ]. The metrics used for evaluation were: a detection score, used for evaluation of a D-phase, and a creation score, used for evaluation of a C-phase. The top three finalists for the detection phase employed CNN-based classifiers EfficientNet-B3, Efficientnet-B0 and EfficientNetV2. 96 40 52 97 Detecció de l'escorça ( ) metric captura la capacitat dels models de classificar correctament les imatges falses presentades a la fase C anterior contra un conjunt d'imatges reals en el conjunt de dades de prova CelebDF. ), 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 ( el soroll) penalitza les imatges sorolloses, les altres tres parts de l'equació es relacionen amb el següent : “ ”. 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 Una meta-revisió d'enquestes relacionades amb Deepfake Aquesta secció presenta una meta-revisió de 12 treballs d'enquesta relacionats amb deepfake seleccionats, dels quals vuit es publiquen en anglès. , , , – , , ] and four published in Chinese [ , , , ]. It covers the following aspects in a systematic manner: definitions and scope, performance metrics, datasets, challenges/competitions/benchmarks, performance comparison, key challenges and recommendations. 16 45 46 64 66 71 73 7 38 41 59 The meta-review aims at drawing some high-level insights for monitoring future development of deepfake-related technologies and their applications. 6.1 Definitions and Scope As we discussed in Section , among researchers, practitioners and law makers there is no universally accepted definition of “deepfake” as a term. This is also reflected in how the authors of the 12 survey papers considered this aspect. Most authors talked about the history of deepfakes and pointed out that the term reflects the combination of “deep learning” and “fake”, but some used a broader definition, e.g., Lyu [ ] defined deepfake as “ Alguns autors també es van referir a legislacions relacionades amb deepfake, però cap d'ells va assenyalar que les definicions en algunes d'aquestes legislacions són completament diferents de les definicions més tècniques que impliquen l'ús de l'aprenentatge profund. ] focused on speech forgery and did not explicitly highlight “deepfake”. 1.1 45 high quality fake videos and audios generated by AI algorithms 59 In terms of the scope, while some authors (correctly) considered all types of media that can be produced by deepfake-related techniques [ , , , ], some considered only a narrow scope, e.g., authors of [ , , , ] considered only videos, and only authors of [ , ] have considered images and videos. Another phenomenon we observed is that many authors focused more on face images and videos, and authors of three surveys [ , , ] even limited the definition of “deepfake” to such a narrow scope: 38 41 45 65 7 64 71 73 16 66 16 64 71 Deshmukh i Wankhade [16] la van definir com "una tecnologia que crea imatges falses o vídeos d'éssers humans objectius intercanviant els seus rostres [per] un altre personatge dient o fent coses que no són absolutament fetes per ells i els humans comencen a creure en tal fals que no sempre es reconeix amb l'ull humà quotidià ". Younus and Hasan [ ] considered deepfake as a technique allowing “ ”; and 71 any computer user to exchange the face of one person with another digitally in any video Tolosana et al. [ ] defined it as “ ”. 64 a deep learning based technique able to create fake videos by swapping the face of a person by the face of another person Such unnecessarily narrow definitions and scopes can lead to confusion and do not help exchanges between researchers and practitioners working on different types of deepfakes. We call on more researchers to accept a broader definition of “deepfake” so that highly realistic/natural media of any kind generated by a sophisticated automated method (often AI-based) is considered deepfake. Here, we provide two examples of such a broader definition: the image2image (or pixel2pixel) technique [ ] that allows the production of deepfake images and videos of any objects (e.g., the “horse2zebra” deepfake image shown in Figure ), and the the so-called “deepfake geography [ ]”, where AI-based techniques are used to generate realistic-looking satellite images. 80 7 77 Another important fact missed or not sufficiently discussed by authors of all the 12 surveys is that deepfake techniques can be used for positive applications, e.g., creative arts, entertainment and protecting online users’ privacy. We call for more researchers and practitioners to follow the proposal in the 2020 Tencent AI White Paper [ ] to start using the more neutral-sounding term “deep synthesis”. Accordingly,we can use different words for different types of data generated using “deep synthesis” techniques, e.g., “deep art”, “deep animation”, “deep music”, and “deepfake”. While authors of the 12 survey papers did not recognise the positive applications of “deepfake” technologies, some other researchers did, e.g., organisers of the Voice Conversion Challenge 2020 qui va dir que la tecnologia VC (per a la profunda falsificació del discurs) “ ”. 60 99 is useful in many applications, such as customizing audio book and avatar voices, dubbing, movie industry, teleconferencing, singing voice modification, voice restoration after surgery, and cloning of voices of historical persons 6.2 Performance Metrics Surprisingly, none of the 12 surveys have covered performance metrics explicitly. Some directly used performance metrics to explain and compare performance of covered deepfake generation and detection methods. The most used performance metrics include accuracy, ERR, and AUC. This may be explained by the page constraints of such survey papers, which did not allow the authors to extend their coverage significantly to cover performance metrics systematically. The subjective quality of deepfakes is an area least covered by the surveys, which seems related to an unbalanced coverage on deepfake generation and deepfake detection in terms of performance evaluation and comparison (the former much less than the latter). 6.3 Datasets Many of the 12 survey papers list a number of deepfake-related datasets, but none of them have coverage as complete as ours shown in Section . For instance, none of the surveys have covered the Voice Conversion Challenge 2016/2018/2020 datasets and the ASVspoof 2019/2021 datasets are covered briefly only in two surveys [ , ]. In addition, more recent deepfake datasets especially those released in 2021 are also not covered by any of the surveys. We believe that our Section is the most comprehensive review of deepfake-related datasets so far. 4 38 59 4 Some survey papers include datasets that are likely deepfakes, e.g., Verdoliva [ ] 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., [ ]) es va esmentar ASVspoof 2015 conjunt de dades, però no vam veure l'ús de l'aprenentatge profund per generar dades utilitzades en el conjunt de dades. 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 , we believe that our coverage of deepfake-related challenges, competitions and benchmarks in Section is also the most comprehensive so far. 5 5 4 5 6.5 Performance Comparison Most surveys have a good coverage of related methods for deepfake generation and detection, but only some explicitly covered performance comparison between different methods [ , , ]. 38 46 64 Among all the survey papers, Li et al. [ ] conducted the most comprehensive study on performance of different deepfake detection methods. In addition to showing the performance metrics of a number of deepfake detection methods in Table 3 of [ ], they also looked at general characteristics and issues of different types of deepfake detection methods, as shown in Table . Furthermore, they also looked at research on robustness of deepfake detection methods against adversarial samples, referring to some work that showed a lack of such robustness. 38 38 4 Due to quality issues of many deepfake-related datasets (discussed in Section ), we need to treat any performance metrics and comparison of different detection methods with caution. Without testing all methods on a sufficiently large, diverse and high-quality deepfake dataset, the performance comparison results can be misleading. This highlights the importance of having more challenges, competitions and benchmarks to encourage performance comparison on standard datasets and using consistent performance metrics. 4.6 6.6 Recomanacions i reptes 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 [ , , , , , , , , Tenint en compte la cursa d'armes entre la generació de deepfake i la detecció, aquesta direcció de recerca probablement seguirà sent el tema més calent en la investigació de deepfake. 7 16 38 41 45 59 65 66 71 Un parell d’anàlisis [ , ] 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 The authors of three surveys, 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 Tao et al. [ ] suggested that low-cost deepfake generation/detection should be considered as a future research direction. This is a valid recommendation since lightweight methods will allow less powerful computing devices (e.g., IoT devices) to benefit from such technologies. 59 Two Chinese surveys [ , ] also mentioned the need to have new deepfake-related legislations combating malicious use of deepfakes and the need to train end users such as journalists. This is likely an area where interdisciplinary research can grow. 38 41 There are also other ad-hoc recommendations given by the authors of some surveys. For example, Lyu [ ] argued that deepfake detection should be considered a (more complicated) multi-class, multi-label and local detection problem. Tolosana et al. [ ] va discutir direccions de recerca específiques per a diferents mètodes de generació de fals profund (síntesi facial, intercanvi d'identitat, manipulació d'atributs i intercanvi d'expressió). Liang et al. [ ] and Li et al. [ ] recommended more active defence mechanisms such as using digital watermarking and blockchain technologies to build trustworthy media frameworks against deepfakes. 45 64 41 38 7 Conclusion The rapid growth in the capability to manipulate media or create synthetic media which look realistic and natural paved the way for deepfakes. At first, this paper adopted a critical approach to look at different definitions of the term “deepfake”. In that regard, we point out the different contradicting definitions and call for the wider community to consider how to define a new term that has a more consistent scope and meaning. For instance, replacing “deepfake” by “deep synthesis” can be more inclusive by embracing positive applications of deepfake techniques, e.g., in entertainment and for simulation purposes. This paper provided a comprehensive overview of multiple aspects of the deepfake ecosystem drawing from the research literature and other online sources published in two languages: English and Chinese. It covers commonly used performance metrics and standards, related datasets, challenges, competitions and benchmarks. It also presents a meta-review of 12 selected deepfake-related survey papers published in 2020 and 2021, covering not only the above mentioned aspects, but also highlighting key challenges and recommendations. References [1] Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. 2018. MesoNet: A Compact Facial Video Forgery Detection Network. In . IEEE, 1–7. Proceedings of the 2018 IEEE International Workshop on Information Forensics and Security https://doi.org/10.1109/WIFS.2018.8630 761 [2] Henry Ajder, Giorgio Patrini, Francesco Cavalli, and Laurence Cullen. 2019. The State of Deepfakes: Landscape, Threats, and Impact. Deeptrace. , 27 pages. https://sensity.ai/reports/ [3] Zahid Akhtar and Tiago H. Falk. 2017. Audio-Visual Multimedia Quality Assessment: A Comprehensive Survey. 5 (2017), 21090–21117. IEEE Access https://doi.org/10.1109/ACCESS.2017.2750918 [4] Ahmed Ali and Steve Renals. 2018. Word Error Rate Estimation for Speech Recognition: e-WER. In . 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-desinformació-amenaça / [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 pàgines 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, and Cristian Canton Ferrer. 2020. The DeepFake Detection Challenge (DFDC) Dataset. 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 n.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, and Christian Theobalt. 2021. Videoforensicshq: Detecting High-Quality Manipulated Face Videos. In . 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, i Ziwei Liu. 2021. ForgeryNet: Un benchmark versàtil per a l'anàlisi integral de la falsificació. . 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, 2886–2895. Procediment de la Conferència IEEE/CVF de 2020 sobre visió informàtica i reconeixement de patrons https://doi.org/ 10.1109/CVPR42600.2020.00296 [28] Nal Kalchbrenner, Erich Elsen, Karen Simonyan, Seb Noury, Norman Casagrande, Edward Lockhart, Florian Stimberg, Aaron van den Oord, Sander Dieleman, and Koray Kavukcuoglu. 2018. Efficient Neural Audio Synthesis. https://doi.org/10.48550/ARXIV.1802.08435 [29] Takuhiro Kaneko i Hirokazu Kameoka. 2017. conversió de veu paral·lela lliure de dades utilitzant xarxes adversàries consistents en cicles. 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, Article 163 (2018), 14 pages. ACM Transactions on Graphics https://doi.org/ 10.1145/3197517.3201283 [35] Pavel Korshunov and S´ebastien Marcel. 2019. Vulnerability Assessment and Detection of Deepfake Videos. In . IEEE, 1–6. Proceedings of the 2019 International Conference on Biometrics https://doi.org/10.1109/ICB45273.2019.8987375 [36] Patrick Kwon, Jaeseong You, Gyuhyeon Nam, Sungwoo Park, and Gyeongsu Chae. 2021. KoDF: A Large-scale Korean DeepFake Detection Dataset. In IEEE, 10724-10733. Procediment de la Conferència Internacional IEEE/CVF de 2021 sobre visió informàtica https://doi.org/10.1109/ICCV 48922.2021.01057 [37] Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. 2020. FaceShifter: Towards High Fidelity And Occlusion Aware Face Swapping. arXiv:1912.13457. https://arxiv.org/abs/1912.13457 [38] Xurong Li, Shouling Ji, Chunming Wu, Zhenguang Liu, Shuiguang Deng, Peng Cheng, Min Yang, and Xiangwei Kong. 2021. Survey on Deepfakes and Detection Techniques. 32, 2 (2021), 496–518. Diari de programari 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 a 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, i Siwei Lyu. 2020. Celeb-DF: Un conjunt de dades desafiant a gran escala per a la forense de DeepFake. IEEE 3204-3213. Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition https://doi.org/10.1109/CVPR42600. 2020.00327 [41] Ruigang Liang, Peizhuo Lv, Yue Zhao, Peng Chen, Hao Xing, Yingjun Zhang, Jizhong Han, Ran He, Xianfeng Zhao, Ming Li i Kai Chen. 5, 2 (2020), 1 a 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 i Frank A. Russo. 2018. The Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS): Un conjunt dinàmic, multimodal d'expressions facials i vocals en anglès nord-americà. 13, 5 (2018), 35 pàgines PloS one [43] Chen-Chou Lo, Szu-Wei Fu, Wen-Chin Huang, Xin Wang, Junichi Yamagishi, Yu Tsao i Hsin-Min Wang. 2021. MOSNet: Avaluació objectiva basada en l'aprenentatge profund per a la conversió de veu. arXiv:1904.08352. https://arxiv.org/pdf/1904.08352.pdf Jaime Lorenzo-Trueba, Junichi Yamagishi, Tomoki Toda, Daisuke Saito, Fernando Villavicencio, Tomi Kinnunen i Zhenhua Ling. 2018. El repte de conversió de veu 2018: Promoure el desenvolupament de mètodes paral·lels i no paral·lels. . 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 pàgines Procediments de la Conferència Internacional de l'IEEE sobre Multimèdia Expo 2020 https://doi.org/10.1109/ICMEW46912.2020.9105991 [46] Yisroel Mirsky i Wenke Lee. 2021. la creació i detecció de deepfakes: una enquesta. 54, 1, Article 7 (2021), 41 pages. ACM Computing Investigació https://doi.org/10.1145/3425780 ¿Podem transformar automàticament el discurs gravat en dispositius de consum comuns en entorns del món real en un discurs de qualitat de producció professional? - Un conjunt de dades, insights i reptes. 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 Proen¸ca, and Julian Fierrez. 2020. GANprintR: Improved Fakes and Evaluation of the State of the Art in Face Manipulation Detection. 14, 5 (2020), 1038 i 1048. IEEE Journal of Selected Topics en Processament de Senyals 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. https://doi.org/10.48550/ARXIV.1609.03499 [51] Debajyoti Pal i Tuul Triyason. 2018. Una enquesta sobre enfocaments normalitzats cap a l'avaluació de la qualitat de l'experiència per als serveis de vídeo: una perspectiva ITU. 2018, article 1391724 (2018), 25 pàgines. International Journal of Digital Multimedia Broadcasting 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 i Wanyi Zhuang. https: //arxiv.org/abs/2106.01217 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 i Weiming Zhang. https://doi.org/10.48550/ARXIV.2005.05535 [54] Stanislav Pidhorskyi, Donald A. Adjeroh, i Gianfranco Doretto. 2020. Adversarial Latent Au-to-codificadors. En IEEE, 10 pàgines 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 i Matthias Nießner. 2018. FaceForensics: Un conjunt de dades de vídeo a gran escala per a la detecció de falsificació en cares humanes. https://doi.org/10.48550/ARXIV.1803.09179 [56] Andreas R ̈ossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, i Matthias Nießner. 2019. FaceForensics++: Aprendre a detectar imatges facials manipulades. IEEE, 1 a 11. Procediments de la Conferència Internacional sobre Visió Informàtica 2019 https://doi.org/10.110 9 / ICCV.2019.00009 [57] Anindya Sen. 2021. Art and Artificial Intelligence: How ‘Deepfakes’ Can Help Create Authentic Museum Experiences! Medium Blog. https://medium.com/art-world-zen/art-and-artificia l-intelligence-how-deepfakes-can-help-create-authentic-museum-experiences-65d8aa El 29C [58] Kou Tanaka, Hirokazu Kameoka, Takuhiro Kaneko, i Nobukatsu Hojo. 2019. WaveCycleGAN2: Post-filtre neural del domini del temps per a la generació d'ones de parla. https://doi.org/10.48550/A RXIV 1904.02892 [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 i 38. Revista de Ciberseguretat http://jcs.iie.ac.cn/xxaqxb/ch/reader/view abstract.aspx?file no=20200204&flag=1 [60] Tencent. 2020. Artificial Intelligence White Paper. https://tech.sina.com.cn/roll/2020-07-14/doc-iivhvpwx5201226.shtml [61] Justus Thies, Michael Zollh¨ofe, and Matthias Niessner. 2019. Deferred Neural Rendering: Image Synthesis using Neural Textures. 38, article 66 (2019), 12 pàgines. 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 i Junichi Yamagishi. 2016. . International Speech Communication Association, 1632–1636. Proceedings of Interspeech 2016 https://doi.org/10.21437/Interspeech.2016-1066 [63] Massimiliano Todisco, Xin Wang, Ville Vestman, Md Sahidullah, Hector Delgado, Andreas Nautsch, Junichi Yamagishi, Nicholas Evans, Tomi Kinnunen, and Kong Aik Lee. 2019. ASVspoof 2019: Future Horizons in Spoofed and Fake Audio Detection. arXiv:1904.05441. https://arxiv.org/ pdf/1904 05441.pdf [64] Ruben Tolosana, Ruben Vera-Rodriguez, Julian Fierrez, Aythami Morales, and Javier Ortega-Garcia. 2020. Deepfakes and beyond: A Survey of face manipulation and fake detection. 64 (2020), de 131 a 148. Informació de fusió https://doi.org/10.1016/j.inffus.2020.06.014 [65] Xin Tong, Luona Wang, Xiaoqin Pan, i Jingya Wang. 2020. Una visió general de Deepfake: L'espasa de Damocles en AI. . IEEE, 265–273. Proceedings of the 2020 International Conference on Computer Vision, Image and Deep Learning https://doi.org/10.1109/CVIDL51233.2020.00-88 [66] Luisa Verdoliva. 2020. Media Forensics and DeepFakes: An Overview. 14, 5 (2020), 910–932. IEEE Journal of Selected Topics en Processament de Senyals 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, Hirokazu Kameoka, Ingmar Steiner, Driss Matrouf, Jean-Fran¸cois Bonastre, Avashna Govender, Srikanth Ronanki, Jing-Xuan Zhang, and Zhen-Hua Ling. 2020. ASVspoof 2019: A Large-scale Public Database of Synthesized, Converted and Replayed Speech. 64 (2020), 27 pages. Computer Speech & Language https://doi.org/10.1016/j.csl.2020.101114 [68] Mirjam Wester, Zhizheng Wu, i Junichi Yamagishi. 2016. Anàlisi del desafiament de conversió de veu 2016 Resultats d'avaluació. en Associació Internacional de Comunicació del Discurs, 1637-1641. Proceedings of the Interspeech 2016 Conference https://doi.org/10.21437/Interspeech.201 6-1331 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, and Tomoki Toda. 2020. Voice Conversion Challenge 2020 – Intra-lingual Semi-parallel and Cross-lingual Voice Conversion –. In Associació Internacional de Comunicació de Discurs, 80–98. Procediments del taller conjunt per al repte Blizzard i el repte de conversió de veu 2020 https://doi.org/10.21437/VCC BC.2020-14 [71] Mohammed A. Younus i Taha M. Hasan. 2020. Visió abreujada de les tècniques de detecció de vídeos profundament falsos. IEEE, 115 a 120 Procediment de la 6a Conferència Internacional d'Enginyeria 2020 https://doi.org/10.1109/IEC49899.2020.9122916 [72] Guangtao Zhai i Xiongkuo Min. 2020 Avaluació de la qualitat de la imatge perceptiva: una enquesta. 63, Article 211301 (2020), 52 pàgines. Science China Information Sciences 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 a 70. Procediment de la III Conferència Internacional IEEE 2020 sobre Tecnologia de la Informàtica i la Comunicació https://doi.org/10.1109/CCET5090 1.2020.9213159 de la Comissió [74] Yuxuan Zhang, Huan Ling, Jun Gao, Kangxue Yin, Jean-Francois Lafleche, Adela Barriuso, Antonio Torralba, i Sanja Fidler. 2021. DatasetGAN: fàbrica de dades etiquetades eficients amb mínim esforç humà. . IEEE, 10140–10150. Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition https://doi.org/10.1109/CVPR46437.2021.01001 [75] ![](file:///C:/Users/user/AppData/Local/Temp/msohtmlclip1/01/clip_image051.gif)Yuanhan Zhang, ZhenFei Yin, Yidong Li, Guojun Yin, Junjie Yan, Jing Shao, i Ziwei Liu. 2020. CelebA-Spoof: Dataset Anti-spoofing de cara a gran escala amb anotacions riques. En Pujol, 70 i 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. https://arxiv.org/pdf/2102.12642.pdf [77] Bo Zhao, Shaozeng Zhang, Chunxue Xu, Yifan Sun, i Chengbin Deng. 2021. Deep Fake Geography? quan les dades geoespacials es troben amb la intel·ligència artificial. 48, 4 (2021), 338–352. Cartografia i ciència de la informació geogràfica https://doi.org/10.1080/15230406.2021.1910075 [78] Peng Zhou, Xintong Han, Vlad I. Morariu, i Larry S. Davis. 2017. xarxes neuronals de dos fluxos per a la detecció de cara tampered. IEEE, de 1831 a 1839. Procediments de la Conferència IEEE de 2017 sobre visió informàtica i tallers de reconeixement de patrons https://doi.org/10.1109/CVPRW.2017.229 [79] Tianfei Zhou, Wenguan Wang, Zhiyuan Liang, i Jianbing Shen. 2021. . IEEE, 5774–5784. Procediment de la Conferència IEEE/CVF de 2021 sobre visió informàtica i reconeixement de patrons https://doi.org/10.1109/CVPR46437.2021.00572 [80] Jun-Yan Zhu, Taesung Park, Phillip Isola, i Alexei A. Efros. 2017. traducció d'imatge a imatge sense parella utilitzant xarxes adversàries coherents en cicle. . IEEE, 2242–2251. Procediments de la Conferència Internacional IEEE de 2017 sobre visió informàtica https://doi.org/10.1109/ICCV.2 017.244 [81] Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma i Yu-Gang Jiang. 2020. WildDeepfake: Un desafiant conjunt de dades del món real per a la detecció de Deepfake. ACM, 2382 – 2390 Procediment de la 28a Conferència Internacional de Multimèdia de l'ACM 2020 https://doi.org/10.1145/339417 1.3413769 Aquest document està disponible en arxiu sota la llicència CC by 4.0 Deed (Attribution 4.0 International). Aquest document està disponible en arxiu sota la llicència CC by 4.0 Deed (Attribution 4.0 International).