TYPE Review

PUBLISHED 04 September 2024

DOI 10.3389/fdata.2024.1400024

Deepfake: definitions,
OPEN ACCESS performance metrics and
standards, datasets, and a
EDITED BY
Nikolaos Pitropakis,
Edinburgh Napier University, United Kingdom

REVIEWED BY
Riccardo Spolaor,
meta-review
Shandong University, China
Christos Chrysoulas,
Edinburgh Napier University, United Kingdom Enes Altuncu, Virginia N. L. Franqueira and Shujun Li*
*CORRESPONDENCE
Institute of Cyber Security for Society (iCSS) & School of Computing, University of Kent, Canterbury,
Shujun Li
United Kingdom
S.J.Li@kent.ac.uk

RECEIVED 12 March 2024

ACCEPTED 12 August 2024
PUBLISHED 04 September 2024
Recent advancements in AI, especially deep learning, have contributed to a
significant increase in the creation of new realistic-looking synthetic media
CITATION
Altuncu E, Franqueira VNL and Li S (2024) (video, image, and audio) and manipulation of existing media, which has led to
Deepfake: definitions, performance metrics the creation of the new term “deepfake.” Based on both the research literature
and standards, datasets, and a meta-review.
and resources in English, this paper gives a comprehensive overview of deepfake,
Front. Big Data 7:1400024.
doi: 10.3389/fdata.2024.1400024 covering multiple important aspects of this emerging concept, including (1)
COPYRIGHT
different definitions, (2) commonly used performance metrics and standards, and
© 2024 Altuncu, Franqueira and Li. This is an (3) deepfake-related datasets. In addition, the paper also reports a meta-review
open-access article distributed under the of 15 selected deepfake-related survey papers published since 2020, focusing
terms of the Creative Commons Attribution
License (CC BY). The use, distribution or
not only on the mentioned aspects but also on the analysis of key challenges
reproduction in other forums is permitted, and recommendations. We believe that this paper is the most comprehensive
provided the original author(s) and the review of deepfake in terms of the aspects covered.
copyright owner(s) are credited and that the
original publication in this journal is cited, in
accordance with accepted academic practice. KEYWORDS
No use, distribution or reproduction is
deepfake, survey, definition, datasets, standards, performance metrics
permitted which does not comply with these
terms.

1 Introduction
Recent advancements in AI and machine learning have increased the capability to
produce more realistic media, e.g., video, image, and audio. Especially, state-of-the-art deep
learning methods enabled the generation of “deepfakes,” manipulated or synthetic media
the realness of which are not easily recognisable by the human eye. Although deepfake is a
relatively new phenomenon (having first appeared at the end of 2017), its growth has been
remarkable. According to the 2019 and 2020 Deeptrace (now, Sensity) reports on the state
of deepfake (Ajder et al., 2019), the number of deepfake videos on the English-speaking
internet grew from 7,964 (December 2018) to 14,678 (July 2019) to 85,047 (December
2020), representing a 968% increase from 2018 to 2020. By 2024, the number of available
tools for deepfake generation has reached to over 10,000 (Sensity, 2024).
In this work, we review the existing deepfake-related research ecosystem in terms of
various aspects, including performance metrics, standards, and datasets. Furthermore,
we provide a meta-review of 15 selected deepfake-related survey papers which covers
several additional aspects other than the mentioned ones in a systematic manner, such as
performance comparison, key challenges, and recommendations.
Despite being a hugely popular term, there is a lack of consensus on the definition
of “deepfake” and the boundary between deepfakes and non-deepfakes is not clear cut. For
this survey, we adopt a relatively more inclusive approach to cover all forms of manipulated
or synthetic media that are considered deepfakes in a broader sense. We also cover closely

Frontiers in Big Data 01 frontiersin.org

Altuncu et al. 10.3389/fdata.2024.1400024

related topics including biometrics and multimedia forensics, • 2018 Malicious Deep Fake Prohibition Act1 :
since deepfakes are often used to launch presentation attacks §1041.(b).(2): “the term ‘deep fake’ means an audiovisual
against biometrics-based authentication systems and detection of record created or altered in a manner that the record would
deepfakes can be considered part of multimedia forensics. A falsely appear to a reasonable observer to be an authentic record
more detailed discussion on different definitions of “deepfake” is of the actual speech or conduct of an individual.”
given next. • 2019 DEEP FAKES Accountability Act2 :
§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
1.1 Definitions of the term deepfake speech or conduct substantially derivative thereof
(A) which appears to authentically depict any speech or
As its name implies, the term “deepfake” is derived from the conduct of a person who did not in fact engage in such speech
combination of “deep” [referring to deep learning (DL)] and “fake.” or conduct; and
It is normally used to refer to the manipulation of existing media (B) the production of which was substantially dependent
(image, video, and/or audio) or the generation of new (synthetic) upon technical means, rather than the ability of another person
media using DL-based approaches. The most commonly discussed to physically or verbally impersonate such person.”
deepfake data are fake face images, fake speech forgeries, and
fake videos that combine both fake images and fake speech As we can see from the above legal definitions of “deepfake,” the
forgeries. While having “fake” in the word indicates manipulated use of DL as a technology is not mentioned at all. The focus here is
or synthesised media, there are plenty of benign applications of the on “authenticity”, “impersonation” and (any) “technical means.”
deepfake technology, e.g., for entertainment and creative arts. With
this respect, another term “deep synthesis” has been proposed as a
more neutral-sounding alternative (Tencent, 2020). This new term,
however, has not been widely adopted. 1.2 Scope and contribution
In addition to the lack of a universal definition, as mentioned
already, the boundary between deepfakes and non-deep fakes is Based on the above discussion on definitions of deepfake,
actually not clear-cut. There are at least two important aspects we we can see it is not always straightforward or meaningful to
should consider, one on detection of and the other on creation of differentiate deepfakes from non-deep fakes. In addition, for our
deepfakes. focus on performance evaluation and comparison, the boundary
First, the detection of deepfakes often follows very similar between deepfakes and non-deep fakes is even more blurred. This
approaches to the detection of traditional fakes generated without is because DL is just a special (deeper) form of machine learning
using DL techniques. Advanced detection methods have also (ML), and as a result, DL and non-deep ML methods share many
started leveraging DL to improve their performance, but they common concepts, metrics and procedures.
do not necessarily need to know how a target media is created Despite the fact that deepfake may be understood in a much
(deep or not). To some extent, one could argue that detecting broader sense, in this work, we have a sufficiently narrower focus
deepfakes does not involve developing deepfake-specific methods to avoid covering too many topics. We, therefore, decided to define
(even though some researchers choose to do so), but a more robust the scope of this survey as follows:
and universal detector that can handle any (deep or not) fake
media. This can be seen in two closely related topics: biometrics and • For metrics and standards, we chose to include all commonly
multimedia forensics. For biometrics, there is a trend of using deep used ones for evaluating general ML methods and those
learning techniques to generate fake biometric signals (e.g., face specifically defined for evaluating deepfake creation or
images and videos) for biometric spoofing or presentation attacks. detection methods.
For multimedia forensics, deepfake-based forgeries have become a • For datasets, we considered those related to fake media
new threat to the traditional problem of “forgery detection.” For covered in the deepfake-related survey papers and those with
both topics, the detection of biometric spoofing and multimedia an explicit mention of the term “deepfake” or a comparable
forgeries have evolved to consider both deep and non-deep fakes. term.
Second, one may argue that the word “deep” in “deepfake” does • For the meta-review, we considered only survey papers whose
not necessarily refer to the use of “deep learning”, but any “deep” authors explicitly referred to the term “deepfakes” in the
(i.e., sophisticated) technology that creates a very believable fake metadata (title, abstract, and keywords).
media. For instance, Brady (2020) considered deepfake as audio-
visual manipulation using “a spectrum of technical sophistication ... In this paper, we aim to make the following contributions:
and techniques.” They also introduced two new terms, Shallowfake
and Cheapfake, referring to “low-level manipulation of audio-visual • We discuss existing definitions of the term “deepfake” and
media created with (easily) accessible software [or no software] propose a more inclusive definition.
to speed, slow, restage or re-contextualise content.” This broader
understanding of “deepfake” has also been adopted by lawmakers
for new legislations combating malicious deepfakes. For instance, 1 https://www.congress.gov/bill/115th-congress/senate-bill/3805
the following two United States acts define “deepfakes” as follows: 2 https://www.congress.gov/bill/116th-congress/house-bill/3230

Frontiers in Big Data 02 frontiersin.org

Altuncu et al. 10.3389/fdata.2024.1400024

• We present an overview of the available deepfake-related 3 Deepfake-related performance

standards and metrics for evaluating deepfake generation or metrics and standards
detection, which have been generally overlooked by previous
surveys. The covered metrics include general AI metrics as In this survey, we focus on performance evaluation and
well as several deepfake-specific metrics for objective and comparison of deepfake generation and detection methods. The
subjective evaluation. metrics used for such performance evaluations are at the core
• We comprehensively cover a wide range of deepfake datasets, of our discussions. In this section, we review the performance
considering different modalities—image, video, audio, and metrics that are commonly used to evaluate deepfake generation
text. We believe this paper offers the most comprehensive and detection algorithms. Note that all metrics covered in this
review of deepfake-related datasets so far. section are also commonly used for evaluating the performance of
• We provide a meta-review of 15 deepfake survey papers similar systems that are not for generating or detecting deepfakes.
to draw some high-level insights for monitoring the future Therefore, this section can be seen as a very brief tutorial on general
development of deepfake-related technologies and their performance metrics.
applications. In the last subsection, we also briefly discuss how the related
performance metrics are covered in formal standards. By “formal
standards,” we refer to standards defined following a formal
1.3 Paper organisation procedure, often by one or more established standardisation
bodies such as the International Organization for Standardization
The rest of the paper is as organised as follows. In Section 2, (ISO)3 and the International Electrotechnical Commission (IEC).4
we mention how we collected the survey papers covered in Note that we consider a broad range of documents defined
this paper. Then, Section 3 reviews existing deepfake-related to be standards by standardisation bodies, e.g., International
performance metrics and standards, followed by Section 4 covering Telecommunication Union (ITU)5 recommendations and ISO
deepfake datasets. In Section 5, we provide a meta-review technical reports (TRs).
of the survey papers collected. Finally, the paper concludes
with Section 6.
3.1 The confusion matrix
2 Methodology Deepfake detection is primarily a binary classification problem.
A binary classifier takes an input that is actually positive or actually
Research papers covered in this survey (i.e., the deepfake- negative and outputs a binary value denoting it to be predicted
related survey papers) were identified via systematic searches on the positive or predicted negative. For example, a deepfake detection
Scopus scientific database. The following search query was used to system will take a suspected image as the input that may be actually
perform the searches on Scopus: fake or actually real and output predicted fake or predicted real.
A fundamental tool used in evaluating a binary classifier is the
(deepfake* OR deep-fake* OR “deep fake*”) AND (review OR
confusion matrix that summarises the success and failure of the
survey OR overview OR systemati* OR SoK)
classification model. On one axis are the two actual values and
on the other axis are the two predicted values. The classification
The searches returned 117 survey papers in English, published
is successful/correct/true (true positive and true negative) when the
between 2020 and 2024 (inclusive). Out of these papers, 15
actual and the predicted values match. It is failed/incorrect/false
papers were selected for consideration in the meta-review.
(false positive and false negative) when the actual and predicted
During the selection process, all the papers were carefully
values do not match. Table 1 shows the confusion matrix for a
reviewed, and only the ones having a substantial comparative
binary deepfake classifier (detector). The two cells in green, TP (the
angle, e.g., those with performance comparison and/or covering
number of true positives) and TN (the number of true negatives),
different datasets, tools, challenges, competitions, metrics, etc.,
indicate correct prediction results, and the two cells in red, FN
were included. Furthermore, for the papers with similar coverage,
(the number of false negatives), and FP (the number of false
those published in more decent venues (e.g., higher-ranked
positives), indicate two different types of errors when making
journals or more well-known conferences) and/or more cited by
incorrect prediction results.
other studies were preferred. Finally, we ensured that the final
set of papers cover publications from each year between 2020
and 2024.
Deepfake-related datasets were compiled based on the
3.2 Precision and recall
selected survey papers and identified deepfake-related challenges,
Based on the four fundamental values introduced in
competitions, and benchmarks. Relevant standards were identified
Section 3.1, i.e., TP, TN, FP, and FN, we define two important
mainly via research papers covered in this survey, the co-authors’
personal knowledge, and Google Web searches. For performance
metrics, we covered those commonly used based on relevant 3 https://www.iso.org/
standards, the survey papers, and the identified challenges, 4 https://www.iec.ch/
competitions, and benchmarks. 5 https://www.itu.int/

Frontiers in Big Data 03 frontiersin.org

Altuncu et al. 10.3389/fdata.2024.1400024

TABLE 1 Confusion matrix for a binary classifier for detecting deepfake.

3.4 True and false negative rates
Fake (predicted) Real (predicted)
Similar to true and false positive rates, we can define two other
Fake (actual) TP FN rates focusing on negative predicted results: true negative rate
Real (actual) FP TN (TNR) indicating the fraction of the predicted negative samples
among the actually negative samples, and false negative rate (FNR)
indicating the fraction of the predicted negative samples among the
performance metrics for a binary classifier—precision actually positive samples, as shown in Equations 6, 7.
and recall.
Precision of a binary classifier is defined as the fraction of TN
TNR = (6)
actually positive samples among all the predicted positives. In the TN + FP
confusion matrix, it is the fraction of true samples in the first
column. It can be formally defined as Equation 1.
FN
TP FNR = (7)
precision = (1) FN + TP
TP + FP
When the “natural” ratio between positive and negative samples
is significantly different from the test set, it is often useful to 3.5 Sensitivity and specificity
adjust the weight of the false positives, which leads to the weighted
precision (wP) defined in Equation 2, where α > 0 is a weight In some applications of binary classifiers, especially in biology
determined by the ratio between the negative and positive samples. and medicine, the TPR and the TNR are more commonly used, and
TP they are often called sensitivity (TPR) and specificity (TNR). The
wP = (2) focus of these two terms is on the two types of correctness of the
TP + αFP
predicted results. These are less used in deepfake-related research,
Recall of a binary classifier is the fraction of predicted hence, we will not refer to them in the remainder of this paper.
positive samples among the actually positive samples, as shown
in Equation 3. In the confusion matrix, it is the fraction of true
samples in the first row.
3.6 Equal error rate
TP
recall = (3)
TP + FN Focusing on error rates means that we need to consider the
Let us consider an example binary classifier that predicts if FPR and the FNR. These two rates normally conflict with each
an image from a database containing both deepfake and real other so that reducing one rate normally leads to an increase in the
(authentic) images is fake or not. Precision of the classifier is the other. Therefore, rather than trying to reduce both error rates at the
fraction of correctly classified images among all images classified same time, which is normally impossible, the more realistic task in
as deepfake. On the other hand, recall is the fraction of deepfake practical applications is to find the right balance so that they are
images identified by the classifier, among all deepfake images in the both below an acceptable threshold.
database. In some applications, such as biometrics, people are particularly
interested in establishing the so-called equal error rate (EER) or
crossover error rate (CER), the point where the FPR and the
3.3 True and false positive rates FNR are equal. The EER/CER is not necessarily a good metric for
some applications, especially when the two types of errors are of
Focusing on predicted positive samples, we can also define two different levels of importance, e.g., for detecting critical deepfakes
metrics: true positive rate (TPR), also called correct detection rate (e.g., fake news that can influence how people cast their votes)
(CDR), as the fraction of the predicted positive samples among we can often tolerate more false positives (false alarms) than false
the actually positive samples and false positive rate (FPR), also negatives (missed alarms).
called false alarm rate (FAR), as the fraction of the predicted
positive samples among the actually negative samples, as shown
in Equations 4, 5. In the confusion matrix, TPR is the fraction of 3.7 Accuracy and F-score
predicted positive samples in the first row and FPR is the fraction
of predicted positive samples in the second row. Note that TPR is In addition to the EER/CER, there are also other metrics
basically a different name for recall (Equation 3). that try to reflect both types of errors, in order to give a more
balanced indication of the overall performance of a binary classifier.
TP The two most commonly used are accuracy and F-score (also
TPR = (4)
TP + FN called F-measure). Both metrics can be defined based on the four
fundamental values (TP, TN, FP, and FN).
FP Accuracy of a binary classifier is defined as the fraction of
FPR = (5)
FP + TN correctly predicted samples (true positives and true negatives)

Frontiers in Big Data 04 frontiersin.org

Altuncu et al. 10.3389/fdata.2024.1400024

among the total number of samples that have been classified, as

shown in Equation 8.

TP + TN
accuracy = (8)
TP + TN + FP + FN
The F-score of a binary classifier is actually a family of metrics.
Its general form can be described based on a parameter β as defined
in Equation 9.

precision · recall
Fβ = (1 + β 2 ) · (9)
β2 · precision + recall
The most widely used edition of all F-scores is the so-called F1-
score, which is effectively the F-score with β = 1. More precisely,
it is defined as shown in Equation 10.

precision · recall 2TP

F1 = 2 · = (10)
precision + recall 2TP + FP + FN FIGURE 1
A representative ROC curve showing how TPR and FPR change w.r.t.
the (hidden) threshold t. The area under the (ROC) curve (AUC) is
shown in grey.
3.8 Receiver operating characteristic curve
and area under curve

Receiver operating characteristic (ROC) curves are loss. For a binary classification with a true label y ∈ {0, 1} and an
commonly used to measure the performance of binary classifiers estimated probability p = Pr(y = 1), the log loss per sample is the
that output a score (or probability) of prediction. negative log-likelihood of the classifier given the true label, defined
Consider the following. Let S be the set of all test samples and as shown in Equation 12.
let the output scores f (s) (for all s ∈ S) lie in the interval [a, b] on
the real line. Let t ∈ [a, b] be a prediction threshold for the model, Llog (y, p) = −(y log(p) + (1 − y) log(1 − p)) (12)
and assume that the classifiers works as follows for all s ∈ S:
( Given a testing set with n samples, the log loss score of a binary
positive, if f (s) ≥ t, and classifier can be calculated using Equation 13, where yi is 1 if the
class(s) = (11)
negative, otherwise. i-th sample is true and 0 if false, and ŷi is the predicted probability
of yi = 1.
Considering Equation 11, it is easy to see that, for t = a, all the
samples will be classified as positive, leading to FN = TN = 0 n
1X
so TPR = FPR = 1; while for t = b, all the samples will be LL = − [yi log(ŷi ) + (1 − yi ) log(1 − ŷi )] (13)
n
classified as negative, leading to FP = TP = 0 so TPR = FPR = 0. i=1
For other threshold values between a and b, the values of TPR
and FPR will normally be between 0 and 1. By changing t from
a to b continuously, we can normally get a continuous curve that 3.10 Extension to multi-class classifiers
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. All metrics that are defined based on the four basic values TP,
For a random classifier, assuming that f (s) distributes uniformly TN, FP, and FN can be easily extended to multi-class classification
on [a, b] for the test set, we can mathematically derive its ROC curve by considering the prediction to be true or false individually with
being the TPR = FPR line, whose area under the ROC curve (AUC) respect to each class. For example, if the system is classifying
is 0.5. For a binary classifier that performs better than a random animals (cats, dogs, horses, lions, tigers, etc.), then a true positive
predictor, we can also mathematically prove that its AUC is always prediction of an image to be of a cat, would simultaneously be true
higher than 0.5, with 1 being the best possible value. Note that no negative predictions for the remaining classes (dogs, horses, lions,
binary classifier can have an AUC below 0.5, since one can simply tigers, etc.). If an image of a cat is incorrectly predicted to be that
flip the prediction result to get a better predictor with an AUC of a dog, it would be a false negative with respect to a cat, a false
of 1 − AUC. The relationship between the ROC and the AUC is positive with respect to a dog, and a true negative with respect to all
graphically illustrated in Figure 1. other classes.

3.9 Log loss 3.11 Deception success rate

Another widely used performance metric for binary classifiers DSR aims to measure to what extent a deepfake detection
that can return a probability score for the predicted label is log model can be fooled by the generated deepfakes. It was used for

Frontiers in Big Data 05 frontiersin.org

Altuncu et al. 10.3389/fdata.2024.1400024

the evaluation of the Audio Deep Synthesis Detection (ADD) 2022 “perfect-quality” signal is available as the reference), reduced-
Challenge (Yi et al., 2022a) and defined as follows: reference (RR) metrics (when some features of the original “perfect-
quality” signal are available as the reference), and no-reference
W (NR) metrics (when the original signal is unavailable or such an
DSR = (14)
A∗N original signal does not exist). For deepfakes, normally NR or
In Equation 14, W refers to the total number of incorrect RR metrics are more meaningful because the “fake” part of the
detection samples by all the detection models under the condition word means that part of the whole data does not exist in the real
of achieving each individual EER performance, A is the count of all world, hence a full reference cannot be obtained. RR metrics are
the evaluation samples, and N is the number of detection models. 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
3.12 Perceptual quality assessment metrics
judge would rate it in a controlled subjective PQA user study.
PQA is a very active research area and many PQA metrics
By definition, the main goal of deepfakes is to make it hard
have been proposed, some of which have been widely used in real-
or impossible for human consumers (listeners or viewers) to
world products and services, e.g., mean squared error (MSE), peak
distinguish fake media from real media. Therefore, when evaluating
signal-to-noise ratio (PSNR), and structural similarity index
the quality of deepfake media, the quality perceived by human
measure (SSIM) for FR PQA of digital images and videos defined
consumers of the media is key. This calls for a subjective assessment
as in Equations 15–17, respectively, where X = {xi }ni is the
of the perceptual quality of the deepfake media as the “gold
reference (the original signal), Y = {yi }ni is the signal whose visual
standard.” The most widely used subjective perceptual quality
quality is assessed, n is the number of pixels in X and Y, L is the
assessment (PQA) metric for audio-visual signals is mean opinion
maximum possible pixel value of X and Y (e.g., 255 for 8-bit grey-
score (MOS), which has been widely used by the signal processing
scale images), c1 = (k1 L)2 and c2 = (k2 L)2 ) are two stabilising
and multimedia communication communities, including digital
parameters (k1 = 0.01 and k2 = 0.03 by default). For more about
TV and other multimedia-related consumer applications. As its
PQA metrics for different types of multimedia signals, we refer
name implies, MOS is calculated by averaging the subjective
readers to some relevant surveys (Akhtar and Falk, 2017; Pal and
scores given by a number of human judges, normally following
Triyason, 2018; Zhai and Min, 2020).
a numerical scale between 1 and 5 or between 0 and 100. MOS
has been used in some deepfake-related challenges and also for n
X
evaluating and comparing the quality (realness/naturalness) of MSE(X, Y) = (yi − xi ) (15)
deepfake datasets (see Section 4.7). i=1
As a general subjective PQA metric, MOS has been
standardised by the ITU.6 There are also ITU standards defining
L2
 
more specific subjective Video Quality Assessment (VQA) metrics PSNR(X, Y) = 10 log10 (16)
MSE
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.”7 (2µx µy + c1 )(2σxy + c2 )
SSIM(X, Y) = (17)
Note that the ITU standards focus more on traditional perceptual (µ2x + µ2y + c1 )(σx2 + σy2 + c2 )
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 3.13 Multimodal alignment assessment
realness and naturalness of the created media, i.e., how real and metrics
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 With the increasing amount of research on leveraging
aspect of perceptual quality. multimodalities in deepfake generation (e.g., text-to-image, text-
One major problem of subjective PQA metrics like MOS is the to-video, and audio-to-video), new evaluation metrics became
need to recruit human judges and to have a well-controlled physical needed to assess to what extent the covered modalities are
testing environment and protocol, which are not easy for many aligned. The development of such metrics is challenging as
applications. To help reduce the efforts and costs of conducting they require a comprehensive and fine-grained cross-modal
PQA-related user studies, various objective PQA metrics have been understanding (Huang et al., 2023).
proposed, where the term “objective” refers to the fact that such While MAA metrics have been used to assess different tasks
metrics are human-free, i.e., automatically calculated following (e.g., image-caption alignment), they can also be considered an
a computational algorithm or process. Depending on whether a alternative to PQA metrics for deepfake evaluation. For instance,
reference exists, such objective PQA metrics can be largely split with the introduction of state-of-the-art text-to-image models, such
into three categories: full-reference (FR) metrics (when the original as DALL-E,8 which enables easier generation of deepfake images
from natural language descriptions, text-to-image alignment
6 https://www.itu.int/rec/T-REC-P.800.1-201607-I/en
7 https://www.itu.int/rec/T-REC-P.910-200804-I/en 8 https://openai.com/index/dall-e/

Frontiers in Big Data 06 frontiersin.org

Altuncu et al. 10.3389/fdata.2024.1400024

metrics can be utilised to assess whether the generated image In terms of performance metrics, two other ongoing work items
aligns the provided description. With this respect, several metrics of the ISO/IEC JTC 1/SC 42 that deserve attention are as follows:
based on vision-language models, which can simultaneously learn
from images/videos and texts [e.g., BLIP (Li J. et al., 2022) and • ISO/IEC DTS (Draft Technical Specification) 4,213
CLIP (Radford et al., 2021)], have been proposed. Some examples of “Information technology—Artificial Intelligence—
such metrics include ClipScore (Hessel et al., 2021), disentangled Assessment of machine learning classification performance”12
BLIP-VQA (Huang et al., 2023), and BLIP-CLIP (Chefer et al., • ISO/IEC AWI (Approved Work Item) TS (Technical
2023) for text-to-image generation, as well as X-CLIP (Ni et al., Specifications) 5,471 “Artificial intelligence—Quality
2022) for text-to-video generation. evaluation guidelines for AI systems”13
In addition to vision-language models, MAA metrics have also
made use of multimodal large language models by generating a While the ISO/IEC JTC 1/SC 42 was created very recently,
chain-of-thought (Wei et al., 2022) through prompts. For example, another standardisation subcommittee under ISO/IEC JTC1 has a
X-IQE (Chen et al., 2023) leverages MiniGPT-4 (Zhu et al., 2024) much longer history of nearly 20 years: the ISO/IEC JTC 1/SC 3714
with prompts prepared with the help of art professionals to evaluate that focuses on biometrics-related technology. This standardisation
the fidelity, alignment, and aesthetics of the generated images. subcommittee is highly relevant for deepfake since deepfake faces
Another example is T2VScore-A (Wu et al., 2024), which uses can be used to spoof biometrics-based user authentication systems.
GPT-3.5 to identify questions, choices, and answers to evaluate In this context, the following three standards are of particular
alignment between a generated video and a given text prompt. It is relevance:
then measured based on the accuracy of visual question answering. ISO/IEC 19795-1:2021 “Information technology—Biometric
While existing MAA metrics mostly focus on the semantic performance testing and reporting—Part 1: Principles and
alignment of modalities, temporal alignment is also worth framework:”15 This standard covers general metrics about
considering for the MAA task. With this respect, AV-Align (Yariv evaluating biometric systems. Two major metrics in this context
et al., 2024) is a metric for assessing temporal alignment between are false accept rate (FAR) and false reject rate (FRR), which
audio-video pairs. It is based on energy peaks of both modalities, refer to the standard FPR and FNR, respectively. This standard also
i.e., the highest mean magnitude of optical flow for video frames deprecates the use of single-number metrics including the EER and
and the onset of the audio waveform. AUC (which were widely used in biometrics-related research in the
past).
ISO/IEC 30107-1:2016 “Information technology—Biometric
presentation attack detection—Part 1: Framework:”16 This
3.14 More about standards standard defines a general framework about presentation attack
detection (PAD) mechanisms, where the term “presentation
Many of the basic performance metrics described in this attack” refers to the “presentation of an artefact or of human
section have been widely used by deepfake researchers as de characteristics to a biometric capture subsystem in a fashion intended
facto standards, e.g., EER, log loss and MOS have been widely to interfere with system policy.” It focuses on biometric recognition
used in deepfake-related challenges. Also, the combination of systems, where a PAD mechanism is a binary classifier trying to
precision, recall and F1-score has been widely used to assess the predict presentation attacks (also called attack presentations, e.g.,
performance of binary classifiers. While there have been a number fake faces) as positive and bona fide (real) presentations as negative.
of ITU standards on PQA to date, there do not seem to be ISO/IEC 30107-3:2017 “Information technology—Biometric
many standardisation efforts on the performance metrics for the presentation attack detection—Part 3: Testing and reporting:”17
evaluation of binary classifiers. This was the case until at least This standard defines a number of special performance metrics
2017 when ISO and IEC jointly set up the ISO/IEC JTC 1/SC for evaluating PAD mechanisms standardised in the ISO/IEC
42,9 a standardisation subcommittee (SC) focusing on AI under 30107-1:2016. Three such metrics look at error rates: attack
ISO/IEC JTC 1,10 the joint technical committee for standardising presentation classification error rate (APCER) referring to the
“information technology.” standard FPR, normal/bona fide presentation classification error
One recent effort that ISO/IEC JTC 1/SC 42 made is to rate (NPCER/BPCER) referring to the standard FNR, and average
produce the ISO/IEC TR 24029-1:2021 “Artificial Intelligence classification error rate (ACER) that is defined as the average
(AI)—Assessment of the robustness of neural networks—Part 1: of the APCER and the NPCER/BPCER. Such metrics have been
Overview,”11 a technical report (TR) that systematically covers used in biometrics-related challenges such as Face Anti-spoofing
many commonly used performance assessment concepts, methods, (Presentation Attack Detection) Challenges.18 When deepfake
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 12 https://www.iso.org/standard/79799.html

learning models. 13 https://www.iso.org/standard/82570.html

14 https://www.iso.org/committee/313770.html
15 https://www.iso.org/standard/73515.html
9 https://www.iso.org/committee/6794475.html 16 https://www.iso.org/standard/53227.html
10 http://www.iso.org/iso/jtc1_home.html 17 https://www.iso.org/standard/67381.html
11 https://www.iso.org/standard/77609.html 18 https://sites.google.com/qq.com/face-anti-spoofing/

Frontiers in Big Data 07 frontiersin.org

Altuncu et al. 10.3389/fdata.2024.1400024

images or videos are used to spoof a biometric system, such deepfake image datasets, deepfake video datasets, deepfake
standardised metrics will become relevant. audio/speech datasets, deepfake text datasets, and hybrid deepfake
datasets (mainly mixed image and video datasets).
Note that many datasets of real (authentic) media were also
3.15 Discussion: performance metrics and used by deepfake researchers for two purposes. First, any detectors
standards would need both fake and real media to demonstrate their
performance. Second, real media have also been used to train
This section provides a comprehensive summary of deepfake generators as the training set. In this section, we include
performance metrics used for evaluating and benchmarking only datasets containing deepfake media, some of which contain
deepfake generators and detectors. It is rare that all such metrics both deepfake and real media.
are used for a specific application. Instead, one or several are chosen Some datasets, especially those created for deepfake-related
based on specific needs. For a deepfake detection system as a binary challenges and competitions, have separate subsets for training
classifier, many researchers have chosen to use overall metrics and evaluation (testing) purposes. The split is necessary for such
such as accuracy, AUC, EER and log loss, but the combination challenges and competitions, but not very useful for people who
of precision, recall and F1-score is also common. However, there just want to use such datasets. Therefore, in this section when
is a growing interest in using more deepfake-focused metrics, introducing such datasets we will ignore that level of details and
including PQA and MAA metrics, especially for evaluating focus on the total number of data including the number of real and
deepfake generation. Other than general evaluation metrics, some fake samples.
deepfake-related challenges and competitions have introduced
their own specific metrics, such as DSR. Furthermore, there exist
several metrics specific to certain deepfake-related tasks, including
4.1 Deepfake image datasets
code generation, animation generation, and synthetic data
generation (Bandi et al., 2023). The use of different performance
Table 2 shows basic information about the image datasets
metrics can make the comparison of different reported results
covered.
more difficult, so we hope the expected new ISO/IEC standard
SwapMe and FaceSwap dataset (Zhou et al., 2017): This dataset
particularly ISO/IEC 4213 will help.
contains 4,310 images, including 2,300 real images and 2,010 fake
It is worth mentioning that, in addition to evaluating the
images created using FaceSwap22 and the SwapMe iOS app (now
performance of deepfake detectors, the introduced performance
discontinued). The fake images were generated by tampering with
metrics for evaluating binary classifiers can also be used to evaluate
one face in each authentic image with face swapping. The selected
the performance of deepfake generation methods by considering
images cover diverse events, genders, ages, and races.
how deepfake detectors fail. For instance, organisers of the Voice
Fake Faces in the Wild (FFW) dataset (Khodabakhsh et al.,
Conversion Challenge 2018 and 2020 used this approach to
2018): This dataset contains 131,500 face images, including 78,500
benchmark how well voice conversion (VC) systems can generate
bona fide images extracted from 150 videos in the FaceForensics
high-quality fake speech samples.
dataset and 53,000 fake images extracted from 150 fake videos
Another point we would like to mention is that for deepfake
collected from YouTube. The fake images involve both tampered
videos there are two levels of performance metrics: those at the
images and those generated using CGI.
frame level (metrics of each frame), and those at the video level
generated.photos datasets23 : This is a number of commercial
(metrics for the whole video). Generally speaking, the latter can
datasets provided by Generated Media, Inc., with up to nearly
be obtained by averaging the former for all frames, potentially
2.7 million synthetic face images generated by StyleGAN. A free
following an adaptive weighting scheme, so that more important
edition with 10,000 128x128 synthetic images is made available for
(key) frames will be counted more.
academic research. The website also provides an interactive face
generator24 and an API.25 The generated.photos datasets have a
good diversity: five age groups (infants, children, youth, adults,
4 Deepfake-related datasets middle-aged), two genders (male and female), four ethnicities
(white, black, Latino, Asian), four eye colours (brown, grey, blue,
In this section, we cover all deepfake-related datasets we
green), four hair colours (brown, black, blond, grey), three hair
identified from the meta-review of deepfake-related survey papers,
length (short, medium, long), facial expressions, three head poses
deepfake-related challenges, three online collections of deepfake-
(front facing, left facing, right facing), two emotions (joy and
related datasets on GitHub,19 – 21 and the co-authors’ personal
neutral), two face styles (natural, beautified).26
collections. We explain the datasets covered in five categories:

19 https://github.com/592McAvoy/fake-face-detection#user-content-i- 22 https://github.com/MarekKowalski/FaceSwap/
dataset 23 https://generated.photos/datasets
20 https://github.com/santaboi/Diffusion-Deepfake-Detection- 24 https://generated.photos/face-generator/new
Datasets_2023 25 https://generated.photos/api
21 https://github.com/media-sec-lab/Audio-Deepfake-Detection?tab= 26 According to a number of research papers we read, an earlier 100K-
readme-ov-file#Datasets Faces dataset was released by generated.photos for academic research in

Frontiers in Big Data 08 frontiersin.org

Altuncu et al. 10.3389/fdata.2024.1400024

TABLE 2 Deepfake-related image datasets.

Dataset Size Year Generation method

SwapMe & FaceSwap 4,310 images 2017 Face swapping

Fake Faces in the Wild (FFW) 53,000 images 2018 Fake YouTube videos

generated.photos datasets 2.7 million images Since 2018 StyleGAN

MesoNet Deepfake Dataset 19,509 images 2018 Face extraction from forged videos

100K-Generated-Images 100,000 images 2019 A GAN generator

Ding et al.’s swapped face dataset 420,053 images 2019 Face swapping

iFakeFaceDB 87,000 images 2019 StyleGAN + GANprintR

Faces-HQ 40,000 images 2019-20 Collection from other datasets

CelebA-Spoof 625,537 images 2020 Face spoofing

Diverse Fake Face Dataset (DFFD) 299,039 images 2020 Multiple facial manipulation methods

DiffusionForensics 615,200 images 2023 Pretrained diffusion models

DeepFakeFace (DFF) 120,000 images 2023 Diffusion models + Face manipulation

methods

MesoNet Deepfake dataset (Afchar et al., 2018): This dataset Faces-HQ (Durall et al., 2019): This dataset includes 40,000
includes 19,457 face images,27 including 7,948 deepfake images high-resolution images, half real and half deepfake. The images
generated from 175 forged videos collected online and 11,509 real were collected from four sources: the CelebA-HQ dataset,30
face images collected from various online sources. The face images the Flickr-Faces-HQ dataset (see text footnote 28 ), the 100K-
were extracted from the collected videos by utilising a visual object Faces dataset31 (not available any longer, see the description of
detection method, and around 50 faces per scene were obtained. generated.photos datasets), and thispersondoesnotexist.com.
100K-Generated-Images (Karras et al., 2019): This dataset CelebA-Spoof (Zhang Y. et al., 2020): This dataset includes
includes 100,000 synthesised face, bedroom, car and cat images by 625,537 synthesised face images of 10,177 celebrities, with 43 rich
a GAN generator trained based on real images in the FFHQ28 and attributes on face, illumination, environment and spoof types.
LSUN29 datasets (three object types—bedrooms, cars and cats—for The real images were selected from the CelebA dataset.32 The 43
the latter). Note that the name “100K-Generated-Images” was not a attributes include 40 for real images, covering all facial components
proper one as the authors (Karras et al., 2019) just used this to name and accessories (e.g., skin, nose, eyes, eyebrows, lip, hair, hat, and
a sub-folder of their Google Drive shared space, but it was used in eyeglass), and 3 for fake images, covering spoof types, environments
one of the survey papers (Tong et al., 2020). and illumination conditions.
Ding et al. (2020)’s swapped face dataset: This dataset Diverse fake face dataset (Dang et al., 2020): This dataset
contains 420,053 images of celebrities, including 156,930 real contains 299,039 images, including 58,703 real images sampled
ones downloaded using Google Image API and 263,123 fake from three datasets [FFHQ (see text footnote 28 ), CelebA (see text
face-swapped ones created using two different methods (Nirkin’s footnote 32 ), and FaceForensics++33 ] and 240,336 fake ones in four
method and Auto-Encoder-GAN). While the former method main facial manipulation types (identity swap, expression swap,
consisted of multiple techniques pipelined, the latter was fully attribute manipulation, and entire synthesis). The images cover two
automated based on a CNN architecture. genders (male and female), a wide age group (the majority between
iFakeFaceDB (Neves et al., 2020): This dataset includes 21 and 50 years old), and both low- and high-quality levels.
87,000 224 × 224 face images, generated by processing some DiffusionForensics (Wang et al., 2023a,b): This dataset
StyleGAN-generated synthetic images using the GAN-fingerprint contains 615,200 diffusion-generated images, sourced from LSUN-
Removal approach (GANprintR) proposed by Neves et al. (2020). Bedroom (see text footnote 29 ), ImageNet (Deng et al., 2009),
It is the replaced version of the FSRemovalDB dataset, which and CelebA-HQ (see text footnote 30 ) datasets. The dataset
contains 150,000 face images generated using an earlier version of covers images that belong to one of the following image
GANprintR. generation methods—unconditional, conditional, and text-to-
image generation, and contains three fields—source image,
reconstructed image, and the image referring to the error between

2018, which was used by many researchers. This dataset is not currently
available any longer.
27 Table 2 of the paper shows the dataset size is 19,509, but the dataset 30 https://drive.google.com/open?id=0B4qLcYyJmiz0TXY1NG02bzZVRGs
downloaded from pCloud contains just 19,457 images. 31 https://generated.photos/
28 https://github.com/NVlabs/ffhq-dataset 32 http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html
29 https://github.com/fyu/lsun 33 https://github.com/ondyari/FaceForensics

Frontiers in Big Data 09 frontiersin.org

Altuncu et al. 10.3389/fdata.2024.1400024

these images. Moreover, 11 different generators were leveraged for subjects of different ages, ethnic groups and genders, and 795
the generation of the fake images. deepfake videos synthesised from these real videos.
DeepFakeFace (DFF) (Song et al., 2023a,b): This dataset Celeb-DF v2 (Li Y. et al., 2020): This dataset contains 6,229
contains 30,000 real and 90,000 fake celebrity images. The real face videos of celebrities, including 590 real videos collected from
images were retrieved from the IMDB-Wiki (Rothe et al., 2015) YouTube with subjects of different ages, ethnic groups and genders,
dataset, and the fake ones were generated by utilising three different and 5,639 deepfake videos synthesised from these real videos. The
methods—Stable Diffusion v1.5 and Stable Diffusion Inpainting deepfake videos were generated by swapping faces for each pair of
diffusion models, as well as the InsightFace toolbox containing the 59 celebrities, using an improved DeepFake synthesis algorithm.
several face manipulation algorithms. DeepFake detection dataset (Dufour and Gully, 2019): 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 AI36 and JigSaw.37 The deepfake videos
4.2 Deepfake video datasets were generated by using publicly available face-swapping methods
although no more details were provided.
Table 3 shows basic information about the video datasets DeeperForensics-1.0 (Jiang et al., 2020): This dataset contains
covered. 60,000 indoor face videos (with 17.6 million frames) generated
DeepfakeTIMIT (Korshunov and Marcel, 2019): This dataset by face swapping, covering 100 subjects, four skin tones (white,
contains 620 deepfake face videos, generated by GAN-based face black, yellow, and brown), two genders (male and female), different
swapping without manipulation of audio, covering 32 subjects and age groups (20–45), 26 nationalities, 7 different angles, 8 face
two quality levels (high and low). The videos in the dataset are expressions, and different head poses.
recordings of people facing the camera and reciting predetermined DFDC (Deepfake Detection Challenge) full
short English sentences. dataset (Dolhansky et al., 2020): This dataset contains 128,154 face
FaceForensics (FF) (Rössler et al., 2018): This dataset contains videos of 960 subjects, including 23,654 real videos from 3,426 paid
1,004 face videos with over 500,000 frames, covering various actors and 104,500 deepfake videos created using eight different
quality levels and two types of facial manipulation using the methods (DF-128, DF-256, MM/NN face swap, NTH, FSGAN,
Face2Face approach. This dataset is now replaced by the larger StyleGAN, refinement, and audio swap).
FaceForensics++ dataset (see below). WildDeepfake (Zi et al., 2020): This dataset contains 7,314 face
FaceForensics++ (FF++) (Rössler et al., 2019): This dataset sequences extracted from 707 deepfake videos that were collected
contains 5,000 face videos with over 1.8 million manipulated completely from the Internet. It covers diverse scenes, multiple
frames, including 1,000 real videos (with 509,914 frames) persons in each scene and rich facial expressions. Different from
downloaded from YouTube, and 4,000 fake videos created using other deepfake video datasets, WildDeepfake contains only face
four face manipulation methods (Deepfakes, Face2Face, FaceSwap sequences, not the full videos. This makes the dataset more like
and NeuralTextures). The videos cover two genders (male and between an image dataset and a video one. We decided to keep
female), and three quality levels (VGA/480p, HD/720p, and it in the video category since the selection process was still more
FHD/1080p). video-focused.
UADFV dataset (Li et al., 2018): This dataset contains 98 face FFIW10K (Face Forensics in the Wild) dataset (Zhou et al.,
videos, half (49) are real ones downloaded from YouTube, and 2021): This dataset contains 10,000 real and 10,000 high-quality
the other half are fake ones generated using the FakeApp mobile forgery videos, with video- and face-level annotations. The dataset
application (which is now discontinued). The video dataset was focuses on a more challenging case for forgery detection: each
created to demonstrate a deepfake video detection method based on video involves one to 15 individuals, but only some (a minority
the detection of eye-blinking behaviours, so all the videos contain of) faces are manipulated. The deepfake videos were generated
at least one eye-blinking event. All fake videos were created by by utilising three different face-swapping methods—two learning-
swapping the original face in each of the real videos with the face based methods, DeepFaceLab and FSGAN, as well as a graphic-
of the actor Nicolas Cage,34 thus, only one subject is represented. based method, FaceSwap.
Deep fakes dataset (Ciftci et al., 2020): This dataset contains Korean DeepFake Detection Dataset (Kwon et al., 2021): This
142 “in the wild” deepfake portrait videos, collected from a dataset contains 37,942 videos of paid subjects (395 Koreans and
range of online sources including news articles, online forums, 8 Southeastern Asians), including 62,166 real videos and 175,776
mobile apps, and research presentations. The videos are diverse, fake ones created using six methods, including three face-swapping
covering the source generative model, resolution, compression, methods (i.e., FaceSwap, DeepFaceLab, and FSGAN), one video-
illumination, aspect ratio, frame rate, motion, pose, cosmetics, driven face reenactment method [i.e., First Order Motion Model
occlusion, content, and context. (FOMM)], and two audio-driven face reenactment method [i.e.,
Celeb-DF v135 : This dataset contains 1,203 face videos of Audio-driven Talking Face HeadPose (ATFHP) and Wav2Lip]. The
celebrities, including 408 real videos collected from YouTube with videos cover a balanced gender ratio and a wide range of age groups.

34 https://en.wikipedia.org/wiki/Nicolas_Cage
35 https://github.com/yuezunli/celeb-deepfakeforensics/tree/master/ 36 https://ai.googleblog.com/
Celeb-DF-v1 37 https://jigsaw.google.com/

Frontiers in Big Data 10 frontiersin.org

Altuncu et al. 10.3389/fdata.2024.1400024

TABLE 3 Deepfake-related video datasets.

Dataset Size Year Language/ethnicity Generation method

DeepfakeTIMIT 620 videos 2018 English GAN-based face swapping

FaceForensics 1,004 videos 2018 Not specified Face manipulation

UADFV 98 videos 2018 Not specified Multiple GANs

FaceForensics++ 5,000 videos 2019 Not specified Multiple face manipulation

methods

Deep Fakes Dataset 142 videos 2020 Not specified Collection

Celeb-DF v1 1,203 videos 2020 Multiple ethnic groups Face swapping

Celeb-DF v2 6,229 videos 2020 Multiple ethnic groups Face swapping

DFD 3,363 videos 2019 Not specified Face swapping

DeeperForensics-1.0 60,000 videos 2020 26 nationalities Face swapping

DFDC (full) 128,154 videos 2020 Not specified Face/Audio swapping + GANs

WildDeepfake 7,314 face sequences 2020 Not specified Online collection

FFIW10K 20,000 videos 2021 Multilingual Face swapping

KoDF 37,942 videos 2021 Korean subjects Face swapping + Face reenactment

VideoForensicsHQ 1,737 videos 2021 Not specified Deep Video Portraits (DVP)

DF-W 1,869 videos 2021 Not specified Online collection

FMFCC-V 82,392 videos 2022 Asian subjects Face swapping

DF-Platter 133,260 videos 2022 Indian subjects Face swapping

AV-Deepfake1M 1,146,760 videos 2023 Not specified Face reenactment + Text-to-speech

DFDM 6,450 videos 2022 Not specified Face swapping

VideoForensicsHQ (Fox et al., 2021): This dataset contains Chinese) individuals. Then, the synthesised videos were generated
1,737 videos with 1,666,816 frames, including 1,339,843 real frames by leveraging four face-swapping methods—Faceswap, Faceswap-
and 326,973 fake frames generated using Deep Video Portraits GAN, DeepFaceLab, and Recycle-GAN.
(DVP) (Kim et al., 2018), i.e., a method that enables to transfer DF-Platter (Narayan et al., 2023a,b): This dataset contains
head pose, facial expression, and eye motion of the source while 764 real videos of 454 Indian individuals, and 133,260 deepfake
preserving the target’s identity and appearance. The original videos videos generated using three state-of-the-art synthesis methods:
were obtained from three sources: the dataset used in (Kim FSGAN (Nirkin et al., 2019), FaceSwap, and FaceShifter (Li L.
et al., 2019), the Ryerson Audio-Visual Database of Emotional et al., 2020). The videos were collected in the wild, particularly
Speech and Song (RAVDESS) (Livingstone and Russo, 2018), and from YouTube, considering many diversity factors, such as gender,
YouTube. Most videos have a resolution of 1280 × 720. orientation, skin tone, face size (counted in pixels), lighting
DF-W (Pu et al., 2021a,b): This dataset contains 1,869 real- conditions, background, and in the presence of occlusion. Each
world deepfake videos collected from two online video portals: video lasts approximately 20 seconds in duration.
YouTube (1,062) and Bilibili38 (807). The authors also collected the AV-Deepfake1M (Cai et al., 2023a,b): This dataset contains
same number of real videos from 6 research community datasets over 1 million videos (286,721 real and 860,039 fake),
for the results reported in their paper (Pu et al., 2021a), but they corresponding to 1,886 hours of audio-visual data, generated
chose not to release such videos as part of DF-W. from 2,068 unique subjects. The dataset covers different video
FMFCC-V (Li G. et al., 2022a,b): This dataset contains 38,102 manipulation techniques, including fake audio over fake visuals,
deepfake videos and 44,290 pristine videos, corresponding to over fake audio over real visuals, and real audio over fake visuals. The
23 million frames. It was created by a group of Chinese researchers fake audios were generated by an identity-dependent text-to-
from the State Key Laboratory of Information Security, Institute speech method, VITS, while the TalkLip model was used for face
of Information Engineering, Chinese Academy of Sciences, for reenactment to generate lip-synchronised fake visual frames.
the accompanying video track of the first Fake Media Forensics Deepfakes from different models (Jia et al., 2022a,b): This
Challenge of the China Society of Image and Graphics. The dataset contains 6,450 face-swap deepfake videos generated by five
source videos were collected from 83 paid Asian (likely all different Autoencoder models based on the Faceswap39 software
(i.e., Faceswap, Lightweight, IAE, Dfaker, and DFL-H128). For the

38 A Chinese video portal: https://www.bilibili.com/. 39 https://github.com/deepfakes/faceswap

Frontiers in Big Data 11 frontiersin.org

Altuncu et al. 10.3389/fdata.2024.1400024

generation of deepfake videos, real videos in the Celeb-DF dataset (with 4 × 4 and 4 × 6 source-target speaker pairs, respectively).
were used. The dataset is available upon request. 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
4.3 Deepfake audio/speech datasets evaluation. Compared to the 2018 challenge, DL-based VC systems
were overwhelmingly used by almost all participating teams (i.e.,
Table 4 shows basic information about the audio/speech WaveNet and WaveGAN were among the most used DL-based
datasets covered. building blocks).
Voice conversion (VC) is a technology that can be used to A major set of deepfake speech datasets were created for
modify an audio and speech sample so that it appears as if spoken the ASVspoof (Automatic Speaker Verification Spoofing and
by a different (target) person than the original (source) speaker. Countermeasures) Challenge43 (2015–2021, held biannually). The
Obviously, it can be used to generate deepfake audio/speech datasets for the 2019 and 2021 challenges contain speech data that
samples. The biennial Voice Conversion Challenge40 that started can be considered deepfakes.
in 2016 is a major challenge series on VC. Datasets released from ASVspoof 2019 challenge database (Wang et al., 2020): This
this challenge series are very different from other deepfake datasets: dataset is based on the Voice Cloning Toolkit (VCTK) corpus,44 a
the deepfake data is not included in the original dataset created by multi-speaker English speech database captured from 107 speakers
the organisers of each challenge, but in the participant submissions (46 males and 61 females). Two attack scenarios were considered:
(which are retargeted/fake utterances produced by VC systems built logical access (LA) involving spoofed (synthetic or converted)
by participants). The challenge datasets also include the evaluation speech, and physical access (PA) involving replay attacks of
(listening-based) results of all submissions. Some fake utterances previously recorded bona fide recordings). For our purpose in
may be produced by DL-based VC systems, so we consider all this survey, the LA scenario is more relevant. The LA part of the
datasets from this challenge series relevant for the purpose of this dataset includes 12,483 bona fide (real) utterances and 108,978
survey. spoofed utterances. Some of the spoofed speech data for the LA
Voice conversion challenge 2016 database (Toda et al., 2016): scenario were produced using a generative model involving DL-
The original dataset created by the challenge organisers was derived based techniques such as long short-term memory (LSTM),45
from the DAPS (Device and Produced Speech) Dataset (Mysore, WaveNet (Oord et al., 2016), WaveRNN (Kalchbrenner et al.,
2015). It contains 216 utterances (162 for training and 54 for 2018), and WaveCycleGAN2 (Tanaka et al., 2019). Note that the
testing) per speaker from 10 speakers. Participating teams (17) challenge organisers did not use the term “deepfake” explicitly,
developed their own VC systems for all 25 source-target speaker despite the fact that the DL-generated spoofed speech data can be
pairs and then submitted generated utterances for evaluation. At considered as deepfakes.
least six participating teams used DL-related techniques (LSTM, ASVspoof 2021 challenge—logical access database (Delgado
DNN) in their VC systems (see Table 2 of the result analysis et al., 2021a; Liu et al., 2023): This dataset contains 16,492 bona
paper41 ), so the submitted utterances can certainly be considered fide and 148,148 spoofed speech data for the logical access (LA)
deepfakes. task. The bona fide speech data in the dataset were sourced from
Voice conversion challenge 2018 database (Lorenzo-Trueba the ASVspoof 2019 LA evaluation database, and the spoofed data
et al., 2018): The original dataset created by the challenge organisers were generated by leveraging 13 voice conversion, text-to-speech,
was also based on the DAPS dataset. It contains 116 utterances (81 and hybrid spoofing attack algorithms.
for training and 35 for testing) per speaker from 12 speakers in two ASVspoof 2021 challenge—speech deepfake
different tasks (called Hub and Spoke). Participating teams (23 in database (Delgado et al., 2021b; Liu et al., 2023): In 2021,
total, all for Hub and 11 for Spoke) developed their own VC systems the challenge included an explicitly defined track on deepfake, but
for all 16 source-target speaker pairs and then submitted generated the task description suggests that the organisers of the challenge
utterances for evaluation. Compared to the 2016 challenge, more considered a broader definition of the term “deepfake” by looking
participating teams used DL-related techniques (e.g., WaveNet, at spoofing human listeners rather than ASV (Automatic Speaker
LSTM, DNN, CycleGAN, DRM – deep relational models, and Verification) systems. The dataset includes 20,637 bona fide and
ARBM – adaptive restricted Boltzmann machines) in their VC 572,616 spoofed speech data. Other than the ASVspoof 2019 LA
systems. evaluation database, the bona fide speech data in this dataset were
Voice conversion challenge 2020 database (Yi et al., 2020): also sourced from VCC 2018 and VCC 2020 datasets. When it
This dataset is based on the Effective Multilingual Interaction comes to spoofed data generation, over 100 different spoofing
in Mobile Environments (EMIME) dataset,42 a bilingual attack algorithms were used.
(Finnish/English, German/English, and Mandarin/English) ReMASC (Gong et al., 2019a,b): This dataset contains 9,240
database. It contains 145 utterances (120 for training and 25 genuine and 45,472 replayed recordings of voice commands in
for testing) per speaker from 14 speakers for two different tasks English performed by English, Chinese, and Indian native speakers.
Although the dataset mainly focuses on replay attacks rather than

40 http://www.vc-challenge.org/
41 http://www.vc-challenge.org/vcc2016/papers/ 43 https://www.asvspoof.org/
SSW9_VCC2016_Results.pdf 44 https://doi.org/10.7488/ds/1994
42 https://www.emime.org/participate/emime-bilingual-database.html 45 https://www.cs.toronto.edu/~graves/phd.pdf

Frontiers in Big Data 12 frontiersin.org

Altuncu et al. 10.3389/fdata.2024.1400024

TABLE 4 Deepfake-related audio datasets (VC: Voice Conversion, TTS: Text-to-speech, AS: Audio Splicing, SA: Spoofing Attack, VS: Voice Synthesis, SE:
Speech Enhancement, CMF: Copy-Move Forgery).

Dataset Size Year Language Generation method

VCC 2016 dataset 3,078 utterances 2016 English VC

VCC 2018 dataset 2,582 utterances 2018 English VC

VCC 2020 dataset 3,505 utterances 2020 English, Finnish, VC

German, Chinese

ASVspoof 2019 dataset 121,461 utterances 2019 English VS, VC

(LA task)

ASVspoof 2021 dataset 164,640 utterances 2021 English VC, TTS, Hybrid SA
(LA task)

ASVspoof 2021 dataset 593,253 utterances 2021 English SA

(DF task)

ReMASC 54,712 recordings 2019 English VS

FSD 650 songs 2023 Chinese VC, VS

DECRO 118,382 utterances 2023 English, Chinese SA

WaveFake 134,268 audios 2021 English, Japanese TTS

HAD 160,836 audios 2021 Chinese TTS, AS

ADD 2022 dataset 154,949 audios 2022 Chinese VS, VC

CMFD 5,600 audios 2022 English, Chinese CMF

In-the-Wild 31,779 audios 2022 English Online collection

SceneFake 84,480 audios 2022 English SE

EmoFake 88,200 audios 2022 English VC

PartialSpoof 121,461 audios 2021–2022 English VC, TTS, AS

CFAD 331,600 audios 2023 Chinese TTS, AS

ADD 2023 dataset 517,068 utterances 2023 Chinese VC, TTS, AS

MLAAD 163.9 hours of synthetic 2024 23 languages TTS

voice

synthesised speech, it also contains fake videos generated using clips in English and Japanese, which corresponds to 175
two speech synthesis methods. The dataset also includes a quick h of audio data. The reference datasets for real videos
evaluation set, which is a small but representative dataset with include LJSPEECH for English and JSUT for Japanese.
around 2,000 samples, that can be used for the quick evaluation of For the generation of fake audio clips, six different GAN-
developed audio deepfake detection models. based architectures (i.e., MelGAN, PWG, Multi-band
Fake song detection dataset (Xie et al., 2023a,b): This dataset MelGAN, Full-band MelGAN, HiFi-GAN, and WaveGlow)
contains 200 real and 450 fake songs in Chinese. To generate fake were leveraged.
songs, initially, fake singing voices were generated by five state- Half-truth audio detection dataset (Yi et al., 2021): This
of-the-art singing voice synthesis (DiffSinger) and singing voice dataset focuses on the attack scenarios in which the attacker
conversion methods (RVC and three variations of SO-VITS). Then, hides small fake audio clips into real speech audio. The dataset
instrumental tracks were extracted from real songs and combined comprises 53,612 real audio clips and 107,224 fake audio clips
with fake singing voices. in Chinese. The fake audio clips in the dataset cover both fully
DEepfake CROss-lingual dataset (Ba et al., 2023a,b): This fake and partially fake samples. While the fully fake samples
dataset consists of English and Chinese speech samples. The were generated utilising a text-to-speech method, namely GST
English part of the dataset contains 12,484 real and 42,800 fake Tacotron,46 the latter set of videos were based on the manipulation
utterances while the Chinese part consists of 21,218 real and of the audio segments corresponding to selected keywords in the
41,880 fake utterances. For the real speech samples, six Chinese original utterances.
recording datasets and the ASVspoof2019 LA (Wang et al., 2020) ADD 2022 dataset (Yi et al., 2022a): This is the original
dataset were used. The fake utterances, however, were generated dataset of the first audio deep synthesis detection challenge,
by leveraging 10 different spoofing attack algorithms, including held in 2022. The dataset covers data for the three tracks of
the challenge—low-quality fake audio detection (LF), partially
HiFiGAN, Multiband-MelGAN, PWG, Tacotron, FastSpeech2,
fake audio detection (PF), and audio fake game (FG). In
StarGANv2, VITS, NVCNet, Baidu, and Xunfei.
WaveFake (Frank and Schönherr, 2021a,b): This dataset
contains 16,283 real audio clips and 117,985 fake audio 46 https://github.com/syang1993/gst-tacotron

Frontiers in Big Data 13 frontiersin.org

Altuncu et al. 10.3389/fdata.2024.1400024

total, the dataset comprises 16,257 real audio samples and Dataset as well as more fake audios generated by manipulating the
138,692 fake audio samples in Chinese. The dataset is based original utterances with real or synthesised audios.
on common Mandarin speech datasets, including AISHELL- Multi-language audio anti-spoofing dataset (Müller et al.,
1, AISHELL-3, and AISHELL-4, and the fake audios were 2024): This dataset is based on MAILABS Speech Dataset and
generated using mainstream speech synthesis and voice covers 163.9 h of synthetic voice. While the real video clips in the
conversion methods. dataset involve human speeches in eight languages (i.e., English,
Copy-Move forged dataset (Su et al., 2023a,b): This dataset French, German, Italian, Polish, Russian, Spanish, and Ukrainian),
contains 3,600 real audio samples and 2,000 fake audio samples the dataset was enriched with computer-generated audio clips in
in two languages – English and Chinese. The dataset was based on 23 languages. For the generation of fake clips, 54 text-to-speech
Librispeech and Chinspeech speech datasets, and the fake samples models comprising 21 different architectures were utilised.
were generated by applying copy-move forgery to the samples of
these two datasets.
In-the-Wild (Müller et al., 2022): This dataset focuses on
detecting audio deepfakes for politicians and celebrities. It 4.4 Deepfake text datasets
comprises 19,963 real speeches and 11,816 fake audio samples of
58 public figures, derived from publicly available sources. In total, Table 5 shows basic information about the text datasets covered.
the dataset corresponds to 20.8 h of benign and 17.2 h of spoofed gpt-2-output-dataset (OpenAI, 2019): This dataset contains
audio data. outputs generated by GPT-2 trained on the OpenAI’s Webtext
SceneFake (Yi et al., 2022b, 2024): This dataset focuses on training set. It comprises 250K training, 5K test, and 5K validation
detecting the manipulation of acoustic scenes, e.g., airport, street, samples for each of the nine different GPT-2 models trained, in
or shopping, of audios. It consists of 19,838 real and 64,642 fake total, 2,340,000 text samples.
audio samples in English. The dataset was constructed based on TweepFake (Fagni et al., 2021a,b): This dataset consists of
ASVspoof 2019 LA (Wang et al., 2020) dataset and the acoustic 25,572 tweets collected from 23 bots, imitating 17 human accounts.
scene dataset from DCASE202247 challenge. The fake samples were The bots leveraged different methods for generating fake tweets—
generated by leveraging different speech enhancement models, most were based on GPT-2 and RNN although other methods were
including SSub, MMSE, Wiener, FullSubNet, WaveU-Net, and also used, including Markov Chains and LSTM.
GCRN. TuringBench (Uchendu et al., 2021): Focusing on the tasks
Chinese fake audio detection dataset (Ma et al., 2023a,b): This of the Turing Test and Authorship Attribution for language
dataset consists of 115,800 real and 115,800 fake audio samples in models, this dataset is based on 10,000 news articles (mostly
Chinese. The audio samples in the dataset were generated by using about politics). The titles of these articles were used to prompt 19
the samples taken from 1302 speakers. The real audio samples were different LLMs (e.g., GPTs, GROVER, CTRL, XLM, PPLM etc.) to
sourced from six different datasets, and twelve types of fake audio generate synthetic news articles similar to the original ones. After
samples were generated from the real samples. preprocessing, the size of the final dataset was reduced to 168,612.
EmoFake (Zhao et al., 2023a,b): This dataset was designed GeneratedTextDetection (Liyanage et al., 2022a,b): This
for the detection of the changed emotion state of original dataset contains two different corpora – 100 research papers
audios. It comprises 35,000 real and 53,200 fake audio samples collected from arXiv.org with 100 fully synthetic papers generated
in English and supports five emotion states – neutral, happy, using GPT-2 and 100 original abstracts from arXiv.org preprints
sad, angry, and surprise. The fake audio samples were generated with 100 abstracts semi-automatically manipulated (i.e., keeping
by leveraging seven open-source emotion voice conversion some sentences from the original content) using a GPT-2-based
models—VAW-GAN-CWT, DeepEST, Seq2Seq-EVC, CycleGAN- model, Arxiv-NLP.
EVC, CycleTransGAN, EmoCycle-GAN, and StarGAN-EVC. ToxiGen (Hartvigsen et al., 2022): Being constructed for hate
PartialSpoof (Zhang et al., 2021a,b): Similar to HAD, this speech detection, this English dataset contains 137,000 benign
dataset helps the detection of partially fake audio samples. It covers and 137,000 toxic sentences, both generated using GPT-3. The
12,483 real and 108,978 fake samples in English. The real samples sentences in the dataset mention 13 different minority groups
in the dataset were collected from the ASVspoof 2019 LA dataset. without using explicit language.
Then, fake audios were generated by spoofing speech segments of MAGE (formerly, DeepfakeTextDetect) (Li et al., 2024):
the bona fide audio through voice conversion, text-to-speech, and This dataset used different types of real texts collected from
audio splicing. existing datasets, including opinion statements, news articles,
ADD 2023 Dataset (Yi et al., 2023): This is the dataset for answers to questions, stories, sentence sets for reasoning, Wikipedia
the second Audio Deepfake Detection Challenge. The challenge paragraphs, and abstracts of scientific articles. Then, synthetic texts
had three main tracks – audio fake game (FG), manipulation were generated using 27 LLMs with three different prompts. In
region location (RL), and deepfake algorithm recognition (AR). In total, it contains 447,674 human-written and machine-generated
total, the dataset involves 243,194 real utterances and 273,874 fake texts.
utterances. The dataset contains utterances from the ADD 2022 AI-Writer (Pu et al., 2023): This dataset utilised AI-Writer,48
i.e., an online commercial service for generating fake news articles,

47 https://dcase.community/challenge2022/index 48 https://ai-writer.com/

Frontiers in Big Data 14 frontiersin.org

Altuncu et al. 10.3389/fdata.2024.1400024

TABLE 5 Deepfake-related text datasets.

Dataset Size Year Language Generation method

gpt-2-output-dataset 2,340,000 texts 2019 Not specified GPT-2

TweepFake 25,572 tweets 2021 English Fake account collection

TuringBench 168,612 news articles 2021 English 19 LLMs

GeneratedTextDetection 200 scientific papers + 2022 English GPT-2

200 abstracts

ToxiGen 274,000 sentences 2022 English GPT-3

MAGE 447,674 texts 2023 English 27 LLMs

AI-Writer 2,000 news articles 2023 Not specified Text-generation service

ArticleForge 2,000 news articles 2023 Not specified Text-generation service

Kafkai 2,000 articles 2023 Not specified Text-generation service

RedditBot 1,774 comments 2023 Not specified GPT-3 powered bot

IDMGSP 29,000 scientific papers 2023 English Context-free grammar, LLMs

AIRABIC 1,000 texts 2023 Arabic ChatGPT

HC3 125,230 answers to 2023 English, Chinese ChatGPT

37,175 questions

Deepfake-BG 9,824 social media posts 2023 Bulgarian GPT-2, ChatGPT

HC3 Plus 214,498 texts 2024 English, Chinese ChatGPT

CHEAT 50,699 abstracts 2024 English ChatGPT

to generate 1,000 synthetic articles from 1000 real news titles passages from books and news articles, covering both classical and
sampled from the RealNews dataset. modern standard Arabic. Moreover, the dataset contains different
ArticleForge (Pu et al., 2023): Similar to AI-Writer, this text variations, including single and multi-paragraph compositions,
dataset was constructed using a commercial text generation service, bullet points, and passages with in-text citations.
ArticleForge,49 which requires a set of keywords to generate fake Human ChatGPT comparison corpus (Guo et al., 2023a,b):
news articles. The dataset contains 1,000 real and 1,000 synthetic This bilingual (i.e., English and Chinese) dataset involves questions
news articles. as well as answers to them provided by humans and ChatGPT
Kafkai (Pu et al., 2023): This dataset leveraged the Kafkai to enable a comparison between human and ChatGPT answers
commercial text generation service, which generates synthetic to the same questions. In total, it covers 37,175 questions, 80,805
articles from a chosen category and an initial priming text. For the human answers, and 44,425 ChatGPT answers. The questions and
construction of the dataset, 100 real articles on 10 of the 25 available human answers were collected from four English and six Chinese
categories, such as cyber security, SEO, and marketing, were used to question-answering datasets. In addition, more question-answer
generate 1,000 fake articles. pairs were constructed by crawling concept-explanation pairs from
RedditBot (Pu et al., 2023): This dataset is based on a GPT- online encyclopaedias, Wikipedia50 and BaiduBaike.51
3 powered bot posting comments on a popular subreddit on Deepfake-BG (Temnikova et al., 2023): This dataset consists
Reddit.com. 887 comments posted by the bot were collected as of 4,912 human-written and 4,912 synthetic social media posts
synthetic comments while the same amount of real comments were in Bulgarian. The human-written posts were randomly selected
sampled randomly from the forum threads with a bot comment. from multiple large datasets. The other posts were generated by
Identifying machine-generated scientific papers (Mosca et al., leveraging two different LLMs—a new Bulgarian GPT-2 model,
2023): This dataset contains 16,000 real scientific papers retrieved called GPT-WEB-BG, and ChatGPT for Bulgarian.
from arXiv.org and 13,000 fake papers generated by utilising HC3 Plus (Su et al., 2024): This is an extended version
different methods. These methods involve LLMs, i.e., GPT-2, GPT- of the HC3 dataset, covering translation, summarisation, and
3, ChatGPT, and Galactica, as well as SCIgen which is based on paraphrasing tasks. It contains 144,528 English texts and 69,970
context-free grammars. In total, the dataset consists of 25 million Chinese texts, all generated by using ChatGPT. In terms of types
tokens. of texts, the dataset involves news articles, social media posts, and
AIRABIC (Alshammari and EI-Sayed, 2023a,b): This Arabic questions.
dataset includes 500 human-written texts and 500 ChatGPT-
generated texts. The human-written texts were sourced from
50 https://www.wikipedia.org/

49 https://www.articleforge.com/ 51 https://baike.baidu.com/

Frontiers in Big Data 15 frontiersin.org

Altuncu et al. 10.3389/fdata.2024.1400024

CHatGPT-writtEn AbsTract (Yu et al., 2024a,b): This dataset generated datasets are fake since they are not produced from real-
contains 15,395 human-written abstracts and 35,304 ChatGPT- world scenes. With this respect, we cover some state-of-the-art
written abstracts. While human-written abstracts were collected deepfake dataset generators that can be used to obtain new deepfake
from IEEE Xplore, ChatGPT was utilised to generate three types datasets, here.
of synthetic abstracts—fully synthetic based on title and keywords, DatasetGAN (Zhang Y. et al., 2021): This generator
polished version of human-written abstracts, and a mixture of focuses on generating image-annotation pairs. It initially
human-written abstracts and their polished versions. requires synthesising a small number of images with a GAN
architecture and their annotation by humans. Then, an ensemble
of multilayer perceptron (MLP) classifiers is trained on the
pixel-wise feature vector of the used GAN architecture. This
4.5 Hybrid deepfake datasets
trained classifier is used as the label synthesis branch of the
GAN architecture.
Table 6 shows basic information about the hybrid datasets
BigDatasetGAN (Li D. et al., 2022): This is an extended version
covered.
of DatasetGAN, enhancing the ImageNET dataset, covering 1
NIST OpenMFC (Open Media Forensics Challenge)
million images, with pixel-wise labels. It leverages two ImageNET-
Datasets52 : These datasets were created by the DARPA Media
pretrained models, BigGAN and VQGAN, to generate 5 samples
Forensics (MediFor) Program53 for the 2020 OpenMFC.54 There
for each of the 1,000 classes in ImageNET. Similar to DatasetGAN,
are two GAN-generated deepfake datasets, one with more than
these samples are manually labelled and used to train a classifier for
1,000 deepfake images and the other with over 100 deepfake videos.
pixel-level label synthesis.
The datasets were made available to registered participants of the
DatasetDM (Wu et al., 2023): This aims to enhance the
competition only.
previous generators in terms of the produced image quality
ForgeryNet (He et al., 2021): This dataset is named as
and generalisability. It makes use of text-to-image diffusion
“a versatile benchmark for comprehensive forgery analysis.” It
models (e.g., Stable Diffusion) to generate more realistic synthetic
contains 2,896,062 images and 221,247 videos, including 1,457,861
images with several perception annotations, including depth,
fake images and 121,617 fake videos. The videos and images cover
segmentation, and human pose estimation. It also requires
seven image-level and eight video-level manipulation approaches,
much fewer samples to be manually labelled, compared to the
36 different types of perturbations and more mixed perturbations,
previous generators.
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 manipulation approaches
belong to five main generation methods—face reenactment, face 4.7 Subjective quality of deepfakes in
editing, face transfer, face swapping, and face stacked manipulation. different databases
The dataset is being used for supporting the Face Forgery Analysis
Challenge 202155 at the SenseHuman 2021 (3rd Workshop on As mentioned in Section 3.12, subjective quality evaluation is
Sensing, Understanding and Synthesizing Humans),56 co-located at necessary to evaluate the realness, realisticness, and naturalness
the ICCV 2021 conference.57 of deepfake media. While there has been very limited work on
Homologous deepfake dataset (Xidian University, Modern this topic, in 2020, Jiang et al. (2020) conducted a user study on
Image Processing Lab, 2024): This dataset contains 6,802 fake the realness of deepfake videos. They recruited 100 professional
images, 400 fake videos, and 100 real videos. According to the participants (most of whom are computer vision researchers),
developers of the dataset, it is the first facial dataset for Chinese who were asked to evaluate the realness of 30 randomly selected
people. The dataset covers four types of deepfake multimedia, videos from 7 deepfake video datasets (DeeperForensics-1.0,
including full-face generation, attribute editing, face reenactment, UADFV, DeepFake-TIMIT, Celeb-DF, FaceForensics++, Deep Fake
and face swap. 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
4.6 Deepfake dataset generators
agree). Table 7 shows the results. Interestingly, we can see a huge
difference between the realness levels of different datasets. What
Despite not being datasets per se, dataset generators are
is probably quite surprising is that FaceForensics++, one of the
systems for producing large datasets more automatically, including
most widely used deepfake datasets, has a very low MOS score
generating deepfake datasets. One may argue the automatically
and less than 9% of participants considered the 30 selected videos
as real.
52 https://mfc.nist.gov/#pills-data
53 https://www.darpa.mil/program/media-forensics
54 https://mfc.nist.gov/ 4.8 Discussion: datasets
55 https://competitions.codalab.org/competitions/33386
56 https://sense-human.github.io/ Among all deepfake image and video datasets, a significant
57 http://iccv2021.thecvf.com/ majority are about face images and videos. This is not surprising

Frontiers in Big Data 16 frontiersin.org

Altuncu et al. 10.3389/fdata.2024.1400024

TABLE 6 Hybrid deepfake-related datasets.

Dataset Size Year Language/Ethnicity Generation method

NIST open media Over 1,000 images and 2020 Not specified GAN
forensics challenge over 100 videos
datasets

ForgeryNet dataset 2,896,062 images and 2021 Not Specified Face reenactment, Face editing,
221,247 videos Face transfer, Face swapping, Face
stacked manipulation

Homologous deepfake 6,802 images and 500 2024 Chinese Face swapping, Face reenactment,
dataset videos Attribute editing, GAN

TABLE 7 Human-judged subjective quality (realness) of deepfake videos

emerged in 2021: a number of very recent datasets started focusing
in 7 datasets.
on more realistic deepfakes (i.e., in the wild) or more specified areas
Dataset MOS 4+ ratings (%) of deepfakes (e.g., FFIW10K focusing on multiple faces in the same
video, and KoDF focusing on Korean faces). This trend shows that
DeeperForensics-1.0 3.806 64.1%
the deepfake research community has grown significantly in the
Celeb-DF 3.723 61.0% past few years so narrower topics have also started gaining attention
DFDC 2.539 23% and interest from some researchers.
The introduction of conversational AI systems, especially
Deep Fake Detection 2.518 21.9%
ChatGPT, appeared as a game-changer for deepfake generation
UADFV 2.249 14.1% due to their high usability and accessibility. They have increasingly
DeepFake-TIMIT 2.205 12.3% been used by researchers to generate deepfake datasets although
FaceForensics++ 1.874 8.4%
their current usage is mostly limited to generating deepfake
texts. However, we believe that new image, video, audio, and
The MOS scores were not reported by Jiang et al. (2020), but calculated by us based on the
raw data shown in Table 3 of Jiang et al. (2020). hybrid deepfake datasets can be constructed with such systems,
considering the multimodal capabilities of the state-of-the-art
generative AI models, e.g., GPT-4o.58

since face swapping, face attribution manipulation, and fully

synthesised face images are among the hottest topics within 5 A meta-review of deepfake-related
deepfake research and real-world applications. We hope more non-
face deepfake image and video datasets can be produced to support
surveys
a broader range of research activities on deepfake.
This section presents a meta-review of 15 selected deepfake-
Several datasets mentioned in this survey overlooked reporting
related survey papers published in English (Lyu, 2020; Tolosana
the language(s) and/or ethnicity(ies) they cover, which could be
et al., 2020; Tong et al., 2020; Verdoliva, 2020; Younus and Hasan,
quite useful information in many applications. For those reporting
2020; Zhang T. et al., 2020; Deshmukh and Wankhade, 2021;
the covered language(s), the majority focused on English, followed
Mirsky and Lee, 2021; Nguyen et al., 2022; Rana et al., 2022; Seow
by Chinese. This indicates the need for new deepfake datasets
et al., 2022; Heidari et al., 2023; Khanjani et al., 2023; Masood
covering underrepresented, and especially low-resource, languages
et al., 2023; Sandotra and Arora, 2024). It covers the following
and ethnicities.
aspects in a systematic manner: definitions and scope, performance
The subjective quality results shown in Table 7 indicate that it
metrics, datasets, performance comparison, key challenges and
is important to check the realness of deepfake media to support any
recommendations.
performance evaluation or comparison. To ensure that the quality
The meta-review aims to draw some high-level insights
evaluation of datasets is fair, transparent and reliable, standard
for monitoring the future development of deepfake-related
procedures need defining and a common pool of qualified human
technologies and their applications.
experts should be used.
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 5.1 Definitions and scope
2019 and since 2019. Typically, datasets created before 2019 are
relatively less advanced and smaller, while those created after 2019 As we discussed in Section 1.1, among researchers, practitioners
tend to be larger, more diverse (i.e., covering more attributes), and lawmakers there is no universally accepted definition of
and of higher quality (i.e., produced by more advanced generative “deepfake” as a term. This is also reflected in how the authors of the
models). This can also be seen from the data in Table 7, in which the 15 survey papers considered this aspect. Most authors talked about
top two datasets (DeeperForensics-1 and Celeb-DF) fall within the the history of deepfakes and pointed out that the term reflects the
new generation (2020), while others belong to the old generation.
In addition to the two generations, a newer generation has also 58 https://openai.com/index/hello-gpt-4o/

Frontiers in Big Data 17 frontiersin.org

Altuncu et al. 10.3389/fdata.2024.1400024

combination of “deep learning” and “fake,” but some used a broader data generated using “deep synthesis” techniques, e.g., “deep art,”
definition, e.g., Lyu (2020) defined deepfake as “high quality fake “deep animation,” “deep music,” and “deepfake.” While authors of
videos and audios generated by AI algorithms.” Some authors also most survey papers did not recognise the positive applications of
referred to deepfake-related legislations, but none of them pointed “deepfake” technologies, Seow et al. (2022) and Sandotra and Arora
out that the definitions in some such legislations are completely (2024) covered positive applications, including entertainment,
different from the more technical definitions involving the use of business, education, art, and medicine. Other than that, some
deep learning. No authors discussed the blurred boundary between other researchers also considered such applications, e.g., organisers
deepfakes and non-deepfakes. of the Voice Conversion Challenge 2020 (see text footnote 40 )
In terms of the scope, while some authors (correctly) who said the VC technology (for speech deepfake) “is useful in
considered all types of media that can be produced by deepfake- many applications, such as customizing audio book and avatar
related techniques (Lyu, 2020; Tong et al., 2020; Rana et al., 2022; voices, dubbing, movie industry, teleconferencing, singing voice
Heidari et al., 2023; Masood et al., 2023; Sandotra and Arora, modification, voice restoration after surgery, and cloning of voices of
2024), some considered only a narrow scope, e.g., authors of historical persons.”
Tolosana et al. (2020), Younus and Hasan (2020), and Zhang
T. et al. (2020) considered only videos, authors of Verdoliva
(2020), Deshmukh and Wankhade (2021), Nguyen et al. (2022),
and Seow et al. (2022) have only considered images and videos,
5.2 Performance metrics
and Khanjani et al. (2023) only considered audio deepfakes.
Surprisingly, only two of the 15 surveys (Rana et al., 2022;
Another phenomenon we observed is that many authors focused
Heidari et al., 2023) have covered performance metrics explicitly.
more on face images and videos, and authors of three surveys
Some directly used performance metrics to explain and compare
(Tolosana et al., 2020; Younus and Hasan, 2020; Deshmukh and
the performance of covered deepfake generation and detection
Wankhade, 2021) even limited the definition of “deepfake” to such
methods. The most used performance metrics include accuracy,
a narrow scope:
ERR, and AUC. This may be explained by the page constraints
of such survey papers, which did not allow the authors to
• Deshmukh and Wankhade (2021) defined it as “a technology
extend their coverage significantly to cover performance metrics
which creates fake images or videos of targeted humans by
systematically. From this perspective, our Section 3 aims to fill
swapping their faces [by] another character saying or doing
this gap by providing a comprehensive coverage of relevant
things that are not absolutely done by them and humans start
metrics and standards, including those specific to deepfakes.
believing in such fake as it is not always recognisable with the
The subjective quality of deepfakes is an area least covered by
everyday human eye;”
the surveys, which seems related to an unbalanced coverage
• Younus and Hasan (2020) considered deepfake as a technique
of deepfake generation and deepfake detection in terms of
allowing “any computer user to exchange the face of one person
performance evaluation and comparison (the former much less
with another digitally in any video;” and
than the latter).
• Tolosana et al. (2020) defined it as “a deep learning based
technique able to create fake videos by swapping the face of a
person by the face of another person.”
5.3 Datasets
Such unnecessarily narrow definitions and scopes can lead to
confusion and do not help exchanges between researchers and Many of the 15 survey papers list a number of deepfake-related
practitioners working on different types of deepfakes. datasets, but none of them has coverage as complete as ours shown
We call on more researchers to accept a broader definition in Section 4. Firstly, none of the surveys has covered text datasets,
of “deepfake” so that highly realistic/natural media of any and only three of them (Heidari et al., 2023; Khanjani et al., 2023;
kind generated by a sophisticated automated method (often AI- Masood et al., 2023) mentioned audio datasets. When it comes
based) is considered deepfake. Here, we provide two examples to the coverage of image, video, and audio datasets, most surveys
of such a broader definition: the image2image (or pixel2pixel) only listed more popular ones, instead of a more complete coverage
technique (Zhu et al., 2017) that allows the production of deepfake of the available datasets. For instance, none of the surveys have
images and videos of any objects, and the so-called “deepfake covered the Voice Conversion Challenge 2016/2018/2020 datasets.
geography (Zhao et al., 2021),” where AI-based techniques are used In addition, more recent deepfake datasets especially those released
to generate realistic-looking satellite images. since 2021 are also not covered by any of the surveys. We believe
Another important fact missed or not sufficiently discussed by that our Section 4 is the most comprehensive review of deepfake-
authors of all the surveys, except Sandotra and Arora (2024), is related datasets so far.
that deepfake techniques can be used for positive applications, e.g., Some survey papers include datasets that are likely deepfakes,
creative arts, entertainment and protecting online users’ privacy. e.g., Verdoliva (2020) covered many general fake image datasets
We call for more researchers and practitioners to follow the where the manipulated images were not generated by deep learning
proposal in the 2020 Tencent AI White Paper (Tencent, 2020) or even AI-based methods, and some surveys mentioned ASVspoof
to start using the more neutral-sounding term “deep synthesis.” 2015 datasets but we did not see the use of deep learning for
Accordingly, we can use different words for different types of generating data used in the dataset.

Frontiers in Big Data 18 frontiersin.org

Altuncu et al. 10.3389/fdata.2024.1400024

5.4 Performance comparison and Masood et al. (2023) provided comprehensive discussions on
future trends regarding the understanding, generation, detection,
Most surveys have a good coverage of related methods for and prevention of deepfakes. Regarding preventing deepfakes,
deepfake generation and detection, but only some explicitly covered Heidari et al. (2023) and Khanjani et al. (2023) mentioned that
performance comparison between different methods (Tolosana blockchains and distributed ledger technologies can be leveraged
et al., 2020; Mirsky and Lee, 2021; Seow et al., 2022; Masood et al., for enhanced digital content traceability and identity sovereignty.
2023; Sandotra and Arora, 2024). Finally, Heidari et al. (2023) and Nguyen et al. (2022) underlined
Due to quality issues of many deepfake-related datasets the importance of considering the human aspects of deepfake
(discussed in Section 4.7), we need to treat any performance metrics detection as well as the societal impacts of deepfakes, indicating the
and comparison of different detection methods with caution. need for more interdisciplinary research on the subject.
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
6 Conclusion
more challenges, competitions and benchmarks to encourage
The rapid growth in the capability to manipulate media or
performance comparison on standard datasets and using consistent
create synthetic media which look realistic and natural paved the
performance metrics.
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
5.5 Challenges and recommendations call for the wider community to consider how to define a new
term that has a more consistent scope and meaning. For instance,
The authors of some surveys identified some key challenges and
replacing “deepfake” with “deep synthesis” can be more inclusive
future research directions for the deepfake community.
by embracing positive applications of deepfake techniques, e.g., in
Not surprisingly, how to develop more robust, scalable,
entertainment and for simulation purposes.
generalisable and explainable deepfake detection methods is one of
This paper provided a comprehensive overview of multiple
the most discussed key challenges and also a major future research
aspects of the deepfake ecosystem drawing from the research
direction (Lyu, 2020; Tong et al., 2020; Verdoliva, 2020; Younus and
literature and other online sources. It covers commonly used
Hasan, 2020; Deshmukh and Wankhade, 2021; Rana et al., 2022;
performance metrics, standards, and related datasets. It also
Heidari et al., 2023; Masood et al., 2023). Considering the arms race
presents a meta-review of 15 selected deepfake-related survey
between deepfake generation and detection, this research direction
papers published since 2020, covering not only the above-
will likely remain the hottest topic in deepfake research.
mentioned aspects but also highlighting key challenges and
Some surveys (Verdoliva, 2020; Rana et al., 2022) mentioned
recommendations.
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 Author contributions
audio-visual analysis) to achieve better performance for deepfake
detection. Lyu (2020) suggested that, for the detection of deepfake EA: Data curation, Investigation, Methodology, Validation,
videos, we need to consider video-level detection more, which can Writing – original draft, Writing – review & editing. VF:
be considered fusion of detection results of all video frames. Conceptualization, Data curation, Formal analysis, Funding
The authors of many surveys (e.g., Lyu, 2020; Younus and acquisition, Investigation, Methodology, Project administration,
Hasan, 2020; Deshmukh and Wankhade, 2021; Masood et al., 2023; Resources, Supervision, Validation, Writing – original draft,
Sandotra and Arora, 2024), argued that better (higher-quality, more Writing – review & editing. SL: Conceptualization, Data curation,
up-to-date, and more standard) deepfake datasets are needed to Formal analysis, Funding acquisition, Investigation, Methodology,
develop more effective deepfake detection methods. Lyu (2020) Project administration, Resources, Supervision, Validation,
and Masood et al. (2023) also suggested that we need to consider Writing – original draft, Writing – review & editing.
social media laundering effects in training data and improve the
evaluation of datasets. We agree with them on these points. Finally,
Rana et al. (2022) emphasised the differences in experimental Funding
settings of existing deepfake research and suggested a unique
framework to be developed for the fair evaluation of deepfake The author(s) declare that no financial support was received for
detection methods. the research, authorship, and/or publication of this article.
There are also other ad-hoc recommendations given by the
authors of some surveys. For example, Lyu (2020) argued that
deepfake detection should be considered a (more complicated) Conflict of interest
multi-class, multi-label and local detection problem. Tolosana et al.
(2020) discussed specific research directions for different deepfake The authors declare that the research was conducted in the
generation methods (face synthesis, identity swap, attribute absence of any commercial or financial relationships that could be
manipulation, and expression swap). Similarly, Heidari et al. (2023) construed as a potential conflict of interest.

Frontiers in Big Data 19 frontiersin.org

Altuncu et al. 10.3389/fdata.2024.1400024

Publisher’s note organizations, or those of the publisher, the editors and the
reviewers. Any product that may be evaluated in this article, or
All claims expressed in this article are solely those of the claim that may be made by its manufacturer, is not guaranteed or
authors and do not necessarily represent those of their affiliated endorsed by the publisher.

References
Afchar, D., Nozick, V., Yamagishi, J., and Echizen, I. (2018). “MesoNet: a Dolhansky, B., Bitton, J., Pflaum, B., Lu, J., Howes, R., Wang, M., et al. (2020). The
compact facial video forgery detection network,” in Proceedings of the 2018 DeepFake detection challenge dataset. arXiv preprint arXiv:2006.07397.
IEEE International Workshop on Information Forensics and Security (IEEE), 1–7.
Dufour, N., and Gully, A. (2019). Contributing data to deepfake detection
doi: 10.1109/WIFS.2018.8630761
research. Technical report, Google AI. Available at: https://ai.googleblog.com/2019/09/
Ajder, H., Patrini, G., Cavalli, F., and Cullen, L. (2019). The state of deepfakes: contributing-data-to-deepfake-detection.html (accessed March, 2024)
Landscape, threats, and impact. Technical report, Deeptrace. Available at: https://
Durall, R., Keuper, M., Pfreundt, F.-J., and Keuper, J. (2019). Unmasking deepfakes
sensity.ai/reports/ (accessed March, 2024)
with simple features. arXiv:1911.00686.
Akhtar, Z., and Falk, T. H. (2017). Audio-visual multimedia quality
Fagni, T., Falchi, F., Gambini, M., Martella, A., and Tesconi, M. (2021a).
assessment: a comprehensive survey. IEEE Access 5, 21090–21117.
TweepFake: about detecting deepfake tweets. PLoS ONE 16:e0251415.
doi: 10.1109/ACCESS.2017.2750918
doi: 10.1371/journal.pone.0251415
Alshammari, H., and EI-Sayed, A. (2023b). AIRABIC: Arabic dataset for
Fagni, T., Falchi, F., Gambini, M., Martella, A., and Tesconi, M. (2021b). TweepFake:
performance evaluation of ai detectors. GitHub dataset. Available at: https://github.com/
about detecting deepfake tweets. GitHub dataset. Available at: https://github.com/tizfa/
Hamed1Hamed/AIRABIC (accessed July, 2024).
tweepfake_deepfake_text_detection (accessed July, 2024).
Alshammari, H., and Sayed, A. (2023a). AIRABIC: Arabic dataset for
Fox, G., Liu, W., Kim, H., Seidel, H.-P., Elgharib, M., and Theobalt, C. (2021).
performance evaluation of ai detectors,” in Proceedings of the 2023 International
“Videoforensicshq: detecting high-quality manipulated face videos,” in Proceedings
Conference on Machine Learning and Applications (ICMLA) (IEEE), 864–870.
of the 2021 IEEE International Conference on Multimedia and Expo (IEEE), 1–6.
doi: 10.1109/ICMLA58977.2023.00127
doi: 10.1109/ICME51207.2021.9428101
Ba, Z., Wen, Q., Cheng, P., Wang, Y., Lin, F., Lu, L., et al. (2023a). DEepfake CROss-
lingual (DECRO) evaluation dataset. GitHub dataset. Available at: https://github.com/ Frank, J., and Schönherr, L. (2021a). “WaveFake: a data set to facilitate audio
petrichorwq/DECRO-dataset (accessed March, 2024) deepfake detection,” in Proceedings of the 35th Conference on Neural Information
Processing Systems (NeurIPS 2021) Track on Datasets and Benchmarks, 1–17.
Ba, Z., Wen, Q., Cheng, P., Wang, Y., Lin, F., Lu, L., et al. (2023b). Transferring
audio deepfake detection capability across languages,” in Proceedings of the ACM Web Frank, J., and Schönherr, L. (2021b). WaveFake: a data set to facilitate audio deepfake
Conference 2023 (ACM), 2033–2044. doi: 10.1145/3543507.3583222 detection. GitHub dataset. Available at: https://github.com/RUB-SysSec/WaveFake

Bandi, A., Adapa, P. V. S. R., and Kuchi, Y. E. V. P. K. (2023). The power of Gong, Y., Yang, J., Huber, J., MacKnight, M., and Poellabauer, C. (2019a). ReMASC:
generative AI: A review of requirements, models, input-output formats, evaluation realistic replay attack corpus for voice controlled systems,” in Proceedings of Interspeech
metrics, and challenges. Fut. Internet 15:260. doi: 10.3390/fi15080260 2019, 2355–2359. doi: 10.21437/Interspeech.2019-1541

Brady, M. (2020). Deepfakes: a new desinformation threat? Technical report, Gong, Y., Yang, J., Huber, J., MacKnight, M., and Poellabauer, C. (2019b). ReMASC:
Democracy Reporting International. Available at: https://democracy-reporting.org/ realistic replay attack corpus for voice controlled systems. GitHub dataset. Available at:
dri_publications/deepfakes-a-new-disinformation-threat/ (accessed March, 2024) https://github.com/YuanGongND/ReMASC (accessed July, 2024).
Cai, Z., Ghosh, S., Adatia, A. P., Hayat, M., Dhall, A., and Stefanov, K. Guo, B., Zhang, X., Wang, Z., Jiang, M., Nie, J., Ding, Y., et al. (2023a). How
(2023a). AV-Deepfake1M: A large-scale LLM-driven audio-visual deepfake dataset. close is ChatGPT to human experts? comparison corpus, evaluation, and detection.
arXiv:2311.15308. arXiv:2301.07597.
Cai, Z., Ghosh, S., Adatia, A. P., Hayat, M., Dhall, A., and Stefanov, K. (2023b). AV- Guo, B., Zhang, X., Wang, Z., Jiang, M., Nie, J., Ding, Y., et al. (2023b). Human
Deepfake1M: a large-scale LLM-driven audio-visual deepfake dataset. GitHub dataset. ChatGPT Comparison Corpus (HC3). GitHub dataset. Available at: https://github.com/
Available at: https://github.com/ControlNet/AV-Deepfake1M (accessed July, 2024). Hello-SimpleAI/chatgpt-comparison-detection (accessed July, 2024)
Chefer, H., Alaluf, Y., Vinker, Y., Wolf, L., and Cohen-Or, D. (2023). Attend-and- Hartvigsen, T., Gabriel, S., Palangi, H., Sap, M., Ray, D., and Kamar, E. (2022).
Excite: attention-based semantic guidance for text-to-image diffusion models. ACM “ToxiGen: a large-scale machine-generated dataset for adversarial and implicit hate
Trans. Graph. 42, 1–10. doi: 10.1145/3592116 speech detection,” in Proceedings of the 60th Annual Meeting of the Association for
Computational Linguistics (ACL), 3309–3326. doi: 10.18653/v1/2022.acl-long.234
Chen, Y., Liu, L., and Ding, C. (2023). X-IQE: eXplainable image quality evaluation
for text-to-image generation with visual large language models. arXiv:2305.10843. He, Y., Gan, B., Chen, S., Zhou, Y., Yin, G., Song, L., et al. (2021). “ForgeryNet:
a versatile benchmark for comprehensive forgery analysis,” in Proceedings of the 2021
Ciftci, U. A., Demir, I., and Yin, L. (2020). “FakeCatcher: Detection of synthetic
IEEE/CVF Conference on Computer Vision and Pattern Recognition (IEEE), 4360–4369.
portrait videos using biological signals,” in IEEE Transactions on Pattern Analysis and
doi: 10.1109/CVPR46437.2021.00434
Machine Intelligence.
Heidari, A., Jafari Navimipour, N., Dag, H., and Unal, M. (2023). Deepfake detection
Dang, H., Liu, F., Stehouwer, J., Liu, X., and Jain, A. K. (2020). “On the
using deep learning methods: a systematic and comprehensive review. WIREs Data
detection of digital face manipulation,” in Proceedings of the 2020 IEEE/CVF
Mining Knowl. Discov. 45:e1520. doi: 10.1002/widm.1520
Conference on Computer Vision and Pattern Recognition (IEEE), 5781–5790.
doi: 10.1109/CVPR42600.2020.00582 Hessel, J., Holtzman, A., Forbes, M., Le Bras, R., and Choi, Y. (2021). “CLIPScore:
a reference-free evaluation metric for image captioning,” in Proceedings of the 2021
Delgado, H., Evans, N., Kinnunen, T., Lee, K. A., Liu, X., Nautsch, A., et al. (2021a).
Conference on Empirical Methods in Natural Language Processing (Association for
ASVspoof 2021 challenge - logical access database. Zenodo dataset.
Computational Linguistics), 7514–7528. doi: 10.18653/v1/2021.emnlp-main.595
Delgado, H., Evans, N., Kinnunen, T., Lee, K. A., Liu, X., Nautsch, A., et al. (2021b).
Huang, K., Sun, K., Xie, E., Li, Z., and Liu, X. (2023). “T2I-CompBench: a
ASVspoof 2021 challenge - speech deepfake database. Zenodo dataset.
comprehensive benchmark for open-world compositional text-to-image generation,”
Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. (2009). in Proceedings of the 37th Neural Information Processing Systems Track on Datasets and
“ImageNet: a large-scale hierarchical image database,” in Proceedings of the 2009 Benchmarks (NeurIPS ’23) (Curran Associates, Inc.), 78723–78747.
IEEE Conference on Computer Vision and Pattern Recognition (IEEE), 248–255.
Jia, S., Li, X., and Lyu, S. (2022a). DFDM: Deepfakes from different models. GitHub
doi: 10.1109/CVPR.2009.5206848
dataset. Available at: https://github.com/shanface33/Deepfake_Model_Attribution
Deshmukh, A., and Wankhade, S. B. (2021). “Deepfake detection approaches (accessed March, 2024)
using deep learning: A systematic review,” in Intelligent Computing and Networking:
Jia, S., Li, X., and Lyu, S. (2022b). Model attribution of face-swap deepfake videos.
Proceedings of IC-ICN 2020, volume 146 of Lecture Notes in Networks and Systems
arXiv:2202.12951. doi: 10.1109/ICIP46576.2022.9897972
(Springer), 293–302. doi: 10.1007/978-981-15-7421-4_27
Jiang, L., Li, R., Wu, W., Qian, C., and Loy, C. C. (2020). “DeeperForensics-1.0:
Ding, X., Raziei, Z., Larson, E. C., Olinick, E. V., Krueger, P., and Hahsler, M. (2020).
a large-scale dataset for real-world face forgery detection,” in Proceedings of the 2020
Swapped face detection using deep learning and subjective assessment. EURASIP J. Inf.
IEEE/CVF Conference on Computer Vision and Pattern Recognition (IEEE), 2886–2895.
Secur. 2020, 1–12. doi: 10.1186/s13635-020-00109-8
doi: 10.1109/CVPR42600.2020.00296

Frontiers in Big Data 20 frontiersin.org

Altuncu et al. 10.3389/fdata.2024.1400024

Kalchbrenner, N., Elsen, E., Simonyan, K., Noury, S., Casagrande, N., Lockhart, E., Masood, M., Nawaz, M., Malik, K. M., Javed, A., Irtaza, A., and Malik,
et al. (2018). Efficient neural audio synthesis. arXiv:1802.08435. H. (2023). Deepfakes generation and detection: State-of-the-art, open
challenges, countermeasures, and way forward. Appl. Intell. 53, 3974–4026.
Karras, T., Laine, S., and Aila, T. (2019). “A style-based generator architecture
doi: 10.1007/s10489-022-03766-z
for generative adversarial networks,” in Proceedings of the 2019 IEEE/CVF
Conference on Computer Vision and Pattern Recognition (IEEE), 4401–4410. Mirsky, Y., and Lee, W. (2021). The creation and detection of deepfakes: a survey.
doi: 10.1109/CVPR.2019.00453 ACM Comput. Surv. 54, 1–41. doi: 10.1145/3425780
Khanjani, Z., Watson, G., and Janeja, V. P. (2023). Audio deepfakes: a survey. Front. Mosca, E., Abdalla, M. H. I., Basso, P., Musumeci, M., and Groh, G. (2023).
Big Data 5:1001063. doi: 10.3389/fdata.2022.1001063 “Distinguishing fact from fiction: a benchmark dataset for identifying machine-
generated scientific papers in the LLM era,” in Proceedings of the 3rd Workshop
Khodabakhsh, A., Ramachandra, R., Raja, K., Wasnik, P., and Busch, C. (2018).
on Trustworthy Natural Language Processing (TrustNLP 2023) (ACL), 190–207.
“Fake face detection methods: can they be generalized?” in Proceedings of the
doi: 10.18653/v1/2023.trustnlp-1.17
2018 International Conference of the Biometrics Special Interest Group (IEEE), 1–6.
doi: 10.23919/BIOSIG.2018.8553251 Müller, N., Czempin, P., Diekmann, F., Froghyar, A., and Böttinger, K. (2022). “Does
audio deepfake detection generalize?,” in Proceedings of Interspeech 2022, 2783–2787.
Kim, H., Elgharib, M., Zollhöfer, M., Seidel, H. P., Beeler, T., Richardt, C., et
doi: 10.21437/Interspeech.2022-108
al. (2019). Neural style-preserving visual dubbing. ACM Trans. Graph. 38, 1–13.
doi: 10.1145/3355089.3356500 Müller, N. M., Kawa, P., Choong, W. H., Casanova, E., Gölge, E., Müller, T., et al.
(2024). MLAAD: The multi-language audio anti-spoofing dataset. arXiv:2401.09512.
Kim, H., Garrido, P., Tewari, A., Xu, W., Thies, J., Niessner, M., et al. (2018). Deep
video portraits. ACM Trans. Graph. 37, 1–14. doi: 10.1145/3197517.3201283 Mysore, G. J. (2015). Can we automatically transform speech recorded on common
consumer devices in real-world environments into professional production quality
Korshunov, P., and Marcel, S. (2019). “Vulnerability assessment and detection of
speech?–a dataset, insights, and challenges. IEEE Signal Proc. Lett. 22, 1006–1010.
deepfake videos,” in Proceedings of the 2019 International Conference on Biometrics
doi: 10.1109/LSP.2014.2379648
(IEEE), 1–6. doi: 10.1109/ICB45273.2019.8987375
Narayan, K., Agarwal, H., Thakral, K., Mittal, S., Vatsa, M., and Singh, R. (2023a).
Kwon, P., You, J., Nam, G., Park, S., and Chae, G. (2021). “KoDF: A
Df-Platter database. Public dataset. Available at: https://iab-rubric.org/df-platter-
large-scale korean DeepFake detection dataset,” in Proceedings of the 2021
database (accessed March, 2024)
IEEE/CVF International Conference on Computer Vision (IEEE), 10724–10733.
doi: 10.1109/ICCV48922.2021.01057 Narayan, K., Agarwal, H., Thakral, K., Mittal, S., Vatsa, M., and Singh, R. (2023b).
DF-Platter: multi-face heterogeneous deepfake dataset,” in Proceedings of the 2023
Li, D., Ling, H., Kim, S. W., Kreis, K., Fidler, S., and Torralba, A. (2022).
IEEE/CVF Conference on Computer Vision and Pattern Recognition (IEEE), 9739–9748.
“BigDatasetGAN: Synthesizing imagenet with pixel-wise annotations,” in Proceedings
doi: 10.1109/CVPR52729.2023.00939
of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
(IEEE), 21330–21340. doi: 10.1109/CVPR52688.2022.02064 Neves, J. C., Tolosana, R., Vera-Rodriguez, R., Lopes, V., Proena, H., and Fierrez,
J. (2020). GANprintR: improved fakes and evaluation of the state of the art in
Li, G., Zhao, X., Cao, Y., Pei, P., Li, J., and Zhang, Z. (2022a). “FMFCC-V: an
face manipulation detection. IEEE J. Select. Topics Signal Proc. 14, 1038–1048.
Asian large-scale challenging dataset for deepfake detection,” in Proceedings of the
doi: 10.1109/JSTSP.2020.3007250
2022 ACM Workshop on Information Hiding and Multimedia Security (ACM), 7–18.
doi: 10.1145/3531536.3532946 Nguyen, T. T., Nguyen, Q. V. H., Nguyen, D. T., Nguyen, D. T., Huynh-The,
T., Nahavandi, S., et al. (2022). Deep learning for deepfakes creation and detection:
Li, G., Zhao, X., Cao, Y., Pei, P., Li, J., and Zhang, Z. (2022b). FMFCC-V: an
a survey. Comput. Vis. Image Understand. 223:103525. doi: 10.1016/j.cviu.2022.
Asian large-scale challenging dataset for deepfake detection. GitHub dataset. Available
103525
at: https://github.com/iiecasligen/FMFCC-V (accessed July, 2024).
Ni, B., Peng, H., Chen, M., Zhang, S., Meng, G., Fu, J., et al. (2022). “Expanding
Li, J., Li, D., Xiong, C., and Hoi, S. (2022). “BLIP: bootstrapping language-image
language-image pretrained models for general video recognition,” in Proceedings of
pre-training for unified vision-language understanding and generation,” in Proceedings
the 17th European Conference on Computer Vision (ECCV ’22) (Springer Nature
of the 39th International Conference on Machine Learning (PMLR), 12888–12900.
Switzerland), 1–18. doi: 10.1007/978-3-031-19772-7_1
Li, L., Bao, J., Yang, H., Chen, D., and Wen, F. (2020). “Advancing high
Nirkin, Y., Keller, Y., and Hassner, T. (2019). “FSGAN: subject agnostic face
fidelity identity swapping for forgery detection,” in Proceedings of the 2020
swapping and reenactment,” in Proceedings of the 2019 IEEE/CVF International
IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5073–5082.
Conference on Computer Vision (IEEE), 7183–7192. doi: 10.1109/ICCV.2019.00728
doi: 10.1109/CVPR42600.2020.00512
OpenAI (2019). GPT-2-output-dataset: dataset of GPT-2 outputs for research in
Li, Y., Chang, M.-C., and Lyu, S. (2018). “In ICTU OCULI: exposing AI
detection, biases, and more. GitHub dataset. Available at: https://github.com/openai/
created fake videos by detecting eye blinking,” in Proceedings of the 2018
gpt-2-output-dataset/ (accessed March, 2024)
IEEE International Workshop on Information Forensics and Security (IEEE), 1–7.
doi: 10.1109/WIFS.2018.8630787 Pal, D., and Triyason, T. (2018). A survey of standardized approaches towards the
quality of experience evaluation for video services: an ITU perspective. Int. J. Dig.
Li, Y., Li, Q., Cui, L., Bi, W., Wang, Z., Wang, L., et al. (2024). MAGE: machine-
Multimedia Broadcast. 2018:1391724. doi: 10.1155/2018/1391724
generated text detection in the wild. arXiv:2305.13242.
Li, Y., Yang, X., Sun, P., Qi, H., and Lyu, S. (2020). “Celeb-DF: a large- Pu, J., Mangaokar, N., Kelly, L., Bhattacharya, P., Sundaram, K., Javed, M., et al.
scale challenging dataset for deepfake forensics,” in Proceedings of the 2020 (2021a). “Deepfake videos in the wild: analysis and detection,” in Proceedings of the
IEEE/CVF Conference on Computer Vision and Pattern Recognition (IEEE), 3204–3213. Web Conference 2021 (ACM), 981–992. doi: 10.1145/3442381.3449978
doi: 10.1109/CVPR42600.2020.00327 Pu, J., Mangaokar, N., Kelly, L., Bhattacharya, P., Sundaram, K., Javed, M., et al.
Liu, X., Wang, X., Sahidullah, M., Patino, J., Delgado, H., Kinnunen, T., et al. (2023). (2021b). DF-W: a new deepfake dataset comprising of deepfake videos created and shared
ASVspoof 2021: towards spoofed and deepfake speech detection in the wild. IEEE/ACM by the internet community. GitHub dataset. Available at: https://github.com/jmpu/
Trans. Audio, Speech Lang. Proc. 31, 2507–2522. doi: 10.1109/TASLP.2023.3285283 webconf21-deepfakes-in-the-wild (accessed March, 2024)
Livingstone, S. R., and Russo, F. A. (2018). The ryerson audio-visual database Pu, J., Sarwar, Z., Abdullah, S. M., Rehman, A., Kim, Y., Bhattacharya, P., et
of emotional speech and song (RAVDESS): a dynamic, multimodal set of al. (2023). “Deepfake text detection: Limitations and opportunities,” in Proceedings
facial and vocal expressions in north american english. PLoS ONE 13:e0196391. of the 2023 IEEE Symposium on Security and Privacy (SP) (IEEE), 1613–1630.
doi: 10.1371/journal.pone.0196391 doi: 10.1109/SP46215.2023.10179387
Liyanage, V., Buscaldi, D., and Nazarenko, A. (2022a). A benchmark corpus for the
Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., et al.
detection of automatically generated text in academic publications. arXiv:2202.02013.
(2021). “Learning transferable visual models from natural language supervision,”
Liyanage, V., Buscaldi, D., and Nazarenko, A. (2022b). GeneratedTextDetection. in Proceedings of the 38th International Conference on Machine Learning (PMLR),
GitHub dataset. Available at: https://github.com/vijini/GeneratedTextDetection 8748–8763.
(accessed July, 2024)
Rana, M. S., Nobi, M. N., Murali, B., and Sung, A. H. (2022). Deepfake
Lorenzo-Trueba, J., Yamagishi, J., Toda, T., Saito, D., Villavicencio, F., Kinnunen, T.,
detection: a systematic literature review. IEEE Access 10, 25494–25513.
et al. (2018). “The voice conversion challenge 2018: Promoting development of parallel
doi: 10.1109/ACCESS.2022.3154404
and nonparallel methods,” in Proceedings of the Odyssey 2018 The Speaker and Language
Recognition Workshop (International Speech Communication Association), 195–202. Rössler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies, J., and Nießner, M. (2018).
doi: 10.21437/Odyssey.2018-28 FaceForensics: a large-scale video dataset for forgery detection in human faces. arXiv
Lyu, S. (2020). “Deepfake detection: Current challenges and next steps,” in preprint arXiv:1803.09179.
Proceedings of the 2020 IEEE International Conference on Multimedia Expo Workshops Rössler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies, J., and Nießner,
(IEEE) doi: 10.1109/ICMEW46912.2020.9105991 M. (2019). “FaceForensics++: learning to detect manipulated facial images,” in
Ma, H., Yi, J., Wang, C., Yan, X., Tao, J., Wang, T., et al. (2023a). CFAD: a Chinese Proceedings of the 2019 International Conference on Computer Vision (IEEE), 1–11.
dataset for fake audio detection. arXiv:2207.12308. doi: 10.2139/ssrn.4748856 doi: 10.1109/ICCV.2019.00009
Ma, H., Yi, J., Wang, C., Yan, X., Tao, J., Wang, T., et al. (2023b). CFAD: a Rothe, R., Timofte, R., and Van Gool, L. (2015). “DEX: deep expectation of apparent
Chinese dataset for fake audio detection. GitHub dataset. Available at: https://github. age from a single image,” in Proceedings of the 2015 IEEE International Conference on
com/ADDchallenge/CFAD (accessed July, 2024). Computer Vision Workshop (IEEE), 252–257. doi: 10.1109/ICCVW.2015.41

Frontiers in Big Data 21 frontiersin.org

Altuncu et al. 10.3389/fdata.2024.1400024

Sandotra, N., and Arora, B. (2024). A comprehensive evaluation of feature- Xie, Y., Zhou, J., Lu, X., Jiang, Z., Yang, Y., Cheng, H., et al. (2023a). FSD: An initial
based AI techniques for deepfake detection. Neural Comput. Applic. 36, 3859–3887. chinese dataset for fake song detection. GitHub dataset. Available at: https://github.
doi: 10.1007/s00521-023-09288-0 com/xieyuankun/FSD-Dataset (accessed July, 2024)
Sensity (2024). The state of deepfakes 2024. Technical report, Sensity. Available at: Xie, Y., Zhou, J., Lu, X., Jiang, Z., Yang, Y., Cheng, H., et al. (2023b). FSD: an initial
https://sensity.ai/reports/ (accessed July, 2024) chinese dataset for fake song detection. GitHub dataset. Available at: https://github.com/
xieyuankun/FSD-Dataset (accessed July, 2024).
Seow, J. W., Lim, M. K., Phan, R. C., and Liu, J. K. (2022). A comprehensive overview
of Deepfake: generation, detection, datasets, and opportunities. Neurocomputing 513, Yariv, G., Gat, I., Benaim, S., Wolf, L., Schwartz, I., and Adi, Y. (2024). “Diverse and
351–371. doi: 10.1016/j.neucom.2022.09.135 aligned audio-to-video generation via text-to-video model adaptation,” in Proceedings
of the 38th AAAI Conference on Artificial Intelligence (AAAI ’24) (AAAI), 6639–6647.
Song, H., Huang, S., Dong, Y., and Tu, W.-W. (2023a). DeepFakeFace. GitHub
doi: 10.1609/aaai.v38i7.28486
dataset. Available at: https://github.com/OpenRL-Lab/DeepFakeFace (accessed March,
2024) Yi, J., Bai, Y., Tao, J., Ma, H., Tian, Z., Wang, C., et al. (2021). “Half-Truth: a partially
fake audio detection dataset,” in Proceedings of Interspeech 2021, 1654–1658.
Song, H., Huang, S., Dong, Y., and Tu, W.-W. (2023b). Robustness and
generalizability of deepfake detection: a study with diffusion models. arXiv:2309.02218. Yi, J., Fu, R., Tao, J., Nie, S., Ma, H., Wang, C., et al. (2022a). “ADD 2022: the first
audio deep synthesis detection challenge,” in Proceedings of the 2022 IEEE International
Su, Z., Li, M., Zhang, G., Wu, Q., Li, M., Zhang, W., et al. (2023a). CMFD. GitHub
Conference on Acoustics, Speech and Signal Processing (ICASSP) (IEEE), 9216–9220.
dataset. Available at: https://github.com/WuQinfang/CMFD (accessed March, 2024)
Yi, J., Tao, J., Fu, R., Yan, X., Wang, C., Wang, T., et al. (2023). “ADD 2023:
Su, Z., Li, M., Zhang, G., Wu, Q., Li, M., Zhang, W., et al. (2023b). Robust
the second audio deepfake detection challenge,” in Proceedings of the Workshop on
audio copy-move forgery detection using constant q spectral sketches and GA-SVM.
Deepfake Audio Detection and Analysis (CEUR Workshop Proceedings), 125–130.
IEEE Trans. Depend. Secure Comput. 20, 4016–4031. doi: 10.1109/TDSC.2022.32
15280 Yi, J., Wang, C., Tao, J., Tian, Z., Fan, C., Ma, H., et al. (2022b). SceneFake: an initial
dataset and benchmarks for scene fake audio detection. GitHub dataset. Available at:
Su, Z., Wu, X., Zhou, W., Ma, G., and Hu, S. (2024). HC3 Plus: a semantic-invariant
https://github.com/ADDchallenge/SceneFake (accessed July, 2024)
human ChatGPT comparison corpus. arXiv:2309.02731.
Yi, J., Wang, C., Tao, J., Zhang, C. Y., Fan, C., Tian, Z., et al. (2024). SceneFake: an
Tanaka, K., Kameoka, H., Kaneko, T., and Hojo, N. (2019). WaveCycleGAN2:
initial dataset and benchmarks for scene fake audio detection. Patt. Recogn. 152:110468.
time-domain neural post-filter for speech waveform generation. arXiv:1904.02892.
doi: 10.1016/j.patcog.2024.110468
Temnikova, I., Marinova, I., Gargova, S., Margova, R., and Koychev, I.
Yi, Z., Huang, W.-C., Tian, X., Yamagishi, J., Das, R. K., Kinnunen, T., et al. (2020).
(2023). “Looking for traces of textual deepfakes in Bulgarian on social media,”
“Voice conversion challenge 2020-intra-lingual semi-parallel and cross-lingual voice
in Proceedings of the 14th International Conference on Recent Advances in
conversion.,” in Proceedings of the Joint Workshop for the Blizzard Challenge and Voice
Natural Language Processing (Shoumen, Bulgaria: INCOMA Ltd.), 1151–1161.
Conversion Challenge 2020 (International Speech Communication Association), 80–98.
doi: 10.26615/978-954-452-092-2_122
doi: 10.21437/VCCBC.2020-14
Tencent (2020). Artificial intelligence white paper. Technical report, Tencent.
Younus, M. A., and Hasan, T. M. (2020). “Abbreviated view of deepfake videos
Available at: https://tech.sina.com.cn/roll/2020-07-14/doc-iivhvpwx5201226.shtml
detection techniques,” in Proceedings of the 2020 6th International Engineering
(accessed March, 2024)
Conference (IEEE), 115–120. doi: 10.1109/IEC49899.2020.9122916
Toda, T., Chen, L.-H., Saito, D., Villavicencio, F., Wester, M., Wu, Z.,
Yu, P., Chen, J., Feng, X., and Xia, Z. (2024a). CHEAT. GitHub dataset. Available at:
et al. (2016). “The voice conversion challenge 2016,” in Proceedings of
https://github.com/botianzhe/CHEAT (accessed July, 2024)
Interspeech 2016 (International Speech Communication Association), 1632–1636.
doi: 10.21437/Interspeech.2016-1066 Yu, P., Chen, J., Feng, X., and Xia, Z. (2024b). CHEAT: A large-scale dataset for
detecting ChatGPT-written abstracts. arXiv:2304.12008.
Tolosana, R., Vera-Rodriguez, R., Fierrez, J., Morales, A., and Ortega-Garcia, J.
(2020). Deepfakes and beyond: a survey of face manipulation and fake detection. Inf. Zhai, G., and Min, X. (2020). Perceptual image quality assessment: a survey. Sci.
Fusion 64, 131–148. doi: 10.1016/j.inffus.2020.06.014 China Inf. Sci. 63, 1–52. doi: 10.1007/s11432-019-2757-1
Tong, X., Wang, L., Pan, X., and Wang, J. (2020). “An overview of Zhang, L., Wang, X., Cooper, E., Yamagishi, J., Patino, J., and Evans, N. (2021a). “An
deepfake: the sword of Damocles in AI,” in Proceedings of the 2020 International initial investigation for detecting partially spoofed audio,” in Proceedings of Interspeech
Conference on Computer Vision, Image and Deep Learning (IEEE), 265–273. 2021, 4264–4268. doi: 10.21437/Interspeech.2021-738
doi: 10.1109/CVIDL51233.2020.00-88
Zhang, L., Wang, X., Cooper, E., Yamagishi, J., Patino, J., and Evans, N.
Uchendu, A., Ma, Z., Le, T., Zhang, R., and Lee, D. (2021). “TURINGBENCH: (2021b). PartialSpoof. Zenodo dataset. Available at: https://zenodo.org/records/
a benchmark environment for Turing test in the age of neural text generation,” 5766198 (accessed March, 2024)
in Findings of the Association for Computational Linguistics: EMNLP 2021 (ACL),
Zhang, T., Deng, L., Zhang, L., and Dang, X. (2020). “Deep learning in face
2001–2016. doi: 10.18653/v1/2021.findings-emnlp.172
synthesis: a survey on deepfakes,” in Proceedings of the 2020 IEEE 3rd International
Van Den Oord, A., Dieleman, S., Zen, H., Simonyan, K., Vinyals, O., Graves, A., et Conference on Computer and Communication Engineering Technology (IEEE), 67–70.
al. (2016). WaveNet: a generative model for raw audio. arXiv:1609.03499. doi: 10.1109/CCET50901.2020.9213159
Verdoliva, L. (2020). Media forensics and deepfakes: an overview. IEEE J. Selected Zhang, Y., Ling, H., Gao, J., Yin, K., Lafleche, J.-F., Barriuso, A., et al. (2021).
Topics Signal Proc. 14, 910–932. doi: 10.1109/JSTSP.2020.3002101 “DatasetGAN: Efficient labeled data factory with minimal human effort,” in Proceedings
of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (IEEE),
Wang, X., Yamagishi, J., Todisco, M., Delgado, H., Nautsch, A., Evans, N., et al.
10140–10150. doi: 10.1109/CVPR46437.2021.01001
(2020). ASVspoof 2019: A large-scale public database of synthesized, converted and
replayed speech. Comput. Speech Lang.64:101114. doi: 10.1016/j.csl.2020.101114 Zhang, Y., Yin, Z., Li, Y., Yin, G., Yan, J., Shao, J., et al. (2020). “CelebA-
Spoof: large-scale face anti-spoofing dataset with rich annotations,” in Proceedings
Wang, Z., Bao, J., Zhou, W., Wang, W., Hu, H., Chen, H., et al. (2023a).
of the 2020 European Conference on Computer Vision (Springer), 70–85.
“DIRE for diffusion-generated image detection,” in Proceedings of the 2023
doi: 10.1007/978-3-030-58610-2_5
IEEE/CVF International Conference on Computer Vision (IEEE), 22388–22398.
doi: 10.1109/ICCV51070.2023.02051 Zhao, B., Zhang, S., Xu, C., Sun, Y., and Deng, C. (2021). Deep fake geography?
When geospatial data encounter artificial intelligence. Cartogr. Geogr. Inf. Sci. 48,
Wang, Z., Bao, J., Zhou, W., Wang, W., Hu, H., Chen, H., et al. (2023b). DIRE for
338–352. doi: 10.1080/15230406.2021.1910075
diffusion-generated image detection. GitHub dataset. Available at: https://github.com/
ZhendongWang6/DIRE (accessed July, 2024). Zhao, Y., Yi, J., Tao, J., Wang, C., Zhang, X., and Dong, Y. (2023a). EmoFake: an
initial dataset for emotion fake audio detection. arXiv:2211.05363.
Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., et al.
(2022). “Chain-of-Thought prompting elicits reasoning in large language models,” in Zhao, Y., Yi, J., Tao, J., Wang, C., Zhang, X., and Dong, Y. (2023b). EmoFake: an
Proceedings of the 36th Neural Information Processing Systems (NeurIPS ’22) (Curran initial dataset for emotion fake audio detection. GitHub dataset. Available at: https://
Associates, Inc.), 24824–24837. github.com/zy511361103/GADE (accessed July, 2024).
Wu, J. Z., Fang, G., Wu, H., Wang, X., Ge, Y., Cun, X., et al. (2024). Towards a better Zhou, P., Han, X., Morariu, V. I., and Davis, L. S. (2017). “Two-stream
metric for text-to-video generation. arXiv:2401.07781. neural networks for tampered face detection,” in Proceedings of the 2017 IEEE
Conference on Computer Vision and Pattern Recognition Workshops (IEEE), 1831–1839.
Wu, W., Zhao, Y., Chen, H., Gu, Y., Zhao, R., He, Y., et al. (2023). “DatasetDM:
doi: 10.1109/CVPRW.2017.229
synthesizing data with perception annotations using diffusion models,” in Proceedings
of the 37th International Conference on Neural Information Processing Systems (Curran Zhou, T., Wang, W., Liang, Z., and Shen, J. (2021). “Face forensics in the wild,”
Associates, Inc.), 54683–54695. in Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern
Recognition (IEEE), 5774–5784. doi: 10.1109/CVPR46437.2021.00572
Xidian University, Modern Image Processing Lab (2024). Homologous deepfake
dataset: A self built small-scale, high-quality, and diverse deepfake dataset. GitHub Zhu, D., Chen, J., Shen, X., Li, X., and Elhoseiny, M. (2024). “MiniGPT-4: enhancing
dataset. Available at: https://github.com/mirro-yyf/Homologous_deepfake_dataset vision-language understanding with advanced large language models,” in Proceedings of
(accessed July, 2024) the 12th International Conference on Learning Representations.

Frontiers in Big Data 22 frontiersin.org

Altuncu et al. 10.3389/fdata.2024.1400024

Zhu, J.-Y., Park, T., Isola, P., and Efros, A. A. (2017). “Unpaired image-to- Zi, B., Chang, M., Chen, J., Ma, X., and Jiang, Y.-G. (2020). “WildDeepfake:
image translation using cycle-consistent adversarial networks,” in Proceedings of a challenging real-world dataset for deepfake detection,” in Proceedings
the 2017 IEEE International Conference on Computer Vision (IEEE), 2242–2251. of the 2020 28th ACM International Conference on Multimedia (ACM),
doi: 10.1109/ICCV.2017.244 2382–2390.

Frontiers in Big Data 23 frontiersin.org