Applied Intelligence (2025) 55:239

https://doi.org/10.1007/s10489-024-05997-8

Video authentication detection using deep learning: a systematic

literature review
Ayat Abd‑Muti Alrawahneh1 · Sharifah Nurul Asyikin Syed Abdullah2 · Siti Norul Huda Sheikh Abdullah1 ·
Nazhatul Hafizah Kamarudin1 · Sarah Khadijah Taylor2

Accepted: 30 September 2024

© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024

Abstract
Recent advancements in deep learning have notably influenced research across various data types, with a significant focus
on video authentication. This area has emerged as a crucial aspect of ensuring the integrity and trustworthiness of video
content amidst growing concerns over manipulation and falsification. It is emerging as a field ripe for exploration. This
paper presents a systematic literature review (SLR) on using deep learning techniques for video authentication, addressing
the urgent need for robust methods to verify video integrity amidst increasing manipulation threats. Reviewing literature
from the past five years, this SLR reviews 99 research articles from the last five years and highlights the significant progress
made through deep learning techniques (Convolution Neural Network (CNN), Recurrent Neural Network (RNN), Deep
Neural Network (DNN), and Generative Adversarial Networks (GANs)). It aims to investigate applications, techniques,
datasets, and challenges in video authentication, providing a comprehensive guide for researchers. This study encompasses
a broad range of research articles, identifying key advancements and trends in combating video manipulation and focusing
on maintaining digital media trustworthiness.

Keywords Video authentication · Deep learning · CNN · LSTM · GANs

1 Introduction enforcement to news media and highbrow property control

[5].
In the realm of virtual media, the integrity and authenticity The evolution of the digital era has made video content a
of video content have become paramount, especially in a ubiquitous component of verbal exchange, fact dissemina‑
technology with easily accessible and adaptable multimedia tion, and enjoyment. However, this proliferation has also led
data. Video authentication, a crucial subset of multimedia to extended instances of video manipulation and forgery, ele‑
safety, entails verifying the originality and integrity of video vating extreme concerns over the credibility of video content
content material and ensuring that it has not been tampered circulating in the public domain. These manipulations can
with or altered from its original form [87]. This method is range from easy edits to sophisticated deepfake technology,
not the most fundamental in maintaining the trustworthiness creating convincing yet entirely fabricated video content. In
of virtual media; however, it also performs an essential func‑ any such panorama, video authentication emerges as a cru‑
tion in various packages, ranging from surveillance and law cial device for fighting misinformation, shielding copyright
and highbrow asset rights, and preserving the authenticity
of judicial proof [94].
Moreover, the importance of video authentication extends
* Ayat Abd‑Muti Alrawahneh to social and political spheres. The authenticity of video
[email protected] content can notably affect public opinion, political discourse,
* Siti Norul Huda Sheikh Abdullah or even international relations [64]. Therefore, ensuring the
[email protected] veracity of videos is not only always a technical task but also
1
a societal vitality, which makes the development of solid
Faculty of Information Science and Technology, Universiti
Kebangsaan Malaysia, 43600 Bangi, Malaysia video authentication strategies crucial.
2
CyberSecurity Malaysia, 63000 Cyberjaya, Malaysia

Vol.:(0123456789)
239 Page 2 of 30 A. A.-M. Alrawahneh et al.

Video authentication is the process of ensuring the accu‑ how complex videos are as multimedia data types and high‑
racy and reliability of video material in a changing digital light the need to ensure their integrity and originality due
environment. This process involves integrating technology, to their easy sharing nature, which opens up opportunities
ethics, and legislation to protect the integrity of videos [68, for activities such as video tampering. They stressed that
69]. This systematic literature review (SLR) investigated the minor alterations can significantly change the video agenda,
many deep learning algorithms used in video authenticity underscoring the necessity of video authentication methods
detection, focusing on their efficacy, difficulties, and future Additionally, [47] provided an overview of developed video
potential in preserving the integrity of digital video material. authentication techniques. They shed light on the increasing
This SLR explores the effectiveness, challenges, and number and quality of videos resulting from applying digi‑
future prospects of deep learning techniques in video authen‑ tal processing techniques. This phenomenon highlights the
tication detection, thereby emphasizing their importance in growing field of video forensics, which faces challenges and
maintaining the fidelity of digital video content in today’s opportunities when addressing video tampering.
digital world. This study spans various domains, including To address these challenges, researchers have focused on
video authentication, integrity, content analysis, forensics, developing techniques to verify the authenticity of video
tampering, forgery detection, and deepfake detection. content and detect tampering. They have used methods such
This SLR covers primary research studies from 2019 to as identifying changes in frame manipulation and deepfake
January 31, 2024, and it highlights various research archi‑ using deep learning algorithms. The primary contribution of
tecture methodologies, such as neural networks, convolu‑ this study is to investigate and synthesize current research
tional neural networks (CNNs), recurrent neural networks findings and methodologies related to video authentication
(RNNs), generative adversarial networks (GANs), and long using deep learning techniques. This study also explores
short-term memory (LSTM) networks, for video authentica‑ the effectiveness of various deep learning models, such as
tion detection. CNNs, RNNs, DNNs, and GANs, in detecting manipulated
Manuscript Outline: This paper is organized as follows: or falsified video content, thereby ensuring the integrity and
Section 1 introduces the topic and the importance of video trustworthiness of digital media. This study aims to provide
authentication. Section 2 presents a comprehensive literature a comprehensive overview of state-of-the-art approaches,
review of video authentication techniques. Section 3 details identify challenges and gaps in existing research, and rec‑
the research methodology used in this SLR, including the ommend directions for future investigations to enhance the
research questions (RQs), search methods, and inclusion reliability of video content authenticity in the digital age.
and exclusion criteria. Section 4 discusses the research out‑ Given the significant role that video authentication plays
comes, addresses each RQ, and concludes the review, sum‑ in preserving the integrity of digital content, it is essential to
marizing the key findings and highlighting potential future explore advanced technological approaches that can enhance
research directions. Section 5 explores the application of these detection capabilities. One such promising area is deep
deep learning techniques in video analyses, thereby thor‑ learning, a subset of artificial intelligence that has demon‑
oughly evaluating the existing research using SLR. Finally, strated remarkable success in various digital forensics appli‑
the reference section lists all the cited sources in the paper. cations. The following section provides an overview of how
Finally, the reference section lists all the cited sources in deep learning techniques are applied in digital forensics,
the paper. thereby providing a deeper understanding of their role in
video authentication.
1.1 Importance of video authentication detection
1.2 Deep learning in digital forensics: an overview
The importance of video authentication in science cannot
be overstated. It helps maintain the integrity and reliability Deep learning techniques have gained importance in video
of video evidence used in legal and security contexts. With authentication because they can handle complex data and
the rise of video editing software and related technologies, identify meaningful patterns that human analysts may need
it has become easier to manipulate video content, which has to identify more readily noticeable [6, 85]. The availability
increased the difficulty of detecting tampered videos during of media editing tools has made it easier for people to create
investigations. Video tampering, which involves alterations multimedia content, posing serious challenges in politics,
like editing or deleting content, can seriously undermine the journalism, and the judiciary. Deep learning provides oppor‑
credibility of video evidence. This alteration poses the risk tunities to generate and detect manipulated videos, thereby
of miscarriage of justice and security breaches. revolutionizing the field of forensics.
Numerous studies have emphasized the significance of According to [105] deep learning has affected multimedia
video authentication, and the difficulties associated with forensics because it exhibits better performance and robust‑
detecting tampering in videos. For example, [8] discuss ness than traditional methods when provided with a sufficiently
Video authentication detection using deep learning: a systematic literature review Page 3 of 30 239

large training set. This progress is particularly crucial in real- knowledge. This SLR aims to address existing research
world scenarios in which manipulated images and videos are gaps and explore the potential of deep learning to counter
shared on networks and often undergo forms of postprocessing sophisticated video manipulations such as deepfakes and
that traditional techniques struggle to handle [7, 101]. frame manipulations. We aim to create a resource that sum‑
Furthermore, deep learning does not aid detection but marizes current advancements and provides guidance for
plays a role in creating deepfakes, highlighting this technol‑ future research efforts to improve the effectiveness of video
ogy. As deep learning models become more adept at pro‑ authentication methods in the digital age.
ducing forged faces in videos, there is an increasing need
for approaches capable of distinguishing these forgeries
1.4 Contribution
[59]. [60] researchers demonstrated that combining tem‑
poral information with deep learning techniques can effec‑
This paper is a resource for researchers because it provides
tively detect manipulated facial videos. They used networks
insights into the advancement of learning techniques in
(CNNs) along with innovative fusion strategies to improve
video authentication. It explains the developments, high‑
forensic analysis. This fusion highlights the potential of
lights the challenges, and recommends areas for further
using learning in conjunction with CNNs to identify video
research. Moreover, our work functions as a comprehensive
forgeries and assess their authenticity frame-by-frame. On
resource for academia and industry, providing a comprehen‑
the other hand, Another research paper by [52] discussed
sive compilation of various studies and establishing itself as
the application of learning to detect video tampering. They
a fundamental reference for future research and advancing
used CNNs to classify videos as original or manipulated,
deep learning and video authentication.
demonstrating the ability to learn models to analyze video
To the best of our knowledge, this SLR combines vari‑
content and provide insights for forensic investigations. This
ous deep learning-based video authentication methods in
approach demonstrates how advanced AI technologies like
a single comprehensive document. Our findings indicate
learning can enhance the reliability and efficiency of video
that between 2019 and 2024, video authentication studies
authentication.
improved due to advanced network architectures like CNN,
AI and machine learning techniques have been widely
RNN, GAN, and LSTM.
used in recent years to automate the personalization pro‑
The remainder of this paper is organized as follows: Sec‑
cess for each topic [9]. These continuous advancements
tion 2 systematically reviews and analyzes the relevant litera‑
in the realm of learning offer robust tools that can greatly
ture to shed light on advancements and challenges in video
improve the trustworthiness and effectiveness of forensic
authentication. Section 3 describes the research methodology
video authentication. By leveraging the capabilities of learn‑
used in this study to establish a foundation for our review. Sec‑
ing models, forensic analysts can better tackle the chang‑
tion 4 presents our investigation results and addresses the RQs
ing landscape of video manipulation, ensuring that critical
mentioned in Section 3. These findings are based on evidence
contexts involving video evidence maintain their integrity.
extracted from literature included in our SLR. Section 4 also
After establishing the importance and context of video
discusses our analysis results. Provides an examination of their
authentication, it is imperative to delve deeper into the exist‑
implications and significance and summarizes the findings of
ing literature to gain a comprehensive understanding of the
our study. Finally, Section 5 recommends research directions
current state of video authentication. By examining the vari‑
in video authentication using deep learning methods.
ous studies and advancements in video authentication, we
can identify the strengths and limitations of existing tech‑
niques, emerging trends, and gaps that require attention. This
thorough exploration of the literature will provide a solid 2 Literature survey
foundation for our study and highlight the path forward for
future research endeavors. Although extensive research has been conducted in the realm
of video authentication, a systematic analysis that focuses
1.3 Motivation explicitly on the application of deep learning techniques
remains scarce (Table 1). Within the context of this SLR, it
The motivation behind the proposed SLR for learning in is highlighted that most current research focuses narrowly
video authentication is the rapid advancements in video on specific video authentication elements and often misses
manipulation technologies, which present significant chal‑ the broader capabilities of deep learning. This trend neglects
lenges to maintaining video integrity. Although there has the complete potential of deep learning to transform video
been research into deep learning for video processing, authentication, including diverse techniques, different types
there is a lack of a focused understanding of its applica‑ of manipulation, feature selection, performance metrics, and
tion in video authentication, which results in fragmented various datasets and detection types.
Table 1  A summary of related surveys is included in this study
239

Year Reported Paper Focus Of Study/Highlight Limitations Contributions Of This SLR

2016 [21] -Video forgery detection and authentication mechanisms - It does not cover deep learning techniques, which -A comprehensive review of deep learning tech‑
have been discussed have gained prominence recently niques like CNNs, RNNs, DNNs, and GANs for
-Video forgery can be classified as spatial, temporal, or - It does not cover video manipulation datasets video authentication
Page 4 of 30

spatiotemporal tampering -Examination of video manipulation datasets and

-Authentication techniques include digital signature and their applications in deep learning
watermarking-based methods
2016 [106] Real-time authentication and tampering detection in - The paper might not extensively explore advanced -It explores advanced tampering methods, including
digital video. It surveys different real-time authentication tampering methods, such as deepfake videos or deepfake videos and compression detection using
algorithms and techniques and discusses watermarking compression detection, which pose challenges for deep learning techniques and comprehensively
and tampering detection being standard techniques detection algorithms analyses their effectiveness and challenges
2018 [96] Video content authentication and detection of forgery/tam‑ -It covers advanced tampering techniques, includ‑
-The paper may not explicitly address certain aspects,
pering. It provides a repository of information on tamper such as advanced tampering techniques (e.g., deep‑ ing deepfakes and face and frame manipulation
attacks and passive-blind techniques for detection and fakes and face and frame manipulation detection) detection
explores video and counters anti-forensics -It uses deep learning methods to enhance detection
accuracy and robustness
2021 [88] Video authentication techniques using watermarking -The paper does not address the potential challenges -It addresses the potential challenges and limita‑
methods. It discusses the watermarking algorithm, which or limitations of implementing video authentication tions of implementing deep learning-based video
consists of embedding and extraction processes, and techniques authentication techniques
emphasizes that the watermark should be difficult to -It does not discuss the scalability of these techniques -It discusses the scalability of these techniques for
remove and not noticeable or detectable for large-scale video datasets large-scale video datasets
2021 [95] Comprehensive survey on passive techniques for digital -The paper does not discuss the effectiveness of these -It analyses the effectiveness of deep learning
video forgery detection. It provides a systematic and techniques in real-world scenarios or their perfor‑ techniques in real-world scenarios and their
detailed survey of passive video forgery detection tech‑ mance in different types of video authentication performance across different types of video
niques, categorizing them based on the features/artifacts -The study does not address the potential limitations authentication
used related to the accuracy and reliability of video -It addresses potential limitations related to the
authentication detection methods accuracy and reliability of detection methods
2022 [12] Detection of tampered videos, localization of altered - The paper does not address the challenges associated -It addresses challenges associated with detecting
frames, and reconstruction of tampered videos. It with detecting tampered videos and the localization tampered videos and localizing tampered frames
provides a detailed study of existing methodologies for of tampered frames using deep learning techniques
detecting tampered video localization of altered frames - The study does not explore the potential limitations -It explores potential limitations and proposes solu‑
and reconstruction of tampered video and improves the in accurately localizing tampered frames within tions for accurately localizing tampered frames
reliability factor for digital video in computerized video videos
crime scene investigation
2023 [44] Passive methods for detecting video forgeries. It provides -The paper does not delve deeply into applying deep -It covers the application of deep learning tech‑
an overview of video forgery detection and its types, learning techniques (e.g., convolutional or recurrent niques such as CNNs and RNNs for video authen‑
highlighting that passive video forgery detection tech‑ neural networks) for video forgery detection tication detection
niques can improve accuracy and effectiveness, and that -The paper does not explore the challenges of main‑ -It addresses the challenges of maintaining temporal
frame interpolation and facial features can detect video taining temporal consistency in video sequences. consistency in video sequences using advanced
forgeries Detecting subtle frame-level manipulations across deep learning methods
time remains an active research area -It explores the detection of emerging forgeries like
-The paper does not address emerging types of forgeries deepfakes and compression detections
(e.g., deepfake videos and compression detection).
Keeping up with these developments is essential
A. A.-M. Alrawahneh et al.
 Video authentication detection using deep learning: a systematic literature review Page 5 of 30 239

For example, [102] described the advanced application of

-It explores the detection of emerging forgeries like

tectures like CNNs, RNNs, and hybrid models in
-It covers the effectiveness of deep learning archi‑
deep learning techniques in complex detection tasks outside

-It addresses challenges associated with detecting

frame and face manipulation and compression

subtle frame-level manipulations across time
video authentication. This study uses a multi-input CNN
to detect chatter during the turning process in machines,
combining image and sound signals to enhance detection
accuracy. The multi-input model approach, which integrates
detecting video manipulations

multiple data types, significantly improves performance

compared with traditional single-input models, achieving
Contributions Of This SLR

an impressive accuracy of 98.8%​​. This methodology is par‑

ticularly relevant to video authentication because integrating
various data sources (such as visual frames and audio tracks)
can significantly enhance the robustness and reliability of
detection)

detection algorithms. By leveraging the principles demon‑

strated in the study of [102] video authentication systems
can be designed to use multi-input CNNs, thereby enabling
more comprehensive and reliable detection of sophisticated
-The paper does not address emerging types of forger‑

forgeries like deepfakes. This underscores the potential for

-Broad overview; specific methodologies for some

ies (e.g., frame manipulation, face manipulation,

compression detection). Keeping up with these

cross-domain applications of advanced deep learning tech‑

niques and highlights the importance of integrating diverse
detection techniques may be less detailed

data sources to realize robust detection mechanisms. This

SLR aims to bridge the existing gap by advocating for a
more integrated approach to leverage deep learning across all
facets of video authentication to tackle complex challenges.
developments is essential

The proposed method provides a broader perspective on the

application of deep learning techniques, thereby emphasiz‑
ing their potential in transforming video authentication.
Following this extensive literature review, Section 3 out‑
Limitations

lines the research methodology used for this SLR and pro‑
vides a clear understanding of the motivation behind this
study and the contributions it aims to make. It is, there‑
fore, essential to outline the rigorous methodology used to
approaches. Explores different categories of fake content
detection, discussing both the advantages and drawbacks
A comprehensive survey on state-of-the-art video forgery

conduct this SLR. A well-defined methodology ensures the

detection techniques reviews the current state of video

deepfake content detection using deep learning-based

reliability and validity of the review, guiding the process of

forgery detection using deep learning methods and

A comprehensive review of recent studies related to

selecting and analyzing relevant studies. This SLR provides

a transparent framework that underpins our investigation by
explores challenges and future directions

detailing the RQs, search strategies, inclusion and exclusion

criteria, and quality assessment procedures. In Section 3,
we describe the methodology used to gather and synthesize
existing research on video authentication, ensuring a com‑
prehensive and unbiased review.
Year Reported Paper Focus Of Study/Highlight

of the examined works

3 Research methodology

We present an SLR to evaluate the effectiveness of deep

learning approaches in video authentication. This study
adheres to the established SLR protocols, as [75] outlined.
Table 1  (continued)

This methodology discovers, defines, and analyzes all papers

2023 [17, 70]

related to deep learning for video authentication to provide

2024 [77]

answers to each RQ and identify any existing gaps. This

literature review technique demonstrates the perspectives
gained from deep learning research on video data.
239 Page 6 of 30 A. A.-M. Alrawahneh et al.

3.1 Research questions

To identify the resources used for training and testing models, which is key to understanding
How effective are these deep learning methods according to evaluating metrics in detecting To evaluate the performance of these methods in terms of accuracy, precision, F1 score, and

To guide the academic and industry communities toward developing more robust, efficient,
To highlight existing gaps and difficulties, guiding future research toward addressing these

and scalable deep learning solutions for video authentication, ensuring the integrity and
To categorize and understand the popular deep learning techniques applied in this field,
The vital part of a systematic review is determining the
RQs. We prepare RQs to follow the review process and
stay focused at the beginning of the study. This is a novel
approach to investigate the answers to the listed RQs.
Table 2 lists the RQs arising in this paper.
During the formation of RQs, the following points are
considered:

• The search stage must identify significant studies that

trustworthiness of video content in various applications

address the identified RQs.
• The data extraction stage must extract the data items

recall, crucial for assessing their practical utility

required to answer the RQs.
• The data analysis stage must synthesize the data; thus,
the RQs can be answered.

providing a foundational overview

their applicability and limitations

3.2 Search method

The search methodology adopted in this SLR is executed

in a meticulous and structured manner to ensure the com‑
prehensiveness and relevance of the surveyed literature. To
delineate the scope of our investigation and the parameters

Motivation
guiding our search, we describe our method as follows.

issues
3.2.1 Choose database

What benefits, challenges, and limitations are currently faced in applying deep learning to
We identified and selected multiple academic databases and

What are the common datasets used in deep learning for video authentication detection?

What future directions and innovations are recommended in the literature for enhancing
digital libraries pertinent to our research domain. These
include but are not limited to IEEE, SPRINGER, MDPI,
ASSOC COMPUTING MACHINERY, and Wiley, as shown
RQ1 What are the popular deep learning techniques used in video authentication?

in Table 3. The selection was based on the databases’ cover‑

age of topics related to our research area, accessibility, and
scholarly relevance.

3.2.2 Choose Keywords

Critical to our search method was the development of a com‑

prehensive set of keywords and phrases directly related to
the core themes. These terms are derived from preliminary
deep learning-based video authentication?

literature reviews, expert consultations, and relevant research

frameworks in specified fields. Boolean operators (AND,
authentic and manipulated videos?

OR, NOT) were used to effectively combine these terms,

video authentication detection?

thereby enabling a more targeted search. This SLR system‑

atically explores a broad spectrum of topics within the realm
of video forensics, including “video authentication,” “video
Table 2  Research questions

integrity verification,” “video content analysis,” “video for‑

Research question

gery detection,” “video tampering detection,” “video foren‑

sics,” “video manipulation,” “deepfake,” “video validation,”
and “frame manipulation.” Concurrently, it delves into the
utilization of advanced neural networks and machine learn‑
ing techniques, such as “deep learning,” “CNN,” “RNN,”
RQ2

RQ3

RQ4

RQ5

“GAN,” and “LSTM” to address and counteract video

ID
Video authentication detection using deep learning: a systematic literature review Page 7 of 30 239

Table 3  The databases selected for conducting this systematic literature review
Database Publisher

IEEE [37] INT INFORMATION & ENGINEERING TECHNOLOGY ASSOC [42]

SPRINGER [98] INT JOURNAL COMPUTER SCIENCE & NETWORK SECURITY-IJCSNS [92]
ELSEVIER [26] ScienceDirect [91]
ASSOC COMPUTING J J STROSSMAYER UNIV OSIJEK, FAC ELECTRICAL ENGINEERING [43]
MACHINERY (ACM)
[62]
MDPI [65] KAUNAS UNIV TECHNOLOGY [100]
Wiley [107] NOW PUBLISHERS INC [39]
PEERJ INC [40] PERGAMON-ELSEVIER SCIENCE LTD [58]
Tech Science Press [78] ASSOC ADVANCEMENT ARTIFICIAL INTELLIGENCE [41]
ACADEMIC PRESS INC EUROPEAN ASSOC SIGNAL SPEECH & IMAGE PROCESSING-EURASIP [79]
ELSEVIER SCIENCE
[2]
AIP Publishing [80]

Fig. 1  Keywords in deep learning for video authentication SLR

forgery and manipulation challenges. This dual focus aims to 3.2.3 Choose time range
provide a comprehensive overview of state-of-the-art video
verification methods and the role of cutting-edge AI tech‑ Initially, we extracted an overall total of 205 articles over
nologies in enhancing the authenticity and integrity of video the last five years. Subsequently, 93 publications were elimi‑
content. Additionally, we recommend future research at the nated because their content did not align with our research
nexus of video forensics and deep learning. A visual repre‑ objective. As part of the subsequent procedure, 13 papers
sentation of keywords in deep learning for video authentica‑ that lacked relevant material and were inaccessible were
tion using SLR is shown in Fig. 1. excluded.
 239 Page 8 of 30 A. A.-M. Alrawahneh et al.

Table 4  Inclusion and exclusion criteria Finally, we included 99 publications in the final research,
Criteria Description consisting of conference and journal papers.

Inclusion criteria - Only conference and journal Inclusion and exclusion criteria This study exclusively
publication
includes conference and journal publications published in
- Methodology of work shows
deep learning for video English. Consequently, we have excluded works that have
authentication for any func‑ not been under-reviewed, as well as review papers and
tionality book chapters. Table 4 presents the criteria used to deter‑
- Only image/video data
mine which publications were included and excluded in the
authentication is considered
in the context of deep learn‑ search.
ing research
Exclusion criteria - Articles in other than the Quality assessment To obtain reliable outcomes and conclu‑
English language sions, using SLR in conjunction with high-quality research
- Gray literature, including
is vital. This scenario requires comprehensive planning,
thesis and dissertations
- Research lacking empirical accurate keyword selection, and clearly defined criteria
evidence or without clear for inclusion and exclusion. Subsequently, snowball track‑
methodology or review ing was conducted by reviewing the reference lists of each
papers
primary study. The validity of the research may be further
- Study having content other
than image/video data like analyzed using the criteria outlined in Table 5.
audio and texts As shown in Fig. 2, most studies included in the SLR
-Related research that did not were conducted between 2019 and 2024, with the high‑
use deep learning techniques
est number of publications occurring in 2022. This surge

Table 5  Quality assessment Aspect Elucidation

analysis criteria
Internal scope The goal, assumptions, and background of the research are provided
External validity The findings are applicable in both industry and academics
Construct alignment The relationship among research questions, assessment/measure‑
ment aspects, and outcomes is well defined
Conclusive integrity The findings and analyses are derived from theoretical and empiri‑
cal research and align with the research questions

Fig. 2  Distribution of studies by publication year between 2020–2024

Video authentication detection using deep learning: a systematic literature review Page 9 of 30 239

Fig. 3  Distribution of studies by

publisher

indicates a growing interest in and possible significant devel‑ detection. This article discusses the structure of models and
opments in the field during these years. discusses significant research conducted using these models.
According to the pie chart in Fig. 3, IEEE published 39%
of the studies under review, followed by Springer (18%) 4.1 RQ (1) What are the most effective deep
and Elsevier (11%). Other notable publishers include the learning techniques currently employed
Association for Computing Machinery and the MDPI, each for video authentication, and how do they
contributing 5% of the publications. Wiley and PEERJ INC compare in terms of accuracy, robustness,
hold a 4% share of the company, and Tech Science Press is and computational efficiency?
represented. The category labeled “Other” indicates a diver‑
sity of sources, comprising 9% of all publications. This dis‑ In Section 4, we discuss RQs that survey the video
tribution reflects the dissemination of research in this field authentication field. This section begins by identifying
across various scientific journals and publishers. prevalent video authentication detection techniques in
After establishing a robust and systematic methodology, scholarly literature, illustrated in a taxonomy of the litera‑
this SLR now focuses on research outcomes. This section ture shown in Fig. 5. It acknowledges the predominance
addresses each RQ and provides detailed insights derived of image and video manipulation through deep learning
from the literature. By systematically presenting the find‑ methods and organizes various research works based on
ings, this SLR aims to shed light on the various deep learn‑ the deep learning approaches used. Detailed descriptions
ing techniques used in video authentication; their effective‑ of these methods are provided in subsequent sections of
ness, the datasets and evaluation metrics commonly used, the SLR.
and the benefits, challenges, and limitations encountered
in this field. Additionally, it explores future directions and 4.1.1 Deep neural network ( DNN)
innovations identified in the literature, providing a compre‑
hensive overview of the current state and potential advance‑ A deep neural network (DNN) is a type of artificial neural
ments in deep learning-based video authentication. network with multiple input and output layers. It is used in
various applications, including video processing and authen‑
tication. In the context of video authentication, DNN tech‑
nology can be applied to more realistic image, audio, and
4 Results and analysis video processing, allowing for tasks such as video analysis,
facial detection, recognition, and entity tracking [83].
We examined 99 peer-reviewed research publications [36] focused on using DNNs to detect fake videos, and
focused on video authentication and used deep learning they specifically focused on the ability to recognize deep‑
techniques. This section examines the features, methods, fakes. The study also explored the use of adversarial pertur‑
threats, solutions, and deep learning algorithms mentioned bations to challenge the efficacy of DNN-based deepfake
in the selected articles (Fig. 4). detectors by setting up scenarios for both white and black
Following a thorough investigation, we provide the find‑ box attacks to test the detectors’ robustness. The study
ings derived from the RQs listed in Table 2. This section pre‑ reports an average success rate (SR-C) of 98.07% for Xcep‑
sents a discussion of the models used in video authenticity tionNet and 99.83% for MesoNet in detecting deepfakes
239 Page 10 of 30 A. A.-M. Alrawahneh et al.

Fig. 4  PRISMA flow diagram for this SLR

in compressed video formats, as evaluated on the FaceFo‑ [16] also developed a DNN for classification and imple‑
rensics + + dataset and further verified on the Deepfake mented a multitask cascaded convolution neural network
Detection Challenge (DFDC) dataset. The results indicate for facial detection and a cascaded deep sparse autoencoder
that future advancements should include methods akin to (CDSAE) for feature extraction. A temporal CNN (TCNN)
adversarial training to enhance the resilience of deepfake was used to train the CDSAE. The proposed model exhib‑
detectors against emerging and more sophisticated deepfake ited high accuracy on several datasets: 98.7% on Face2Face,
generation techniques. 98.5% on FaceSwap, and 97.63% on the DFDC (Deepfake
However, [45] used a DNN to analyze biological sig‑ Detection Challenge) dataset. However, this study identi‑
nals for deepfake detection. The proposed DNN method fied a significant overfitting issue in deepfake detection
achieved a high accuracy rate of 98% on the main public approaches, and it faces specific challenges in feature selec‑
dataset, which includes UADFV, FaceForensics + + , and tion and estimation with existing techniques. This high‑
deepfake detection datasets. The study also acknowledges lights the necessity of addressing the overfitting problem to
the difficulty of detecting deepfake videos, particularly improve the reliability of deepfake detection methods.
under complex scenarios, such as erratic character move‑ Despite these advancements, challenges remain, such as
ments and rapid scene changes. To address these challenges, the difficulty of detecting deepfakes under complex scenar‑
future research directions may focus on improving signal ios and the overfitting problem in deepfake detection mod‑
processing techniques, specifically signal enhancement and els. Future research is recommended to focus on enhanc‑
denoising, to enhance the detection of deepfakes under such ing the resilience of deepfake detectors through adversarial
intricate conditions. training, improving signal processing techniques for better
Video authentication detection using deep learning: a systematic literature review Page 11 of 30 239

Fig. 5  Taxonomy of deep learning methods used to detect video authentication

detection under complex conditions, and addressing overfit‑ like ReLU, enable CNN to handle complex patterns, and the
ting to improve the reliability of detection methods. This final layer uses functions like SoftMax to classify the input
study underscores the critical need for continued research into different categories. This architecture allows CNNs to
into using DNNs for video authentication to keep pace with process and analyze visual data efficiently.
evolving video manipulation technologies. The initial idea was supported by a study [27], which
revealed that they proposed a system that uses deep learn‑
4.1.2 Convolution neural network (CNN) ing techniques to detect interframe forgeries in surveillance
videos. The proposed method focuses on detecting various
A Convolutional Neural Network (CNN) is a class of deep types of forgeries, even in postprocessing operations, such as
learning neural networks that is widely used in visual compression, noise, blurring, and brightness modifications.
imagery analysis. CNNs are particularly powerful for image The proposed system achieves superior results compared
and video recognition, image classification, medical image with previous methods based on handcrafted features and
analysis, and many other autos vision tasks. According to has potential applications in real-world surveillance systems.
its capacity to automatically acquire hierarchical features In addition to previous work, [84] emphasized the impor‑
from input data, it can improve computer vision processes. tance of recognizing fake images and videos in social net‑
A CNN is a specialized deep-learning structure that works and media because of the potential for intentional
learns to recognize and categorize image features through manipulation and disinformation. They introduced a deep
a series of layers. These include convolutional layers that learning approach to identify copy-move forgeries, a com‑
apply various filters to detect features, pooling layers that mon type of image manipulation. Two models of copy-move
simplify the feature maps while maintaining essential infor‑ forgery detection based on deep learning are proposed, and
mation, and fully connected layers that interpret these fea‑ their performance is evaluated using data from eight public
tures to make predictions. Nonlinear activation functions, datasets. The models reported 97.0% accuracy and 97.0%
239 Page 12 of 30 A. A.-M. Alrawahneh et al.

F1 score. These findings have potential applications in Authors in [38] found certain differences suggesting a
image and video forensics, offering a promising approach hybrid technique based on Inception-ResNet-V2and Xcep‑
to combating disinformation and ensuring the integrity of tion CNN architectures to detect deepfake videos. They
digital media. These methods contribute to the development evaluated the performance of the proposed technique on
of effective methods for detecting copy-move forgeries, the DFDC dataset and reported high accuracy in detect‑
addressing the need for robust recognition algorithms in the ing deepfake videos, with a maximum precision of 98.5%,
context of digital media and social platforms. recall of 96%, an F1 score of 98%, and support of 96.8%.
Further innovations by [54] used deep learning features They emphasize the importance of developing technologies
and a parallel CNN architecture to detect and locate frame to combat the spread of deepfake videos, highlight poten‑
insertion forgeries in videos. They identified inconsistencies tial ethical implications, and indicate that individuals and
between frames by analyzing the correlation coefficients of organizations should take steps to protect themselves from
deep features, and they proved to be highly accurate, with a the harmful effects of such videos.
99.96% success rate in frame-level detection. For the stand‑ The use of CNNs for video authentication presents prom‑
ard video datasets, VIFFD and SULFA achieved 86.5% and ising avenues to significantly contribute to the detection of
92% accuracy, respectively, at the video level. The proposed sophisticated video manipulations like deepfakes and vari‑
method also helps pinpoint the exact location of the forged ous forms of video forgeries with high accuracy. Innova‑
frames, thereby aiding in the restoration of the original tions in CNN architectures, such as Inception Resnet V2,
video. With the characterization as noninvasive, efficient, Xception, and parallel models, have notably enhanced the
and robust with low time complexity, the VFID-Net method effectiveness of these technologies in identifying fraudulent
is well suited for practical use and contributes significantly videos across diverse datasets.
to the realm of video forensics, offering a reliable solution Despite these advancements, the video authentication
to combat the increasing issue of video tampering. domain using CNNs faces challenges that underscore the
The findings of [90] are in agreement with the past stud‑ importance of ongoing research and development. The
ies, which the proposed model used the Xception CNN increasing complexity of video forgeries demands continu‑
architecture and demonstrated strong performance on key ous evolution of CNN models and training approaches to
datasets, achieving an AUC of 0.931, 0.633, and 0.627 for improve detection precision. More varied and extensive
FaceForensics + + , Celeb-DF, and DFDC, respectively. datasets are required to develop CNN models with robust
They indicated that focusing on individual facial features generalization capabilities for different forgery techniques
rather than an entire face yields more accurate classifica‑ and unknown scenarios. Moreover, integrating multimodal
tion results, which led to an approach that uses parts-based data, including audio and metadata, can provide a compre‑
detectors for in-depth empirical analysis across these data‑ hensive strategy for authenticating videos.
sets. Additionally, the proposed model effectively addresses Future research must address overfitting prevention and
various digital forgery techniques, such as deepfakes and enhance model generalizability to novel manipulations
FaceSwap, to enhance the capabilities of digital media through innovative regularization, data augmentation, and
verification. adversarial training techniques. Ethical considerations and
The above findings are consistent with the study of [111]. the potential for misuse of video authentication technology
They also used a CNN, achieving an F1 score of 93.78% must be considered, as well as ethical guidelines and stand‑
with a single input (In) and an increased accuracy of 2.38% ards. The future path for CNN applications in video authen‑
when the number of inputs was expanded to 13, resulting in tication includes technical advancements, dataset expansion,
an F1 score of 96.06%. The datasets used for this research improved model generalization, and ethical usage considera‑
included XIPH1, XIPH2, and MCL-V. Notably, XIPH1 was tions. Collaboration among academics, industry experts, and
an unseen dataset reserved for testing, whereas XIPH2 and regulatory bodies is essential to address these challenges
MCL-V, cropped datasets, were used during the training and fully realize the potential of CNNs in safeguarding the
phase. The study acknowledged certain limitations, such as integrity of digital media.
the reliance on a limited selection of datasets and the con‑
centration on a specific type of video manipulation known 4.1.3 Recurrent neural network (RNN)
as frame rate up-conversion (FRUC). Future research direc‑
tions will include the expansion of the dataset to encom‑ An RNN, or Recurrent Neural Network, is a type of artificial
pass additional forms of video manipulation while probing neural network in which connections between nodes form
the efficacy of the FCDNet model in practical, real-world a directed graph along a temporal sequence. This connec‑
applications. FRUC was identified as a key form of video tivity allows it to exhibit dynamic temporal behavior for a
manipulation within the scope of this study. time sequence. Unlike feedforward neural networks, RNNs
Video authentication detection using deep learning: a systematic literature review Page 13 of 30 239

use their internal state (memory) to process sequences of and classifies the extracted features. Tested on the TRECVid
inputs. This makes them applicable to unsegmented, con‑ 2001, 2007, and RAI databases, the model achieved an aver‑
nected handwriting and speech recognition. age F1 score of 94.4%, precision of 94.1%, and recall of
The key feature of RNNs is their ability to obtain a deeper 95%. Despite these high-performance metrics, this study
understanding of a sequence and its context compared with identified a need for better generalization across different
other algorithms. This sequence pattern is achieved through video types and conditions, as well as a need to handle
loops, which allow information to be carried across neurons diverse and complex video content more effectively. It also
while processing a sequence of inputs. Thus, they can con‑ highlighted the necessity for more efficient computational
sider the entire history of inputs when making predictions. strategies, particularly in the context of video shot bound‑
However, it can be quite difficult to train basic RNNs ary detection.
effectively due to vanishing or exploding gradients. To Similarly [110], proposed affective RNN incorporates
address this issue, more sophisticated versions of RNNs, LSTM as its recurrent layer used in the model, which has
such as Long Short-Term Memory (LSTM) networks and improved arousal and valence detection accuracy by 0.251%
Gated Recurrent Units (GRUs), have been developed. These and 0.862%, respectively. Furthermore, the Pearson correla‑
architectures are designed to remember long-term informa‑ tion coefficient improved by 2.7% and 3.1% in the arousal
tion and are the gold standards for sequential data, such as and valence domains of the EIMT16 benchmark. The mod‑
time series data, speech, text, financial data, audio, video, els were evaluated on LIRIS-ACCEDE and video emotion
and weather. datasets. It is suggested that future research should focus on
A similar methodology is used by [50], which proposed classifying similar emotions more accurately to enhance the
a detection framework that incorporates a two-step process analysis of affective video content. The proposed approach
to identify synthetic face images. This process employs underscores the importance of the adaptive fusion of modali‑
Pixel RNNs, and a straightforward classifier derived from ties and temporal inputs in affective video content analysis.
the LeNet-5 architecture. The proposed system has dem‑ Research into RNNs for video authentication has dem‑
onstrated a commendable average detection accuracy of onstrated their effectiveness in detecting video forgeries,
95.73%, as evidenced by its performance on the FaceForen‑ including deepfakes and other sophisticated manipulations,
sics dataset. Challenges remain, including the need for bet‑ with high accuracy. Studies have highlighted the potential of
ter generalization across different fake image types and the RNNs, particularly LSTM networks and GRUs, to analyze
evolving sophistication of deepfake techniques. This rapid video temporal sequences, which is crucial for identifying
evolution underscores the ongoing need for advanced detec‑ manipulated content. However, challenges, such as better
tion methods capable of identifying synthetic content with generalization across various video types, handling complex
deepfakes as the primary concern. content, and improving computational efficiency, remain.
In continuation, [72] employed an inflated 3D (I3D) Future work should focus on developing more advanced
and Siamese-based RNN to detect video forgeries. This RNN architectures, enhancing the generalizability of mod‑
method achieved impressive accuracy, precision, recall, F1 els, expanding training datasets to include a broader range
score, and Matthew’s correlation coefficient (MCC) on two of forgeries, and improving computational strategies for
datasets: Image Retrieval and Analysis Tool (VIRAT) and efficient processing. Additionally, incorporating multimodal
Media Forensics Challenge (MFC), with scores of 86.6% data analysis and addressing ethical considerations regard‑
and 93.3% in accuracy, 87.5% and 93.3% in precision, 86.6% ing video authentication technology is essential to ensur‑
and 93.3% in recall, 86.5% and 93.3% in F1 score, and 71.7% ing its appropriate application. Addressing these challenges
and 86.65% in MCC, respectively. The VIRAT action rec‑ requires collaborative research efforts to enhance the reli‑
ognition and MFC datasets were used. A notable limitation ability and effectiveness of RNN-based video authentication
of this study is its sole focus on detecting duplicate frame methods.
video forgeries without addressing other forgery types like
deepfake videos. The results indicate that future research 4.1.4 Generative adversarial networks (GANs)
should extend to other types of forgeries and that improve‑
ments could be made by incorporating diverse deep learn‑ Generative Adversarial Networks (GANs) are deep genera‑
ing techniques and applying the model to larger datasets. tive models that have gained significant attention recently.
The primary objective was to detect duplicate frame video GANs are capable of implicitly learning complex and
forgeries. high-dimensional distributions over images, audio, and
Finding by [19] also points towards the results work in data. GANs comprise two adversarial models: generative
[72]. They employed a Siamese recurrent architecture with and discriminative. The generative model learns to map a
a pre-trained InceptionV3 model and an LSTM network for simple latent distribution to a more complex data distribu‑
feature extraction. The final layer of the model then fuses tion, whereas the discriminative model learns to distinguish
239 Page 14 of 30 A. A.-M. Alrawahneh et al.

between real and generated data [76]. GANs have been suc‑ performance, particularly by focusing on interframe video
cessfully applied in various fields, including image and video forgery detection using deep learning methodologies.
processing, medical image analysis, and unsupervised rep‑ In addition to the work of [48] and [15], Authors [74] con‑
resentation learning. However, major challenges exist when clusively agree GAN is a promising method to examine vari‑
training GANs, such as mode collapse, nonconvergence, ous face swapping methods, including 3D-based, GAN-based,
and instability, due to an inappropriate design of network Face2Face, and Neural Textures, and assessed their detec‑
architecture, the use of objective functions, and the selection tion efficacy. The detection results were highly successful at
of an optimization algorithm. To address these challenges, 99.7%, 89.36%, and 75.19% for general face swaps, 3D-based
several solutions for the better design and optimization of swaps, and FSGAN, respectively. The methods were evaluated
GANs have been investigated on the basis of techniques for on FaceForensics + + , Celeb-DF-v2, and DFDC datasets.
reengineered network architectures, new objective functions, This paper proposes future research directions, recommending
and alternative optimization algorithms [89]. the exploration of additional detection signals, enhancing the
In contrast to previous different deep learning studies, adaptability of current methods to newer forms of manipula‑
[48] implemented a GAN model using an algorithm called tion, and developing techniques to address more sophisticated
Deep Vision, achieving a notable accuracy rate of 87.5% manipulation scenarios. The primary focus of this study is the
in deepfake image and video detection, as tested on Kag‑ detection of deepfake images and videos.
gle datasets. The study proposes an innovative approach to GANs have demonstrated significant promise in video
address the current gaps in pixel-based integrity verification authentication, particularly in detecting sophisticated
algorithms. The proposed method also opens future research video manipulations like deepfakes. Despite their poten‑
avenues to investigate the creation of deepfakes sufficiently tial, GANs face training instability, mode collapse, and the
sophisticated to replicate natural blinking patterns, which need for continuous adaptation to new manipulation tech‑
could circumvent the Deep Vision algorithm’s detection niques. Future work should improve GAN architectures to
capabilities. realize excellent stability and efficiency, develop advanced
Interestingly, this is in contrary to study conducted by detection methods that incorporate additional data modali‑
[15], which explored the capabilities of StyleGAN in the ties, and enhance models’ ability to generalize across vari‑
realm of image anti-forensics, focusing on face anti-foren‑ ous manipulation types. Additionally, ensuring the ethical
sics in images. The proposed model presents notable quan‑ use of GANs and enhancing their robustness against anti
titative results: an average attack success rate of 0.8882, a forensic methods are crucial areas for further development.
Structural Similarity Index of 0.9849, a peak signal-to-noise Addressing these challenges will help leverage GANs more
ratio of 44.5047, and a mean learned perceptual image patch effectively in video authentication and maintain the integ‑
similarity score of 0.0045. The evaluation was performed rity of digital media.
using the CelebA dataset and a GAN-generated facial data‑ On the basis of the included research papers in this SLR,
set. This study underscores the current research emphasis the findings shown in the pie chart “Deep learning methods”
on enhancing the transferability and visual quality of anti- illustrate the proportion of different deep learning techniques
forensic methods, especially their application to black box applied in research. CNNs dominate the field, comprising
detectors, which is an area that requires further improve‑ 61% of the methods used. LSTM networks are the next most
ment. The study also notes that defense mechanisms could used at 15%, followed by RNNs and DNNs, each at 7%.
compromise the adversarial capabilities of such anti-forensic GANs and autoencoders are less commonly used (5% and
methods. Looking ahead, the research recommends focusing 1%, respectively). The category “Other” accounts for 4%
on improving the transferability and increasing the robust‑ of the methods, indicating that additional and less common
ness of anti-forensic methods against such defense solutions. techniques are also in use. Considering all these methods,
The YOLO (V2) neural network combined with CNN Fig. 6 shows the deep learning methods used to detect video
and GAN techniques was used to achieve an average object authentication in this SLR.
detection accuracy of 99% with a confidence score ranging With a thorough understanding of the various deep learning
from 95 to 99% across various videos [83]. Tested on the techniques used in video authentication, this SLR examines
SULFA and SYSU-OBJFORG datasets, the system exhibits the effectiveness of these methods in detecting authentic and
promising results in object-based video forgery detection, manipulated videos. By evaluating different techniques such
specifically targeting copy–move attacks. However, the as frame and face manipulation, compression, and deepfake
proposed system faces limitations in terms of localization detection, this section comprehensively analyzes how well
accuracy, ability to generalize across different datasets, and these deep learning approaches perform in practical applica‑
effectiveness in detecting other forms of video tampering. tions. This examination not only highlights the strengths and
The findings indicate the need for further development to weaknesses of each method but also offers valuable insights
overcome these challenges. Future work should enhance into their real-world applicability and reliability.
Video authentication detection using deep learning: a systematic literature review Page 15 of 30 239

Fig. 6  Deep learning methods

are used to detect video authen‑
tication in this SLR

4.2 RQ (2) How do different deep learning‑based deep learning in identifying and authenticating original video
video authentication techniques perform sequences while pinpointing alterations. The proposed method
in detecting various types of video can be further improved by incorporating other deep learning
manipulations, including deepfakes, frame techniques and exploring the use of larger datasets.
manipulation, and compression detection? Author in [32] also detected frame rate manipulation in
videos, particularly through the use of deep FRUC tech‑
On the basis of the findings from the SLR, as represented niques based on multiscale dual-flow CNN, and achieved a
in the summary of selected papers, here is a discussion precision of 97.53%, recall of 96.34%, and an F1 score of
about the effectiveness of deep learning methods in terms 96.93% using the UCF101 dataset. Additionally, the pro‑
of detecting authenticity and manipulations related to frame posed method was evaluated on a limited dataset, and its
and face manipulation, deepfake, and compression detec‑ generalizability to other datasets and deep FRUC methods
tion. The manipulation detection category taxonomy used requires further investigation. Future research can explore
to detect video authentication in this SLR is illustrated in the combination of different video manipulation detection
Fig. 7. Figure 8 shows the distribution of manipulation methods to improve the overall detection performance.
detection categories for video authentication detection in Each of these studies advances our understanding and
this SLR. capabilities in detecting frame manipulations and outlines
the path forward. Collectively, they emphasize the critical
4.2.1 Frame manipulation detection need for ongoing development in deep learning methodolo‑
gies to adapt to and counteract the evolving landscape of
Frame manipulation detection in digital content is a critical digital video manipulation, highlighting the dynamic inter‑
research area, particularly with the advent of sophisticated tech‑ play between technological advancement and the imperative
niques such as deepfakes. Deep learning methods have emerged for robust detection mechanisms. They are also crucial for
as powerful tools for identifying and mitigating these manipula‑ forensic analysis, ensuring the integrity of video evidence
tions with the aim of protecting the integrity of digital media. by detecting duplicate frame video forgeries that could mis‑
Deep learning methods have shown promising results in represent facts or alter narratives.
detecting frame manipulation, which includes frame cutting,
adding, and altering in videos. Techniques such as I3D and
Siamese-based RNN for detecting duplicate frame video for‑ 4.2.2 Face manipulation detection
geries have demonstrated high effectiveness. The novel method
mentioned by [72] achieved accuracy of 86.6% and 93.3%, Deep learning algorithms have been pivotal in advanc‑
precision of 87.5% and 93.3%, recall of 86.6% and 93.3%, F1 ing face manipulation detection, particularly CNN and
score of 86.5% and 93.3%, and MCC of 71.7% and 86.65% for LSTM networks. These methods leverage facial recogni‑
the VIRAT and MFC datasets, highlighting the capabilities of tion and behavioral patterns to differentiate authentic and
239 Page 16 of 30 A. A.-M. Alrawahneh et al.

Fig. 7  Taxonomy of manipula‑

tion detection categories to
detect video authentication in
this SLR

Fig. 8  Distribution of manipu‑

lation detection categories for
detecting video authentication
in this SLR

manipulated content. For example, the LSTM-based affec‑ become more advanced, detection methods have also evolved,
tive video content analysis highlights the nuanced capabili‑ using sophisticated technologies such as DNNs [71], facial
ties of deep learning in interpreting emotions and detect‑ recognition [3], and the analysis of physiological signals like
ing inconsistencies that may lead to signal manipulation. heart rate, through biological signals in videos [22]. Despite
Although the review indicates substantial success rates, it progress in detection technologies, the challenge of keeping up
also indicates the need for continuous improvement in algo‑ with increasingly sophisticated manipulations persists, necessi‑
rithmic sensitivity to subtle facial manipulations, which is a tating continuous research and development. Researchers also
common challenge in deepfake videos. In a study by [18] the emphasized the importance of ethical use and the implemen‑
proposed method used a metric learning-based framework tation of regulations to prevent misuse of these technologies,
CNN with steganalysis-rich models for face manipulation highlighting the need for responsible handling and protective
detection and achieved accuracy gains of 5%–15% under measures against harmful applications.
unseen manipulations using the FaceForensics + + , Celeb,
DF, and Deepfake TIMIT datasets. Future research should 4.2.3 Compression detection
expand on datasets with more challenging variants and addi‑
tional layers in CNNs to highlight manipulation-invariant Compression detection is an essential area of research that
features. aims to identify manipulations related to video compression.
The effectiveness of face manipulation detection varies with This manipulation can include detecting double compres‑
the manipulation’s complexity and the quality of the altered sion, which is a common trait in tampered videos in which
images or videos. However, as fake creation techniques have a video has been compressed, edited, and then compressed
Video authentication detection using deep learning: a systematic literature review Page 17 of 30 239

again, leaving distinct artifacts that can be analyzed for learning in combating deepfake. However, these methods
authentication purposes. do not work on all datasets. For example, in [38] the results
From the overview, it can be deduced that there is little demonstrate that the DFDC dataset has a precision of 98.5%,
published data on compression detection except for [34] who F1 score of 98%, support of 96.8%, and recall of 96%, which
developed a hybrid neural network that integrates attention- demonstrates that the methods work well and how important
based mechanisms to detect double HEVC compression in it is to have various datasets and ensure that the algorithms
videos. The model achieved an accuracy of 94.07%, a true work on all of them.
negative rate of 95.18%, and a true positive rate of 92.96% Recent studies have demonstrated significant advance‑
on publicly accessible raw videos (YUV sequences). The ments in deepfake detection. For example, [23] used vision
results highlight challenges in identifying relocated I-frames transformers combined with EfficientNetV2 to achieve high
in HEVC compressed videos, and they indicate future direc‑ classification accuracy on datasets such as ForgeryNet, dem‑
tions for improving detection capabilities in such scenarios. onstrating the potential of transformer models in identify‑
A similar methodology is used by [11] but focuses on ing manipulated media. Similarly, approaches that integrate
detecting double compression in videos using a CNN sparse autoencoders with graph-based clustering have dem‑
architecture called MultiFrame-Net, and they achieved an onstrated promise in detecting deepfakes across diverse data‑
accuracy of 88.42% for I-Frame input on the VISION data‑ sets like FFHQ, 100K-Faces, Celeb-DF, and Wild Deep‑
set. The findings underscore the need for more extensive fake. Moreover, the application of cascaded deep sparse
research on social media provenance, especially for videos autoencoders and CNNs enhances the accuracy of deepfake
subjected to double compression, to effectively trace their detection. These models have achieved notable success rates
origin and authenticity. on datasets such as Face2Face, FaceSwap, and the DFDC,
Furthermore, [46] used CNNs to localize tampered areas highlighting the efficacy of deep learning in distinguishing
in videos and achieved accuracy values ranging from 66.53% between authentic and manipulated content [49].
to − 71.18% on the Xiph.org and Vision datasets. This paper Despite these advancements, deepfake detection faces
addresses limitations related to the use of only two datasets several challenges, highlighting the complexity associated
and the challenge of detecting sophisticated tampering meth‑ with combating digital media manipulation. One of the pri‑
ods, indicating further exploration of deep learning tech‑ mary issues is the continuous evolution of deepfake genera‑
niques to realize more accurate localization of tampering in tion techniques, which are becoming increasingly sophisti‑
compressed videos. cated, thereby making detection more difficult [61]. As these
The above studies demonstrate the potential of deep technologies evolve, they generate realistic and harder-to-
learning to detect compression-related video manipula‑ detect manipulations, which require constant updates and
tions, which is an essential aspect of video authentication. enhancements to detection algorithms.
Researchers can discern the authenticity of video content Another significant challenge is the generalization across
by identifying specific artifacts and patterns associated different deepfakes and datasets. Many deep learning mod‑
with compression. An ongoing challenge lies in enhanc‑ els are trained on specific datasets and may underperform
ing the accuracy and reliability of these detection methods, when exposed to deepfakes generated by unknown methods
especially as manipulation techniques become increasingly or data sources [50]. This limitation underscores the need
sophisticated. Future work will likely focus on improving for models that can generalize well across diverse scenarios
deep learning models to address these challenges and ensure and manipulation techniques.
robust video authentication mechanisms despite evolving The adversarial attack problem poses a further challenge,
digital forensics scenarios. where attackers intentionally modify deepfakes to evade
detection. As deepfake detectors become more advanced,
4.2.4 Deepfake detection so do the strategies used by adversaries to circumvent these
systems. Adversarial Deepfakes: Evaluating Vulnerability
Detecting deepfakes is a significant challenge due to the [36]. This arms race necessitates the development of robust
sophisticated AI techniques used to create highly realistic detection mechanisms that can resist such adversarial tactics.
fake videos. This SLR presents several deep learning meth‑ Moreover, the computational cost of training and deploying
ods that are well suited to this task. These include two-step deep learning models for deepfake detection can be prohibi‑
synthetic face image detection methods and adversarial tive, especially in real-time applications [23]. Efficient mod‑
deepfake generation models to evaluate the effectiveness of els that balance accuracy and computational efficiency are
detection. Notably, [50] achieved an average detection accu‑ essential for practical deployment across various platforms
racy of 95.73%, which demonstrates the potential of deep and devices.
239 Page 18 of 30 A. A.-M. Alrawahneh et al.

The overfitting problem arises when a model learns to frame tampering) and desired application scenario (e.g.,
perform exceptionally well on training data but fails to social media content and surveillance footage). The con‑
generalize to unseen data. This challenge is particularly tinuous evolution of video manipulation techniques requires
pertinent in the context of deep learning-based deepfake regular updates to these datasets and the development of new
detection, where the diversity and complexity of manipula‑ benchmarks to keep pace with the advancing technology.
tion techniques continue to grow [16]. Finally, the lack of A comprehensive summary of the various deep learning
large, diverse, publicly available datasets for training and methods used for video authentication detection, highlight‑
testing deepfake detection models has hindered progress ing their evaluation measures, datasets, and manipulation
[81]. Developing comprehensive datasets that cover various types, is provided in Table 6. Additionally, it underscores
manipulation techniques and scenarios is critical for advanc‑ the limitations of the reviewed studies to provide a clearer
ing state-of-the art deepfake detection. understanding of the current state of research in this domain.
In conclusion, although deep learning has shown great Table 6 serves as a valuable resource for identifying the
promise in detecting deepfakes, addressing these challenges strengths and weaknesses of different approaches, thereby
through innovative research, interdisciplinary collaboration, facilitating the development of more robust and effective
and ethical guidelines will be vital to advancing the field video authentication detection techniques.
and ensuring the integrity of digital media. Table 6 shows Building on exploring datasets and evaluation metrics,
the effectiveness of the deep learning methods in terms of this SLR discusses the benefits, challenges, and limitations
detecting authenticity and manipulations related to frame of applying deep learning to video authentication. By delv‑
and face manipulation, compression detection, and deep‑ ing into the practical advantages of deep learning methods
fake detection. Following an assessment of the effectiveness and the obstacles and constraints encountered in real-world
of various deep learning techniques in detecting authentic applications, this section provides a balanced view of the
and manipulated videos, this SLR explores datasets and field’s current state. This discussion illuminates the practi‑
evaluation metrics commonly used in the field. By examin‑ cal implications of deploying deep learning techniques and
ing the diverse datasets that serve as benchmarks for video identifies critical areas for further research and innovation
authentication research and the metrics used to evaluate the to overcome existing challenges.
performance of different techniques, this section provides
a detailed overview of the tools and standards that underlie 4.4 RQ (4) What are the key challenges
this area of study. This exploration not only highlights the and limitations in applying deep learning
resources available to researchers but also underscores the techniques to video authentication,
importance of standardized evaluation methods to advance and what future efforts are recommended
video authentication. in the literature to overcome these obstacles?

4.3 RQ (3) What are the primary datasets An SLR on video authentication using deep learning tech‑
and evaluation metrics used in deep niques provides a comprehensive approach to synthesize
learning research for video authentication, current knowledge, identify research trends, and highlight
and how do these datasets influence model future directions in this rapidly evolving field. This effort
performance and generalizability? is crucial for advancing state-of-the-art video authentica‑
tion and serves a broader spectrum of stakeholders, from
Deep learning algorithms for video authentication detection, academics and students to industry professionals and poli‑
especially for identifying manipulations such as deepfakes cymakers. Figure 11 presents the key benefits of undertaking
and frame tampering, relies on diverse and comprehensive such a review.
datasets to train and benchmark algorithms. Figure 9 shows
the distribution of the datasets used in this SLR. •Consolidation of knowledge: This aggregates exten‑
In addition to these datasets, evaluation metrics in video sive video authentication research, making it easier for
authentication detection often involve accuracy, precision, newcomers and experienced researchers to find pertinent
recall, F1 score, and the area under the receiver operating studies and results in a single document.
characteristic curve. These metrics comprehensively evalu‑ •Identification of trends: By examining the progression
ate a model’s performance in distinguishing between authen‑ and outcomes of various studies, SLR can identify trends
tic and manipulated content. Figure 10 shows the distribu‑ in deep learning techniques that are most effective in the
tion of the datasets used in this SLR. field.
The selection of the dataset and benchmark depends on •Gap analysis: SLR can uncover gaps in current research,
the specific aspects of video authentication being studied, provide direction for future studies, and encourage inves‑
including the type of manipulation (e.g., deepfakes and tigation into underexplored areas.
Table 6  Summary of methods, evaluation measures, datasets, and manipulation types in this SLR
Ref Methods Used Evaluation Measures Dataset Manipulation Type Limitations

[50] RNN- PixelRNN Accuracy: 95.73% FaceForensics + + Deepfake detection Limited to a specific dataset
[19] RNN-LSTM, CNN -Incep‑ Recall: 95%, Precision: TRECVid 2001, 2007, and Video shot boundary detection Potential overfitting on training
tionV3 94.10%, F1 score: 94.4% RAI databases data
[38] CNN- InceptionResNet v2, Recall: 96%, Precision: DFDC Deepfake detection High computational cost
Xception 98.50%, F1 score: 98%
[36] DNN- white-box, black box XceptionNet: 98.07%, Mes‑ FaceForensics + + and DFDC Deepfake detection Limited generalization to new
oNet: 99.83% types of deepfakes
[93] CNN, ConvLSTM Accuracy: 99% FRAUD1 and FRAUD2 Facial video forgery detection Limited dataset variety
[59, 60] CWSA module, Deep CNN- Accuracy: 93.90% FaceForensics + + , Celeb- Interframe manipulations in Performance on low-resolution
EfficientNet B0 DF, and Deepfake Detection facial videos videos not tested
Challenge
[45] DNN Accuracy: 98% UADFV and Face Forensics + Deepfake detection Sensitive to noise in biological
+ signals
[81] Computer Vision-based DNN Accuracy: 98% Face2Face, FaceSwap, Deepfake detection High dependency on large
with Pairwise Learning 100KFaces, and DFDC datasets
[13] FFF-CDG—k-mean- capsuled Accuracy: 81.5%–95.82% FFHQ, 100K-Faces, DFFD, Deepfake detection Complexity in algorithm tuning
dual graph algorithm VGG-Face2, and WildDeep‑
fake
[30] 3DCNN Accuracy: 98%, Recall: 96%, UCF-101 and VIFFD Interframe video forgery Limited real-world applicability
Precision: 95%, F1 score: 97% detection
Video authentication detection using deep learning: a systematic literature review

[48] GANs Accuracy: 87.50% Kaggle Deepfake detection High computational resources
required
CNN Accuracy: 90.4%–98.5%, WLDR, DFDC, DF-TIMIT, Facial manipulations in videos, Potential bias in datasets
Recall: 85.7%–99.5%, Preci‑ FF + + , and Celeb-DF face-swap, deepfakes
sion: 89.7%–99.4%, F1 score:
87.5%–98.9%
[53] MTCNN, XceptionNet, RNN, Accuracy: 90.71% AUC: 99.2% on Celeb-DF Facial manipulations in videos, Generalization to unseen
3DCNN face-swap, deepfakes manipulations
DNN Accuracy: 94.72%, Recall: VTD Video tampering and manipu‑ Requires extensive preprocess‑
95%, Precision: 94.49%, F1 lation detection ing
score: 94.74%
[14] DPN Accuracy: 95.87%, Recall: XIPH, MCL-V, and MVQP Detection of fine-grained Sensitive to video compression
95.47%, Precision: 95.61%, manipulations in H.264 artifacts
F1 score: 95.54% videos
[72] Inflated 3D (I3D)- Siamese- VIRAT: Accuracy: 86.6%, VIRAT and Media Forensics Frame manipulations Limited scalability
based RNN MFC: 93.3%, F1 score: Challenge (MFC)
VIRAT: 86.5%, MFC: 93.3%
[33] CNN- multi-layer multilayer Accuracy: 92% Synthetic and KITTI Deepfake detection Not evaluated on real-world data
Page 19 of 30

perception structure with a

zero-hist layer
239
Table 6  (continued)
239

Ref Methods Used Evaluation Measures Dataset Manipulation Type Limitations

[55] Spatial Restore Detection Accuracy: 92.15%, AUC: FaceForensics + + , Celeb-DF, Facial manipulations in videos, High false positive rate
Framework (SRDF) 97.40% and DFDC face-swap, deepfakes
[32] Multiscale dual-flow CNN Recall: 96.34%, Precision: UCF101 Frame manipulations Limited robustness to diverse
Page 20 of 30

97.53%, F1 score: 96.93% manipulations

[31] GHOST—FaceShifter AEI-Net ID retrieval: 98.67%, Pose: FaceShifter AEI-Net Facial manipulations in videos, Limited dataset for evaluation
3.00 face-swap, deepfakes
[4] PN-CNN MISLnet, Efficient Average accuracy: 69.9%– Daxing, QUFVD Digital forensics and authen‑ Low accuracy on diverse
multiplier method (ADMM) 81.6% ticity verification of video datasets
footage
[82] CNN Accuracy: 86.49%, Perfor‑ DFD, Celeb-DF, and DFDC Facial manipulations in videos, Sensitive to video quality
mance: 74.03% face-swap, deepfakes
[103] CNN Accuracy: 84.50%, AUC: Celeb-DF-v2, DFDC, and Deepfake detection Limited evaluation metrics
0.927 FaceForensics + +
[25] Temporal convolution STConv, F1 score: 89.12, IoU: 85.42 DAVIS and YouTube-VOS Deep video inpainting tech‑ Limited to specific manipulation
HPFCN, STCA​ niques types
[73] CNNs- ResNet-50, Xception‑ AUC for FF + + : 0.994, AUC FaceForensics + + , DeeperFo‑ Facial manipulations in videos, Generalization issues
Net, InceptionNet, Efficient‑ for Celeb-DF: 0.669 rensics-1.0, and Celeb-DF face-swap, deepfakes
Net-v2-L
[99] Convolutional LSTM-based CLRNet: 93.22%, MesoNet: In-domain attacker (D), Deep‑ Facial manipulations in videos, Generalization to new datasets
Residual Network (CLRNet) 80.22% fake-in-the-Wild (DFW), and face-swap, deepfakes
Unknown deepfake genera‑
tion method (D)
[97] Deep CNN Accuracy: 97.56%, Sensitivity: N/A Interframe video forgery Requires further validation
95.48%, Specificity: 96.38% detection
[112] SegNet and t-SNE FF + + : Accuracy: 98.1%, FF + + Facial manipulations in videos, Not tested on diverse datasets
ROC: 0.99 face-swap, deepfakes
[109] Texture map preprocessing, co- Accuracy: 91%, AUC: FaceFo‑ FaceForensics + + and Celeb- Facial manipulations in videos, Limited evaluation metrics
occurrence matrix, and SVM rensics + + : 0.996, Celeb- DF face-swap, deepfakes
DF: 0.718
[57] MTCNN, facial pose estima‑ Accuracy: 0.587, EAO: 0.365, VOT-2015 Video content analysis Limited dataset variety
tion, and ImageNet Failure values: 0.920
A. A.-M. Alrawahneh et al.
Video authentication detection using deep learning: a systematic literature review Page 21 of 30 239

Fig. 9  Distribution of datasets used to detect video authentication in this SLR

Fig. 10  Distribution of evalua‑

tion metrics used to detect video
authentication in this SLR

•Methodological insights: By comparing different meth‑ lines to manage the ethical use of deepfakes and other
odologies, SLR can provide insights into the most robust manipulative technologies.
techniques and under what conditions. •Industry applications: For industry professionals, the
•Technological advancement: The SLR supports the SLR can guide the implementation of video authentica‑
advancement of technology by compiling and reviewing tion methods within their services to enhance security and
the effectiveness of various deep learning models, which trust in digital media platforms.
contributes to the development of sophisticated video •Encouraging collaboration: By highlighting areas
authentication tools. for further research and potential challenges, the SLR
•Educational resource: This resource serves as an edu‑ encourages collaboration among academics, industry
cational resource for students and professionals seeking professionals, and other stakeholders.
to understand the landscape of video authentication tech‑ •Standardization efforts: These findings can contribute
niques. to standardization efforts in video authentication, provid‑
•Policy and ethical guidelines: This review could inform ing a basis for developing common benchmarks and best
policymakers and ethicists about the state of video practices.
authentication, thereby aiding in the creation of guide‑
 239 Page 22 of 30 A. A.-M. Alrawahneh et al.

Fig. 11  Benefits of conducting

a systematic literature review of
video authentication using deep
learning techniques

The application of deep learning techniques to video Finally, this SLR examines the future directions and inno‑
authentication detection is a rapidly evolving research area vations indicated in the literature. By examining emerging
marked by significant advancements and potential. However, trends and cutting-edge advancements, this section provides
this domain faces a series of challenges and limitations that insights into the potential evolution of deep learning-based
must be addressed to further develop and refine these tech‑ video authentication. This analysis highlights promising
nologies. This SLR tackles these challenges, including the areas for future investigation and development and offers
need for dataset diversity and representativeness, the impor‑ a roadmap to enhance the effectiveness and robustness of
tance of temporal information and defense against adversar‑ video authentication techniques. Through this exploration,
ial attacks, the difficulty in detecting complex manipulation this SLR inspires continued innovation and progress in the
techniques, the necessity to broaden the scope of manipu‑ field.
lations covered by current models, and the imperative to
evaluate models against a broader range of manipulations. 4.5 RQ (5) What future trends and innovations
Addressing these issues is essential to enhance the effec‑ are emerging in deep learning‑based video
tiveness, reliability, and robustness of deep learning models authentication, and how might these
in video authentication, which will facilitate advancements advancements address current limitations
in security and trust in digital media. The challenges and and enhance detection capabilities?
limitations associated with applying deep learning to video
authentication detection, as mentioned in this SLR, are sum‑ Based on the challenges and limitations identified in apply‑
marized in Table 7. ing deep learning to video authentication detection, this SLR
The proposed method for video authentication detection indicates several future trends and innovations to enhance
using Deep Neural Networks (DNNs, including CNNs, LSTM the effectiveness of deep learning-based video authentica‑
networks, and GANs has significant potential to enhance the tion technologies:
accuracy and robustness of detection mechanisms. However,
like any advanced technology, such technology comes with •Exploration of sophisticated models and techniques:
certain limitations that must be acknowledged and addressed. Investigate more advanced deep learning models and
Table 8 provides a comprehensive overview of these limita‑ attention mechanisms to improve the accuracy and effec‑
tions, highlighting the challenges associated with computa‑ tiveness of video forgery detection [29, 67].
tional cost, data dependency, integration complexity, scal‑ -Use of advanced feature extraction and aggregation tech‑
ability, generalization, latency in real-time applications, and niques and different attention mechanisms for spatial fea‑
attention mechanisms for CNNs and LSTMs. Understanding ture extraction [32].
these limitations is crucial for developing more effective and -Exploring the use of hierarchical and ensemble models
efficient video authentication systems. for more robust forgery detection [1, 4].
By clearly outlining these limitations, we provide a bal‑ •Cross-dataset generalization:
anced perspective on the capabilities and challenges associ‑ -Develop models that perform well across different data‑
ated with the proposed method. This understanding will help sets to address dataset specificity issues [28].
guide future research and development efforts to overcome -Reducing dataset specificity for deepfakes: develop an
these limitations and enhance the effectiveness of video ensemble learning approach (leveraging multiple CNN
authentication systems.
Table 7  Challenges and limitations in applying deep learning to video authentication detection in this SLR
Challenge Category Key Points Future Efforts References

Dataset diversity and representativeness - Importance of diverse and representative datasets for effective - Develop broader, more inclusive datasets that cover a wide [18]
model training range of video types and manipulations
- Challenges include limited scope, bias, overfitting risks, under‑ - Implement strategies to minimize bias and enhance data privacy
represented manipulations, and data privacy issues
- Essential for preventing model bias and ensuring generalization
Temporal information and adversarial attacks - Necessity of using temporal information for detecting manipula‑ - Investigate and develop new methods to incorporate temporal [56]
tions over time analysis more effectively
- Challenges from adversarial attacks that target temporal aspects, - Enhance model resilience against adversarial attacks targeting
requiring development of robust defences temporal features
Complex manipulation techniques - Difficulty in detecting sophisticated deepfake videos with seam‑ - Advance research in multimodal deep learning to better [24]
less audio–video synchronization understand and analyse the interplay between audio and video
Video authentication detection using deep learning: a systematic literature review

- Need for models that can analyse complex interplays between elements
audio and video elements - Leverage additional features like audio analysis and textual
content
Scope of manipulations - Current focus primarily on facial manipulations, overlooking - Broaden the detection frameworks to include various video [63]
other tampering forms like background alterations, object inser‑ manipulations
tion and removal, and audio changes - Develop models capable of multimodal analysis (visual, audio,
- Expansion of detection frameworks are needed and textual) and understanding context changes
Evaluation on other manipulations - Necessity of testing models on a broader range of manipulations - Enhance datasets with diverse manipulations for testing [10, 72]
than originally trained for - Conduct cross-domain evaluations to identify vulnerabilities
- Expanding the evaluation scope to improve model robustness and improve the security and reliability of authentication
and adaptability technologies
Page 23 of 30
239
 239
Page 24 of 30

Table 8  Strengths and limitations of various deep learning techniques

Technique Strengths Limitations Effectiveness Efficiency

Convolutional neural networks - Excellent at hierarchical feature - Struggle with capturing temporal - Highly effective for spatial feature - Efficient in processing individual
(CNNs) extraction from video frames dependencies across frames extraction and detecting visual frames but can be resource-intensive
- Highly effective in capturing spa‑ - Can become computationally anomalies for high-resolution or large datasets
tial relationships within frames expensive for high-resolution
- Efficient processing of large video videos or large datasets
data
Recurrent neural networks (RNNs) - Designed to capture temporal - Difficult to train because of issues - Highly effective for tasks requir‑ - Less efficient in terms of process‑
including LSTMs dependencies and sequential pat‑ like vanishing and exploding ing understanding of temporal ing speed, particularly for real-time
terns gradients sequences, such as action recogni‑ applications
- Effective for detecting temporal - Require significant computa‑ tion and deepfake detection over
inconsistencies in video streams tional resources for training long time
sequences
Deep neural networks (DNNs) - Versatile and adaptable for various - Require large amounts of labeled - Effective across various tasks -Computationally demanding, affect‑
video authentication tasks data to achieve high performance because of versatility and complex ing efficiency
- Capable of learning complex pat‑ - High computational demands, pattern recognition
terns from video data necessitating powerful hardware
and extensive training time
Generative adversarial networks - Powerful for generating realistic - Challenging to train because of - Effective for both generating - Resource-intensive and can be inef‑
(GANs) synthetic videos and detecting issues like mode collapse and and detecting video forgeries, ficient in terms of training time and
forgeries instability enhancing detection in adversarial computational requirements
- Adversarial training improves - Resource-intensive training process settings
detection capabilities by challeng‑ requiring extensive computational
ing the discriminator with new power and time
synthetic data
A. A.-M. Alrawahneh et al.
Video authentication detection using deep learning: a systematic literature review Page 25 of 30 239

models or combining CNNs with other types of deep Improve the transferability and robustness of GAN
learning architectures to create a more robust system) to against adversarial attacks [15].
improve generalizability across datasets [73].
-Use unsupervised domain adaptation to improve the These new ideas and directions for the future are intended
model’s ability to detect forgeries in unseen domains [53]. to solve the problems currently occurring in deep learning-
•Real-time detection and scalability [35]: based video authentication by improving the models’ reli‑
-Improve the speed and efficiency of detection algorithms ability, improving the datasets, and making it easier to
to realize real-time video authentication. identify changes. Deep learning methods, such as CNN,
-Enhance scalability for practical applications in diverse LSTM, DNN, and GAN, indicate a need for improvement
environments. or enhancement to address video authentication and forgery
•Temporal analysis of enhanced detection: detection challenges. Each method is associated with cer‑
Implement temporal analysis and long-distance interac‑ tain research projects, indicating future directions to enhance
tions to detect sophisticated deepfakes that may involve their capabilities.
subtle manipulations over time [20].
The incorporation of temporal information and contras‑ -CNNs and LSTM networks are highlighted for their
tive loss functions into the CNN-LSTM network archi‑ potential in object-based video forgery detection and
tecture for deepfake video embedding to improve clas‑ localization, and advancements in these models can
sification [86]. improve detection accuracy and efficiency.
•Improvement of network performance via more effec‑ -GAN is mentioned in the context of ensemble learning
tive loss functions and streamlined structures to reduce for face manipulation detection, indicating an interest in
inference times [55]. further exploring GANs’ capabilities in generating and
•Improving low-quality and sophisticated deepfake detection: detecting sophisticated forgeries.
-Improve methods for detecting low-quality deepfakes -DNN is associated with deepfake detection using cas‑
and sophisticated forgeries that may not be easily iden‑ caded deep sparse autoencoders and CNNs, indicating a
tifiable [55]. need to design more complex and layered approaches for
-Explore the use of signal enhancement, denoising, and effective detection mechanisms.
other preprocessing techniques for effective detection
[45]. By focusing on these areas, the field can advance toward
•The incorporation of multimodal cues: more secure and reliable video authentication technologies
-For comprehensive forgery detection, include additional to combat the evolving digital video manipulation landscape.
biometric and nonbiometric cues, such as audio and cam‑ In conclusion, this SLR summarizes the key findings and
era noise patterns [51, 66]. highlights the implications for future research. By providing
-Investigate the balance between local and global features a concise synthesis of the insights gained from the compre‑
in forgery detection, using motion cues and facial decom‑ hensive review of the literature, this section underscores the
position techniques [28]. main contributions of the study. Additionally, it discusses
•Improvement in error rates and computational efficiency the broader implications of these findings relative to video
for systems using the VIOLA–Jones algorithm for face authentication, offers recommendations for future research
detection [104]. directions, and identifies areas that require further investiga‑
Enhancing model robustness and exploring video com‑ tion. This summary consolidates the knowledge presented in
pression and unseen forgery methods to improve detec‑ this SLR and guides subsequent efforts to advance the field.
tion capability [108].
Exploration of parts-based detectors, U-net architecture
variants, and global approaches for improved detection 5 Conclusion
on diverse datasets [90].
Extension of ResNet and LSTM detection methods to This study meticulously explored the application of deep
include relocated I-frames and finer-grained coding infor‑ learning techniques in video analysis and conducted a thor‑
mation analysis [34]. ough evaluation of existing research using SLR. Our review
Exploration of spatial and temporal dependencies using incorporates 99 research articles, all of which have been
other network architectures like RNNs or 3D CNNs [33]. published in peer-reviewed journals or presented at confer‑
•Explore the application of EfficientNet to deepfake ences. Over the last five years, this SLR has been developed
detection, focusing on specialization and adaptability by establishing RQs and systematically addressing them to
[23]. shed light on various video manipulations.
 239 Page 26 of 30 A. A.-M. Alrawahneh et al.

We have discovered that advanced deep learning tech‑ org/​10.​1109/​WIFS4​9906.​2020.​93609​04 WE - Conference Pro‑
niques can significantly enhance the detection of video ceedings Citation Index - Science (CPCI-S)
4. Akbari Y, Al Maadeed S, Elharrouss O, Ottakath N, Khelifi F
authenticity and tampering. A notable finding is the predomi‑ (2024) Hierarchical deep learning approach using fusion layer
nant use of the FaceForensics + + dataset in most experi‑ for Source Camera Model Identification based on video taken
ments, highlighting its importance in the field. Deep learning by smartphone. Expert systems with applications. 238. https://​
models, especially CNNs, are the backbone of most analysed doi.​org/​10.​1016/j.​eswa.​2023.​121603
5. Akhtar N, Saddique M, Asghar K, Bajwa UI, Hussain M, Habib
studies. Following the CNN, other significant models include Z (2022) Digital video tampering detection and localization:
combinations of the CNN with LSTM (CNN-LSTM), DNNs, review, representations, challenges and algorithm. Mathematics
and GANs. Among the evaluated metrics, detection accuracy 10(2):168
is the most frequently used statistical performance measure. 6. Alameri SA, Mohd M (2021) Comparison of fake news detection
using machine learning and deep learning techniques. In: 2021
The evolution of studies focusing on CNNs, CNN-LSTM, 3rd International Cyber Resilience Conference (CRC). IEEE; p.
DNNs, and GANs has provided a clear path for future video 1–6. Available from: https://​www.​ftsm.​ukm.​my/​v5/​file/​resea​rch/​
authenticity analysis and manipulation detection research. techn​icalre​ port/​PS-​FTSM-​2020-​032.​pdf. Accessed 12 Dec 2024
Despite rapid technological advancements in multimedia 7. Alherbawi N, Shukur Z, Sulaiman R (2017) Current techniques
in JPEG image authentication and forgery detection. J Eng Appl
and the proliferation of accessible tools and applications, Sci 12(1):104–112
challenges in video authentication detection remain preva‑ 8. Ali NH, Harun F (2019) Video authentication using shot boundary
lent. We are optimistic that this SLR will serve as a valuable detection technique—An overview. J Phys Conf Ser 1358(1):012061.
resource for the scientific community and aid in developing Available from: https://​iopsc​ience.​iop.​org/​artic​le/​10.​1088/​1742-​
6596/​1358/1/​012061/​meta. Accessed 12 Dec 2024
effective detection techniques and countermeasures. 9. Alrawahneh A, Safei S (2021) A model of video watching
concentration level measurement among students using head
Acknowledgements The authors thank the Ministry of Higher Edu‑ pose and eye tracking detection. J Theor Appl Inform Technol
cation, Malaysia, for funding support Fundamental Research Grant 99(17):4305–4315. Available from: https://​www.​jatit.​org/​volum​
Scheme FRGS/1/2019/ICT02/UKM/02/9. A special appreciation to es/​Vol99​No17/​8Vol9​9No17.​pdf. Accessed 12 Dec 2024
Cybersecurity Malaysia for collaborating with Universiti Kebangsaan 10. Alsaheel A, Alhassoun R, Alrashed R, Almatrafi N, Almallouhi
Malaysia (UKM) under a research grant TT-2024-010. N, Albahli S (2023) Deep Fakes in Healthcare: How Deep Learn‑
ing Can Help to Detect Forgeries. CMC-COMPUTERS MATE‑
Author contributions Ayat and Asyikin developed the search strategy; RIALS & CONTINUA, 76(2), 2461–2482. https://​doi.​org/​10.​
Ayat and Sarah screened the articles and extracted the data; Ayat and 32604/c​ mc.2​ 023.0​ 40257 WE - Science Citation Index Expanded
Dr. Hafizah conducted the quality assessment; Ayat drafted the manu‑ (SCI-EXPANDED)
script; Dr. Huda reviewed and edited the manuscript for critical content. 11. Amerini I, Anagnostopoulos A, Maiano L, Celsi L, IEEE (2021)
Learning double-compression video fingerprints left from social-
Data availability All articles reviewed are publicly available from the media platforms. https://​doi.​org/​10.​1109/​ICASS​P39728.​2021.​94133​
databases listed in the methodology section. Full references are pro‑ 66 WE - Conference Proceedings Citation Index - Science (CPCI-S)
vided in the reference list. 12. Anbu T, Milton Joe M, Murugeswari G (2022) A comprehensive
survey of detection of tampered video and localization of tam‑
Declarations pered frame. Wirel Pers Commun 125(2):1–34. Available from:
https://​link.​sprin​ger.​com/​artic​le/​10.​1007/​s11277-​021-​09227-z.
Ethical and informed consent This systematic review synthesizes Accessed 12 Dec 2024
data from previously published studies; therefore, no primary ethical 13. Arunkumar P, Sangeetha Y, Raja P, Sangeetha S (2022) Deep
approval or informed consent was required. All analysed studies were Learning for Forgery Face Detection Using Fuzzy Fisher Cap‑
assessed for ethical standards and quality. sule Dual Graph. Information Technol Control 51(3):563–574.
https://​doi.​org/​10.​5755/​j01.​itc.​51.3.​31510​WE-​Scien​ceCit​ation​
Competing interests The authors declare that they have no known Index​Expan​ded(SCI-​EXPAN​DED)
competing financial interests or personal relationships that could have 14. Bae W, Nam S, Yu I, Kwon M, Yoon M, Lee H (2021)
appeared to influence the work reported in this systematic review. Dual-path convolutional neural network for classifying fine-
grained manipulations in H264 videos. Multimedia Tools and
Applications 80(20):30879–30906. https://​doi.​org/​10.​1007/​
s11042-​021-​10552-5
References 15. Baek J, Yoo Y, Bae S (2020) Generative Adversarial Ensem‑
ble Learning for Face Forensics. IEEE Access 8:45421–45431.
https://​doi.​org/​10.​1109/​ACCESS.​2020.​29686​12WE-​Scien​ceCit​
1. Abbas Q, Alghamdi T, Alsaawy Y, Alyas T, Alzahrani A, Malik ation​Index​Expan​ded(SCI-​EXPAN​DED)
K, Bibi S (2023) Reducing Dataset Specificity for Deepfakes 16. Balasubramanian S, Kannan R, Prabu P, Venkatachalam K, Tro‑
Using Ensemble Learning. CMC-Computers materials & con‑ jovsky P (2022) Deep fake detection using cascaded deep sparse
tinua. 74(2). https://​doi.​org/​10.​32604/​cmc.​2023.​034482. WE - auto-encoder for effective feature selection. PEERJ COMPUTER
Science Citation Index Expanded (SCI-EXPANDED) SCIENCE, 8. https://​doi.​org/​10.​7717/​peerj-​cs.​1040 WE - Sci‑
2. Academic Press Inc (2024) Elsevier Science. Elsevier. Avail‑ ence Citation Index Expanded (SCI-EXPANDED)
able from: https://​books​ite.​elsev​ier.​com/​broch​ures/​acade​micpr​ 17. Bei Y, Lou H, Geng J, Liu E, Cheng L, Song J, Song M, Feng
ess/. Accessed 12 Dec 2024 Z (2024) A large-scale universal evaluation benchmark for face
3. Agarwal S, Farid H, El-Gaaly T, Lim S, IEEE (2020) Detecting forgery detection. arXiv preprint arXiv:2406.09181. Available
Deep-Fake Videos from Appearance and Behavior. https://​doi.​ from: https://​arxiv.​org/​abs/​2406.​09181. Accessed 12 Dec 2024
Video authentication detection using deep learning: a systematic literature review Page 27 of 30 239

18. Bekci B, Akhtar Z, Ekenel H, IEEE (2020) Cross-Dataset Face Deep Spatio-Temporal Representations in Compression
Manipulation Detection. Domain. IEEE Trans Multimedia 23:3179–3192. https://​doi.​
19. Benoughidene A, Titouna F (2022) A novel method for video org/​10.​1109/​T MM.​2020.​30212​34WE-​Scien​ceCit​ation​Index​
shot boundary detection using CNN-LSTM approach. Int J Mul‑ Expan​ded(SCI-​EXPAN​DED)
timedia Inform Retrieval 11(4):653–667. https://d​ oi.o​ rg/1​ 0.1​ 007/​ 35. Host K, Pobar M, Ivasic-Kos M (2023) Analysis of Move‑
s13735-​022-​00251-8 ment and Activities of Handball Players Using Deep Neural
20. Bonettini N, Cannas E, Mandelli S, Bondi L, Bestagini P, Tubaro S, Networks. J Imaging. 9(4). https://​doi.​org/​10.​3390/​jimag​ing90​
SOCIC (2021) Video Face Manipulation Detection Through Ensem‑ 40080 WE - Emerging Sources Citation Index (ESCI)
ble of CNNs. https://​doi.​org/​10.​1109/​ICPR4​8806.​2021.​94127​11 36. Hussain S, Neekhara P, Jere M, Koushanfar F, McAuley J,
WE - Conference Proceedings Citation Index - Science (CPCI-S) IEEE (2021) Adversarial Deepfakes: Evaluating Vulnerabil‑
21. Christian A, Sheth R (2016) Digital video forgery detection and ity of Deepfake Detectors to Adversarial Examples. https://​
authentication technique-a review. Int J Scientific Res Sci Tech‑ doi.​org/​10.​1109/​WACV4​8630.​2021.​00339 WE - Conference
nol (IJSRST) 2(6):138–143 Proceedings Citation Index - Science (CPCI-S)
22. Ciftci UA, Demir I, Yin L (2020) How do the hearts of deep fakes 37. IEEE (2024) IEEE Xplore. Available from: https://​ieeex​plore.​
beat? Deep fake source detection via interpreting residuals with ieee.​org/. Accessed 12 Dec 2024
biological signals. arXiv e-prints. Available from: https://​ieeex​ 38. Ikram S, Priya V, Chambial S, Sood D, Arulkumar V (2023)
plore.​ieee.​org/​abstr​act/​docum​ent/​93049​09. Accessed 12 Dec 2024 A Performance Enhancement of Deepfake Video Detection
23. Coccomini D, Caldelli R, Falchi F, Gennaro C, Amato G, ACM through the use of a Hybrid CNN Deep Learning Model. Int J
(2022) Cross-Forgery Analysis of Vision Transformers and electrical Computer Eng Systems. 14(2):169–178 WE - Emerg‑
CNNs for Deepfake Image Detection. https://​doi.​org/​10.​1145/​ ing Sources Citation Index (ESCI).
35127​32.​35335​82 WE - Conference Proceedings Citation Index 39. INC NP (2024) Now Publishers Inc. 2024. Available from: https://​
- Science (CPCI-S) www.​nowpu​blish​ers.​com/​Home . Accessed 12 Dec 2024
24. Demir I, Çiftçi U, ACM (2021) Where Do Deep Fakes Look? 40. INC P (2014) PeerJ Inc. Available from: https://​peerj.​com/.
Synthetic Face Detection via Gaze Tracking. https://​doi.​org/​10.​ Accessed 12 Dec 2024
1145/​34480​17.​34573​87WE-​Confe​rence​Proce​eding​sCita​tionI​ 41. Intelligence AAA (2024) Association for the Advance‑
ndex-​Scien​ce(CPCI-S) ment of Artificial Intelligence. Available from: https://​aaai.​
25. Ding X, Pan Y, Luo K, Huang Y, Ouyang J, Yang G, IEEE (2021) org/. Accessed 12 Dec 2024
Localization of Deep Video Inpainting Based on Spatiotemporal 42. International Information and Engineering Technology Asso‑
Convolution and Refinement Network. https://​doi.​org/​10.​1109/​ ciation (IIETA) (2024) IIETA. Available from: https://​iieta.​
ISCAS5​ 1556.2​ 021.9​ 40167​ 5 WE - Conference Proceedings Cita‑ org/. Accessed 12 Dec 2024
tion Index - Science (CPCI-S) 43. JJ Strossmayer University of Osijek, Faculty of Electrical
26. Elsevier (2024) Elsevier. Available from: https://​www.​elsev​ier.​ Engineering (2024) International Journal of Engineering and
com/. Accessed 12 Dec 2024 Computer Science (IJECE). Available from: https://​www.​etfos.​
27. Fadl S, Han Q, Li Q (2021) CNN spatiotemporal features unios.​hr/​ijeces/. Accessed 12 Dec 2024
and fusion for surveillance video forgery detection. SIGNAL 44. Jegaveerapandian LM, Rani AJ, Periyaswamy P, Velusamy
PROCESSING-IMAGE COMMUNICATION, 90. https://​doi.​ S (2023) A survey on passive digital video forgery detection
org/​10.​1016/j.​image.​2020.​116066 WE - Science Citation Index techniques. International Journal of Electrical & Computer
Expanded (SCI-EXPANDED) Engineering (2088–8708), 13(6).
28. Fei J, Xia Z, Yu P, Xiao F (2021) Exposing AI-generated 45. Jin X, Ye D, Chen C (2021) Countering Spoof: Towards
videos with motion magnification. Multimedia Tools and Detecting Deepfake with Multidimensional Biological Sig‑
Applications 80(20):30789–30802. https://​doi.​org/​10.​1007/​ nals. Security Commun Networks. 2021. https://​doi.​org/​10.​
s11042-​020-​09147-3 1155/​2021/​66269​74 WE - Science Citation Index Expanded
29. Ganguly S, Mohiuddin S, Malakar S, Cuevas E, Sarkar R (SCI-EXPANDED)
(2022) Visual attention-based deepfake video forgery detec‑ 46. Johnston P, Elyan E, Jayne C (2020) Video tampering locali‑
tion. Pattern Anal App 25(4):981–992. https://​doi.​org/​10.​1007/​ sation using features learned from authentic content. Neural
s10044-​022-​01083-2 Comput Appl 32(16):12243–12257. https://​doi.​org/​10.​1007/​
30. Gowda R, Pawar D (2023) Deep learning-based forgery identifi‑ s00521-​019-​04272-​zWE-​Scien​ceCit​ation​Index​Expan​ded(SCI-​
cation and localization in videos. Signal Image Video Processing EXPAN​DED)
17(5):2185–2192. https://​doi.​org/​10.​1007/​s11760-​022-​02433-7 47. Joshi V, Jain S (2015) Tampering detection in digital video—a
31. Groshev A, Maltseva A, Chesakov D, Kuznetsov A, Dimitrov review of temporal fingerprints based techniques. In: 2015 2nd
D (2022) GHOST-A New Face Swap Approach for Image and International Conference on Computing for Sustainable Global
Video Domains. IEEE Access 10:83452–83462. https://​doi.​org/​ Development (INDIACom). IEEE; p. 1121–1124. Available
10.​1 109/​ACCESS.​2 022.​3 1966​6 8WE-​Scien​c eCit ​a tion​I ndex​ from: https://​i eeex​p lore.​i eee.​o rg/​a bstr ​a ct/​d ocum​e nt/​7 1004​
Expan​ded(SCI-​EXPAN​DED) 22. Accessed 12 Dec 2024
32. Gu Q, Ding X, Zhang D, Yang C, IEEE (2022) Forgery Detec‑ 48. Jung T, Kim S, Kim K (2020) DeepVision: Deepfakes Detec‑
tion Scheme of Deep Video Frame-rate Up-conversion Based tion Using Human Eye Blinking Pattern. IEEE Access
on Dual-stream Multi-scale Spatial-temporal Representation. 8:83144–83154. https://​doi.​org/​10.​1109/​ACCESS.​2020.​29886​
https://d​ oi.​org/​10.​1109/​Trust​Com56​396.​2022.​00104 WE - Con‑ 60WE-​Scien​ceCit​ation​Index​Expan​ded(SCI-​EXPAN​DED)
ference Proceedings Citation Index - Science (CPCI-S) 49. Kandasamy V, Hubálovsky S, Trojovsky P (2022) Deep fake
33. He P, Li H, Li B, Wang H, Liu L (2020) Exposing Fake Bitrate detection using a sparse auto encoder with a graph capsule
Videos Using Hybrid Deep-Learning Network From Rec‑ dual graph CNN. Peerj Computer Sci. 8. https://​doi.​org/​10.​
ompression Error. IEEE Trans Circuits Syst Video Technol 7717/​p eerj-​cs.​9 53 WE - Science Citation Index Expanded
30(11):4034–4049. https://​doi.​org/​10.​1109/​TCSVT.​2019.​29516​ (SCI-EXPANDED)
30WE-​Scien​ceCit​ation​Index​Expan​ded(SCI-​EXPAN​DED) 50. Khodabakhsh A, Busch C (2020) A generalizable deepfake detec‑
34. He P, Li H, Wang H, Wang S, Jiang X, Zhang R (2021) Frame- tor based on neural conditional distribution modelling. In: 2020
Wise Detection of Double HEVC Compression by Learning International Conference of the Biometrics Special Interest Group
239 Page 28 of 30 A. A.-M. Alrawahneh et al.

(BIOSIG). IEEE; p. 1–5. Available from: https://​ieeex​plore.​ieee.​ 67. Mehra A, Spreeuwers L, Strisciuglio N (2021). Deepfake Detec‑
org/​abstr​act/​docum​ent/​92110​32. Accessed 12 Dec 2024 tion using Capsule Networks and Long Short-Term Memory Net‑
51. Kingra S, Aggarwal N, Kaur N (2023) SiamNet: Exploiting works. https://​doi.​org/​10.​5220/​00102​89004​07041​4WE-​Confe​
source camera noise discrepancies using Siamese Network for rence​Proce​eding​sCita​tionI​ndex-​Scien​ce(CPCI-S)
Deepfake Detection. Information Sciences, 645. https://​doi.​org/​ 68. Mercan S, Cebe M, Aygun RS, Akkaya K, Toussaint E, Danko
10.​1016/j.​ins.​2023.​119341 D (2021) Blockchain-based video forensics and integrity veri‑
52. Koshy L, Ajay S, Paul A, Hariharan V, Basheer A (2021) Video fication framework for wireless Internet-of-Things devices.
forgery detection using CNN. In: 2021 Smart Technologies, Security and Privacy 4(2):e143
Communication and Robotics (STCR). IEEE; p. 1–6. Avail‑ 69. Mizher MA, Ang MC, Abdullah SNHS, Ng KW, Mazhar AA,
able from: https://​ieeex​plore.​ieee.​org/​abstr​act/​docum​ent/​95888​ Mizher MAA (2021) Passive object-based video authentication
60. Accessed 12 Dec 2024 using stereo statistical descriptor on wavelet decomposition.
53. Kumar A, Bhavsar A, Verma R, IEEE (2020) Detecting Deep‑ In: 2021 International Conference on Information Technology
fakes with Metric Learning. https://​doi.​org/​10.​1109/​iwbf4​9977.​ (ICIT). IEEE; p. 791–798. Available from: https://​ieeex​plore.​
2020.​91079​62 WE - Conference Proceedings Citation Index - ieee.​org/​abstr​act/​docum​ent/​94917​47. Accessed 12 Dec 2024
Science (CPCI-S) 70. Mohiuddin S, Malakar S, Kumar M, Sarkar R (2023) A compre‑
54. Kumar V, Gaur M, Kansal V (2022) Deep feature based forgery hensive survey on state-of-the-art video forgery detection tech‑
detection in video using parallel convolutional neural network: niques. Multimedia Tools and Applications 82(22):33499–33539
VFID-Net. Multimedia Tools Applications 81(29):42223–42240. 71. Montserrat D, Hao H, Yarlagadda S, Baireddy S, Shao R, Hor‑
https://​doi.​org/​10.​1007/​s11042-​021-​11448-0 vath J, Bartusiak E, Yang J, Guera D, Zhu F, Delp E, SOCIC
55. Li Y, Bian S, Wang C, Polat K, Alhudhaif A, Alenezi F (2023) (2020) Deepfakes Detection with Automatic Face Weighting.
Exposing low-quality deepfake videos of Social Network Service https://d​ oi.o​ rg/1​ 0.1​ 109/C
​ VPRW5​ 0498.2​ 020.0​ 0342 WE - Confer‑
using Spatial Restored Detection Framework. Expert Systems with ence Proceedings Citation Index - Science (CPCI-S)
Applications. 231. https://​doi.​org/​10.​1016/j.​eswa.​2023.​120646 72. Munawar M, Noreen I (2021) Duplicate Frame Video Forgery
56. Lin D, Tondi B, Li B, Barni M, IEEE (2022) Exploiting temporal Detection Using Siamese-based RNN. Intelligent Automation
information to prevent the transferability of adversarial examples Soft Computing. 29(3):927–937. https://​d oi.​o rg/​1 0.​3 2604/​
against deep fake detectors. https://​doi.​org/​10.​1109/​IJCB5​4206.​ iasc.​2 021.​0 18854 WE - Science Citation Index Expanded
2022.​10007​959 WE - Conference Proceedings Citation Index - (SCI-EXPANDED)
Science (CPCI-S) 73. Nadimpalli A, Rattani A, IEEE (2022) On Improving Cross-
57. Liu B, Wu M, Tao M, Wang Q, He L, Shen G, Chen K, Yan J dataset Generalization of Deepfake Detectors. https://​doi.​org/​10.​
(2020) Video Content Analysis for Compliance Audit in Finance 1109/​CVPRW​56347.​2022.​00019 WE - Conference Proceedings
and Security Industry. IEEE Access 8:117888–117899. https://​ Citation Index - Science (CPCI-S)
doi.​org/​10.​1109/​ACCESS.​2020.​30058​25WE-​Scien​ceCit ​ation​ 74. Nirkin Y, Wolf L, Keller Y, Hassner T (2022) DeepFake Detec‑
Index​Expan​ded(SCI-​EXPAN​DED) tion Based on Discrepancies Between Faces and Their Con‑
58. LTD P-ES (2024) Pergamon-Elsevier Science Ltd. Available text. IEEE Transactions Pattern Anal Machine Intelligence
from: Available from: https://​www.​elsev​ier.​com/. Accessed 12 44(10):6111–6121. https://​doi.​org/​10.​1109/​TPAMI.​2021.​30934​
Dec 2024 46WE-​Scien​ceCit​ation​Index​Expan​ded(SCI-​EXPAN​DED)
59. Lu Y, Liu Y, Fei J, Xia Z (2021) Channel-wise spatiotemporal 75. Okoli C, Schabram K (2015) A guide to conducting a systematic
aggregation technology for face video forensics. Security Com‑ literature review of information systems research. Available from:
mun Networks 2021:1–13 https://d​ 1wqtx​ ts1xz​ le7.c​ loudf​ ront.n​ et/3​ 25066​ 6/O
​ koliS
​ chabr​ am20​
60. Lu Y, Liu Y, Fei J, Xia Z (2021) Channel-wise spatiotemporal 10Spr​outsL​itRev​iewGu​ide-​libre.​pdf. Accessed 12 Dec 2024
aggregation technology for face video forensics. Security and 76. Pan Z, Yu W, Yi X, Khan A, Yuan F, Zheng Y (2019) Recent
Commun Networks. 2021. https://​doi.​org/​10.​1155/​2021/​55249​ Progress on Generative Adversarial Networks (GANs): A Survey.
30 WE - Science Citation Index Expanded (SCI-EXPANDED) IEEE Access 7:36322–36333
61. Lyu S (2020) Deepfake detection: Current challenges and next 77. Passos LA, Jodas D, Costa KA, Souza Júnior LA, Rodrigues D,
steps. In: 2020 IEEE International Conference on Multimedia Del Ser J, Camacho D, Papa JP (2024) A review of deep learn‑
& Expo Workshops (ICMEW). IEEE; p. 1–6. Available from: ing‐based approaches for deepfake content detection. Expert Syst
https://i​ eeexp​ lore.i​ eee.o​ rg/a​ bstra​ ct/d​ ocume​ nt/9​ 10599​ 1. Accessed 41(8):e13570. Available from: https://​onlin​elibr​ary.​wiley.​com/​
12 Dec 2024 doi/​abs/​10.​1111/​exsy.​13570. Accessed 12 Dec 2024
62. Association for Computing Machinery (ACM) (2024) Asso‑ 78. Press TS (2024) Tech Science Press. Available from: https://​
ciation for Computing Machinery. Available from: https://​www.​ www.​techs​cience.​com/. Accessed 12 Dec 2024
acm.​org/. Accessed 12 Dec 2024 79. PROCESSING-EURASIP EASSI (2024) EUROPEAN ASSOC
63. Manisha Li C, Kotegar K (2023) Source Camera Identification SIGNAL SPEECH & IMAGE PROCESSING-EURASIP.
with a Robust Device Fingerprint: Evolution from Image-Based Retrieved 30/06/2024 from https://​euras​ip.​org/
to Video-Based Approaches. Sensors, 23(17). https://​doi.​org/​ 80. Publishing A (2024). AIP Publishing. Available from: https://​
10.​3390/​s2317​7385 WE - Science Citation Index Expanded pubs.​aip.​org/​pages/​about. Accessed 12 Dec 2024
(SCI-EXPANDED) 81. Ram R, Kumar M, Al-shami T, Masud M, Aljuaid H, Abouhaw‑
64. Maras M-H, Alexandrou A (2019) Determining authenticity of wash M (2023) Deep Fake Detection Using Computer Vision-
video evidence in the age of artificial intelligence and in the wake Based Deep Neural Network with Pairwise Learning. Intelligent
of Deepfake videos. Int J Evidence Proof 23(3):255–262 Automation Soft Computing. 35(2):2449–2462. https://​doi.​
65. MDPI (2024) MDPI. Available from: https:// ​ w ww. ​ m dpi.​ org/​10.​32604/​iasc.​2023.​030486 WE - Science Citation Index
com/. Accessed 12 Dec 2024 Expanded (SCI-EXPANDED)
66. Megahed A, Han Q (2022) Identify videos with facial manipula‑ 82. Ranjan P, Patil S, Kazi F, IEEE (2020) Improved Generalizability
tions based on convolution neural network and dynamic texture. of Deep-Fakes Detection Using Transfer Learning Based CNN
Multimedia Tools and Applications 81(30):43441–43466. https://​ Framework. https://​doi.​org/​10.​1109/​ICICT​50521.​2020.​00021
doi.​org/​10.​1007/​s11042-​022-​13102-9 WE - Conference Proceedings Citation Index - Science (CPCI-S)
Video authentication detection using deep learning: a systematic literature review Page 29 of 30 239

83. Raskar P, Shah S (2021) Real time object-based video forgery 101. Teymourzadeh R, Laadi AA, Samir Y, Othman M (2018) Design
detection using YOLO (V2). forensic Sci Int. 327. https://d​ oi.o​ rg/​ an advanced computer-aided tool for image authentication and
10.​1016/j.​forsc​iint.​2021.​110979 classification. arXiv preprint arXiv:1808.02085. Available
84. Rodriguez-Ortega Y, Ballesteros D, Renza D (2021) Copy-Move from: https://​arxiv.​org/​abs/​1808.​02085
Forgery Detection (CMFD) Using Deep Learning for Image and 102. The Ho QN, Do TT, Minh PS, Nguyen V-T, Nguyen VTT (2023)
Video Forensics. Journal of Imaging, 7(3). https://​doi.​org/​10.​ Turning chatter detection using a multi-input convolutional neu‑
3390/​jimag​ing70​30059 WE - Emerging Sources Citation Index ral network via image and sound signal. Machines 11(6):644
(ESCI) 103. Tran V, Lee S, Le H, Kim B, Kwon K, IEEE (2023) Learning
85. Rosli NA, Abdullah SNHS, Zamani AN, Ghazvini A, Othman Face Forgery Detection in Unseen Domain with Generalization
NSM, Tajuddin NAAAM (2021) Comparison multi transfer Deepfake Detector. https://​doi.​org/​10.​1109/​ICCE5​6470.​2023.​
learning models for deep fake image recognizer. In: 2021 3rd 10043​436 WE - Conference Proceedings Citation Index - Sci‑
International Cyber Resilience Conference (CRC). IEEE; p. 1–6. ence (CPCI-S)
Available from: https://​ieeex​plore.​ieee.​org/​abstr​act/​docum​ent/​ 104. Venkatachalam K, Trojovsky P, Hubálovsky S (2023) VIOLA
93925​66. Accessed 12 Dec 2024 jones algorithm with capsule graph network for deepfake detec‑
86. Saif S, Tehseen S, Ali S, Kausar S, Jameel A (2022) Generalized tion. Peerj Computer Sci. 9. https://​doi.​org/​10.​7717/​peerj-​cs.​
Deepfake Video Detection Through Time-Distribution and Met‑ 1313 WE - Science Citation Index Expanded (SCI-EXPANDED)
ric Learning. IT PROFESSIONAL 24(2):38–44. https://​doi.​org/​ 105. Verdoliva L (2018) Deep learning in multimedia forensics. In:
10.​1109/​MITP.​2022.​31683​51WE-​Scien​ceCit​ation​Index​Expan​ Proceedings of the 6th ACM Workshop on Information Hiding
ded(SCI-​EXPAN​DED) and Multimedia Security. ACM; p. 3–3. Available from: https://​
87. Saini P, Ahuja R (2022) Watermarked Hashing As a Video Con‑ dl.a​ cm.o​ rg/d​ oi/a​ bs/1​ 0.1​ 145/3​ 20600​ 4.3​ 20602​ 4. Accessed 12 Dec
tent Authentication Technique. ECS Trans 107(1):5211 2024
88. Saini P, Ahuja R, Kaur A (2021) A review on video authentica‑ 106. Vijayalakshmi V, Vasitha R, Vanitha V, Gayathri M (2016) A
tion technique exploiting watermarking methods. In: 2021 9th survey on real-time authentication and tampering detection in
International Conference on Reliability, Infocom Technologies digital video. In: 2016 World Conference on Futuristic Trends in
and Optimization (Trends and Future Directions) (ICRITO). Research and Innovation for Social Welfare (Startup Conclave).
IEEE; p. 1–5. Available from: https://​ieeex​plore.​ieee.​org/​abstr​ IEEE; p. 1–5. Available from: https://​ieeex​plore.​ieee.​org/​abstr​
act/​docum​ent/​95962​80. Accessed 12 Dec 2024 act/​docum​ent/​75839​26/. Accessed 12 Dec 2024
89. Saxena D, Cao J (2021) Generative adversarial networks (GANs) 107. Wiley (2024) Wiley. Available from: https://​onlin​elibr​ary.​wiley.​
challenges, solutions, and future directions. ACM Computing com/. Accessed 12 Dec 2024
Surveys (CSUR) 54(3):1–42 108. Wu B, Su L, Chen D, Cheng Y (2023) FPC-Net: Learning to
90. Schwarcz S, Chellappa R, Soc IC (2021). Finding Facial Forgery detect face forgery by adaptive feature fusion of patch correlation
Artifacts with Parts-Based Detectors. https://​doi.​org/​10.​1109/​ with CG-Loss. IET Comput Vision 17(3):330–340. https://​doi.​
CVPRW5​ 3098.2​ 021.0​ 0104W ​ E-​Confer​ enceP ​ rocee​ dings​ Citat​ ionI​ org/​10.​1049/​cvi2.​12169
ndex-​Scien​ce(CPCI-S) 109. Xia Z, Qiao T, Xu M, Zheng N, Xie S (2022) Towards DeepFake
91. ScienceDirect (2024) ScienceDirect. Available from: https://​ video forensics based on facial textural disparities in multi-color
www.​scien​cedir​ect.​com/. Accessed 12 Dec 2024 channels. Inf Sci 607:654–669. https://​doi.​org/​10.​1016/j.​ins.​
92. SECURITY-IJCSNS, I. J. C. S. N. (2024) International Journal 2022.​06.​003
of Computer Science and Network Security (IJCSNS). Available 110. Yi Y, Wang H, Li Q (2020) Affective Video Content Analysis
from: http://i​ jcsns.o​ rg/0​ 1_a​ im/0​ 1_a​ im_0​ 1.h​ tm. Accessed 12 Dec With Adaptive Fusion Recurrent Network. IEEE Trans Multime‑
2024 dia 22(9):2454–2466. https://d​ oi.o​ rg/1​ 0.1​ 109/T ​ MM.2​ 019.2​ 9553​
93. Sedik A, Faragallah O, El-sayed H, El-Banby G, Abd El-Samie F, 00WE-​Scien​ceCit​ation​Index​Expan​ded(SCI-​EXPAN​DED)
Khalaf A, El-Shafai W (2022) An efficient cybersecurity frame‑ 111. Yoon M, Nam S, Yu I, Ahn W, Kwon M, Lee H (2022) Frame-
work for facial video forensics detection based on multimodal rate up-conversion detection based on convolutional neural net‑
deep learning. Neural Comput Appl 34(2):1251–1268. https://​ work for learning spatiotemporal features. Forensic Sci Interna‑
doi.​org/​10.​1007/​s00521-​021-​06416-6 tional. 340. https://​doi.​org/​10.​1016/j.​forsc​iint.​2022.​111442
94. Shanableh T (2022) HEVC video encryption with high capacity 112. Yu C, Chen K, Chang C, Ti Y (2022) SegNet: a network for
message embedding by altering picture reference indices and detecting deepfake facial videos. Multimedia Syst 28(3):793–
motion vectors. IEEE Access 10:22320–22329 814. https://​doi.​org/​10.​1007/​s00530-​021-​00876-5
95. Shelke NA, Kasana SS (2021) A comprehensive survey on pas‑
sive techniques for digital video forgery detection. Multimedia Publisher's Note Springer Nature remains neutral with regard to
Tools and Applications 80:6247–6310 jurisdictional claims in published maps and institutional affiliations.
96. Singh RD, Aggarwal N (2018) Video content authentication tech‑
niques: a comprehensive survey. Multimedia Syst 24:211–240 Springer Nature or its licensor (e.g. a society or other partner) holds
97. Singla N, Singh J, Nagpal S (2023) Raven finch optimized deep exclusive rights to this article under a publishing agreement with the
convolutional neural network model for intra-frame video forgery author(s) or other rightsholder(s); author self-archiving of the accepted
detection. Concurrency and computation-practice & experience, manuscript version of this article is solely governed by the terms of
35(3). https://​doi.​org/​10.​1002/​cpe.​7516 such publishing agreement and applicable law.
98. SpringerLink. (2024). SpringerLink. Available from: https://l​ ink.​
sprin​ger.​com/. Accessed 12 Dec 2024
99. Tariq S, Lee S, Woo S, ACM (2021) One Detector to Rule Them
All Towards a General Deepfake Attack Detection Framework.
https://d​ oi.o​ rg/1​ 0.1​ 145/3​ 44238​ 1.3​ 44980​ 9 WE - Conference Pro‑
ceedings Citation Index - Science (CPCI-S)
100. Technology KU (2024) Kaunas University of Technology. Avail‑
able from: https://​www.​itc.​ktu.​lt/​index.​php/​ITC. Accessed 12
Dec 2024
 239 Page 30 of 30 A. A.-M. Alrawahneh et al.

AYAT ABD‑MUTI ALRAWAH‑ Information Science and Technology at Universiti Kebangsaan Malay‑
NEH received her first degree sia after serving as the Chairperson of the Centre for Cyber Security
in Computer Engineering from from 2017 to 2020. Pattern recognition, cybersecurity, and computer
the University of Jordan in 2011. vision are the main areas of her study.
She furthered her master’s study
in artificial intelligence at Uni‑ NAZHATUL HAFIZAH KAMARU‑
versiti Sultan Zainal Abidin DIN currently holds the position
(UniSZA), Malaysia, completing of Senior Lecturer at the Centre
it in 2022. She is currently for Cybersecurity, Faculty of
a research assistant at Universiti Information Science and Tech‑
Kebangsaan Malaysia, contribut‑ nology, Universiti Kebangsaan
ing to deep learning and com‑ Malaysia. She earned her Bach‑
puter vision projects. Ayat has elor of Engineering in Electrical
been actively conduct‑ Engineering and her Master
ing national and international of Engineering in Wireless Secu‑
training sessions in IT and pro‑ rity from Stevens Institute of
gramming, delivering impactful educational courses to diverse audi‑ Technology in New Jersey,
ences. She has served in various capacities, including quality assurance United States of America. In
engineer, IT trainer, and student affairs administrator, showcasing her 2019, she successfully
ability to bridge technical expertise with practical application. Her obtained her PhD in Electrical
research interests include computer vision using AI-driven solutions. Engineering with a specializa‑
She has published in international journals and conferences, contribut‑ tion in Security and Cryptography from UiTM Shah Alam, Malaysia.
ing to the advancement of AI applications in education and Her research interests encompass a diverse range, including authentica‑
technology. tion, network security, and artificial intelligence. Prior to her current
role, she served as an Assistant Professor at UCSI University and as a
Sharifah Nurul Asyikin Syed lecturer at Infrastructure University of Kuala Lumpur (IUKL).
Abdullah is a senior analyst in Throughout her career, she has actively participated in delivering talks
digital forensics at CyberSecu‑ and seminars. She also has contributed to the academic community by
rity Malaysia. She holds a bach‑ publishing articles in international conferences and high-indexed
elor’s degree from the National journals.
University of Malaysia (UKM)
and a Master of Business Admin‑ Sarah Khadijah Taylor is a digital
istration from the University of forensics specialist with more
Malaya (UM). In her role, than 15 years of experience in
she specializes in analyzing digi‑ the field. She is currently work‑
tal evidence, including video and ing for CyberSecurity Malaysia.
image forensics. She also leads She has authored several digital
the Vehicle Forensics Project, forensics guidelines, includ‑
which supports investigations by ing the ISO 27037, INTERPOL
preserving and analyzing data Global Guidelines for Digital
from vehicles at crime scenes. Forensics Laboratories, United
As a subject matter expert, she has presented findings on digital evi‑ Nations Digital Evidence Best
dence in court as an expert witness, contributing to the resolution Practice Guide and INTERPOL
of complex cases. Her education and expertise have positioned her as Guidelines for Seizing Virtual
a key professional in the field of digital forensics. Assets. She has successfully
led several digital forensics labs
into achieving ISO 17025:2017
SITI NORUL HUDA SHEIKH ABDUL‑
accreditation, including Oman, Brunei and Uzbekistan. Apart from
LAH graduated from the Univer‑
that, she is a digital forensics consultant for United Nations Office of
sity of Manchester Institute of
Drugs and Crime (UNODC) and adjunct lecturer for master’s program
Science and Technology in the
in several local universities.
United Kingdom with a first
degree in computing. She contin‑
ued her master degree in artifi‑
cial intelligence at Universiti
Kebangsaan Malaysia. She
later pursued a Ph.D. in Com‑
puter Vision at Universiti
Teknologi Malaysia under Fac‑
ulty of Electrical Engineering.
As Chairperson of the Cyber
Security Academia Malaysia,
she participated in both domestic
and international initiatives such as UNoDC,FIRA, Asian Foundation,
IMWU, DFRS, MIAMI, MACE, and IDB Alumni. She was elected as
the Deputy Dean (Research and Innovation) of the Faculty of