A Systematic Review on Deep Fake Image Generation,

Detection Techniques, Ethical Implications, and

Overcoming Challenges
Weeam Khimi
Effat University, Kingdom of Saudi Arabia
[email protected]

Kholood Albarqi and Kendah Saif

King Abdulaziz University, Kingdom of Saudi Arabia

Salma Elhag
Associate Professor, King Abdulaziz University, Kingdom of Saudi Arabia

Abstract
Deepfake technology, rooted in sophisticated machine learning techniques, utilizes
deep neural networks to create highly realistic fake content such as videos, audio
recordings, and images. This technology has rapidly evolved due to advancements
in deep learning models, computational power, and data availability. The ethical
implications, social impact, misuse, and legal frameworks surrounding Deepfake
technology have been extensively studied. Detection techniques using deep learning
approaches have been developed to combat the challenges posed by Deepfake
content. Recommendations for future research include enhancing detection
techniques, integrating explainable AI, and exploring real-time detection systems.
Industry and policy implications emphasize the need for robust detection
technologies, comprehensive legal frameworks, and collaborative efforts to address
ethical concerns and regulate Deepfake content.

37
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
This systematic review explores the landscape of deep fake image generation,
detection techniques and challenges, in addition to ethical considerations. By
synthesizing existing research, we aim to provide insights into deep-fake
technology's advancements, limitations, and societal implications. This review
underscores the urgent need for interdisciplinary collaboration and robust
frameworks to address the multifaceted issues surrounding deep fakes in the digital
age.
Keywords: Images Manipulation, Deepfake detection, Generative models, Ethical
implications, Misuse, Legal frameworks, Industry implications.
1. Introduction
Deep fake technology represents a significant advancement in the field of artificial
intelligence, specifically in the realm of synthetic media. This technology involves
the use of deep learning algorithms to create highly realistic and often deceptive fake
content, such as videos, audio recordings, or images. The term "deep fake" is derived
from "deep learning" and "fake," highlighting its roots in sophisticated machine
learning techniques.
- Definition and Overview of Deep Fake Technology
Deep fake technology relies on deep neural networks, a subset of machine learning
algorithms inspired by the human brain's neural architecture. These networks are
trained on vast amounts of data to understand patterns and nuances, allowing them
to generate content that mimics the appearance and behavior of real media. The term
is commonly associated with the manipulation of video and audio content, where
individuals or objects can be convincingly superimposed or synthesized within
existing footage. This technology has evolved rapidly, driven by advancements in
deep learning models, increased computational power, and the availability of large

38
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
datasets. Generative models, particularly Generative Adversarial Networks (GANs)
and Variational Autoencoders (VAEs), play a crucial role in creating realistic deep
fake content by generating new data that is indistinguishable from authentic media.
- Significance and Pervasiveness in Various Domains
The significance of deep fake technology extends across various domains, raising
both opportunities and concerns. In the entertainment industry, deep fake
applications have been used for digital doubles of actors, enabling filmmakers to
recreate scenes seamlessly or rejuvenate actors for roles. However, the widespread
use of deep fakes also poses challenges to the integrity of visual media, as the lines
between reality and manipulation become increasingly blurred.
Beyond entertainment, deep fakes have implications in politics, journalism, and
cybersecurity. Politicians and public figures may find themselves targets of
manipulated content, leading to misinformation and potential damage to reputations.
Journalistic integrity is threatened as deep fake technology can be exploited to create
fabricated news stories or alter the context of real events. Additionally, the use of
deep fakes in cybersecurity poses risks, as attackers can manipulate audio or video
to impersonate individuals for malicious purposes. As deep fake technology
continues to advance, society must grapple with the ethical, legal, and societal
implications of its use. Striking a balance between leveraging its positive applications
and mitigating the potential for misuse is crucial for navigating the evolving
landscape of synthetic media.
- Research questions
1- What are the ethical implications of Deepfake?
2- What are used and proposed techniques to detect Deepfake?
3- What are the challenges of Deepfake detection development?
39
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
In order to answer previous questions, we conducted a systematic review
methodology in which we collected sixty papers related to our subject based on
specific criteria, then we synthesized every paper by extracting useful information
and categorizing according to our main questions, finally, we analyzed our findings
to give final answers followed by visualization for easier understanding.
The structure of this paper is as follows: Section one presents the abstract for an
overview about the study. Section two presents an introduction to our topic. Section
three presents the methodology including steps and summary of scientific research
papers. Section four presents. Section five presents analysis and results of our
synthesis. Section six presents conclusions and discussion including gaps and future
research. Section seven presents other information related to the conduction of this
study. Finally, section seven presents all the used references.
2. Methods
In our systematic review, we reviewed many papers, then narrowed it down to sixty
papers. During our synthesis, we ensured that the selected papers met our standards.
We filtered the date to get only recent papers between 2019 and 2024, as the technical
field is rapidly changing; therefore, we need to cope with the most recent changes,
so we excluded papers that were conducted before 2019, as well as papers that did
not clearly mention the date of publication. Moreover, selected papers have been
chosen according to their connectivity to our subject. As we have three main
objectives, we ensured that every single paper could benefit us in one or more
objectives, and we excluded all papers that showed weak connections. Papers were
selected from the Google Scholar engine due to the wide variety provided and the
filtration features that ease the process. Finally, we only selected papers that are
written in English. As English is considered the dominant and standardized language
40
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
in the academic world, we excluded papers that were originally written in any other
language, even if they were translated. Table 1 summarizes our inclusion and
exclusion criteria, as shown:
Table 1. List of Inclusion and Exclusion Criteria

For our systematic review we depended on academic journals collected through

search engines, which represent a combination of primary and secondary sources.
Primary sources helped us highlight latest discoveries, making strong evidence for
the purpose of these papers, and providing authoritative information. While
secondary sources helped us to gain a more comprehensive perception about the
subject and related aspects.
In a review of multiple papers examining ethical implications and effect on media
trust that come along with the creation of Deepfake, tools used to generate Deepfake,
methods investigated to detect Deepfake, and challenges associated with the
development of Deepfake detection we used several terms to extract most related
papers. We illustrated these terms using Boolean operators for clarification, in which
OR connects similar terms while AND connects the flow of different terms, detailed
as follows:
(Images* OR Pictures* OR frames* OR Videos) AND (Detection* OR Reveal* OR
Discover* OR Find) AND (Fake* OR Manipulated* OR Forgery) AND (Technique*
41
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
OR Approach* OR Method) AND (Concerns* OR Issues* OR Implications) AND
(Improve* OR Enhance* OR Mitigate) AND (Prevent* OR Protect* OR Preserve)
AND (Attack* OR Violate* OR Abuse) AND (Ethics* OR Morals) AND
(Challenges* OR Limitations* OR Obstacles) AND (Authenticity* OR Originality*
OR Genuineness) In our pursuit to address the core research inquiries, we
meticulously sifted through fifty papers, each dedicated to unraveling Deepfake
ethical implications, detection techniques and challenges. These selected studies
provide in-depth understanding of the Deepfake field, including implications,
detection techniques and challenges. To comprehensively structure our findings, we
organized the amassed information into distinct categories, namely objective,
methodology, findings, and limitations or future areas of exploration. This strategic
categorization facilitated a thorough investigation into our primary research queries,
ensuring a comprehensive and detailed understanding of the landscape concerning
Deepfake technology.
- Comprehensive Overview: Ethical Implications, Detection Techniques
and Challenges
(1) Ethical Implications and Social Impact of Deep Fake Image Generation
A. Ethical Considerations in Deep Fake Technology
The collection of studies and research on Deepfake technology reveals a multitude
of ethical concerns regarding its potential misuse and harm. Themes such as
deception, manipulation, and the erosion of authenticity in media content recur
throughout the literature [1], [2], [3], [5], [6], [8], [12], [13]. Ethical considerations,
including consent, privacy, and the impact on public trust, emerge as central issues
across various contexts [1], [2], [3], [6], [7], [8], [12], [13], [14], [15], [16], [17].
Notably, non-consensual use, particularly in areas like pornography, raises
significant concerns about privacy invasion and reputational harm [1], [3], [4], [6],
42
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
[9], [10], [15]. Furthermore, scholars stress the need to address broader ethical
implications, including those related to cybercrimes, terrorism, and political
manipulation [3], [6], [9], [15]. The literature highlights the importance of ethical
guidelines, regulatory frameworks, and interdisciplinary collaboration to mitigate
risks, protect individuals, and ensure responsible use of Deepfake technology [3],
[6], [11], [12], [54]. Challenges such as attribution techniques and content detection
underscore the complexity of addressing these ethical concerns [11], [12]. In essence,
the discourse surrounding Deepfakes emphasizes the urgent need for comprehensive
strategies to navigate their ethical dimensions and safeguard societal well-being [52],
[53]. Deepfake technology raises ethical concerns regarding the manipulation of
audio and visual content. The creation and dissemination of Deepfakes without
consent can lead to deception and harm. Ethical guidelines and regulations are
essential to address the potential misuse of Deepfake technology and protect
individuals and society from its negative impacts.
B. Social Impact on Individuals and Society
The array of studies and research provided delve into the profound social
implications of Deepfake technology, highlighting its potential to disrupt trust,
manipulate perceptions, and spread misinformation [1], [2], [3], [4], [5], [6], [7], [8],
[9], [10], [11], [12], [13], [14], [15], [16], [17][52],[53],[54]. Across these studies, a
recurring theme emerges regarding the erosion of trust in media, institutions, and
public figures due to the proliferation of Deepfakes. Moreover, the impact extends
beyond traditional media as Deepfakes infiltrate social interactions and interpersonal
relationships, affecting self-perception and memories. The misuse of Deepfake
technology, including its role in spreading misinformation, inciting violence, and
perpetuating non-consensual pornography, poses significant risks to individuals and
society. Victims of Deepfake pornography, in particular, experience profound
psychological and professional consequences, exacerbating gendered disparities in
43
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
online spaces. Additionally, the continuous exposure to Deepfakes can instill
skepticism and confusion, challenging societal trust and mental well-being. It is
evident that addressing the social impact of Deepfake technology requires
interdisciplinary efforts, awareness campaigns, and educational initiatives to foster
critical thinking and mitigate harm. Deepfake technology's social impact includes
spreading fake news, manipulating public opinion, and undermining trust in media,
necessitating a thorough understanding to address its negative consequences [52],
[53], [Social Impact on Individuals and Society]. Deepfake technology can have a
significant social impact by influencing public perception, spreading
misinformation, and undermining trust. Individuals and society may face challenges
in distinguishing between real and fake content, leading to confusion and potential
harm. The proliferation of deep fakes can disrupt social dynamics, impact
relationships, and have broader implications on media consumption and
communication. Deepfake technology has the potential to manipulate information,
influencing public opinion and eroding trust in media and online content. This
manipulation can lead to misinformation and deception, impacting individuals and
society at large. Researchers and policymakers are actively working to develop
strategies and tools to protect against the harmful effects of Deepfakes, aiming to
safeguard the integrity of information and public discourse. Being targeted by
Deepfakes can lead to loss of trust, credibility, and potential harm to personal and
professional lives. The social impact of Deepfakes underscores the need for
awareness and ethical considerations in their creation and use [54].
C. Misuse and Malicious Applications of Deepfakes
The compilation of studies and research underscores the wide-ranging potential for
malicious misuse of Deepfake technology across various domains, from politics to
personal relationships. Malicious actors exploit Deepfakes for purposes such as
political manipulation, fraud, harassment, and spreading misinformation, posing
44
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
significant threats to national security, public safety, and individual well-being [1],
[2], [3], [4], [5], [6], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16],
[17],[52],[53],[54]. Victims of malicious Deepfakes experience reputational damage,
emotional distress, and personal harm, with particular vulnerabilities highlighted in
cases such as non-consensual sexual imagery. The ease of access to Deepfake
creation tools exacerbates these risks, amplifying concerns about the widespread
dissemination and impact of malicious content. Privacy infringements, defamation,
and psychological harm are among the grave consequences associated with the
misuse of Deepfake technology. To combat these negative ramifications, a multi-
pronged approach encompassing awareness campaigns, detection tools, regulatory
frameworks, and technological solutions is imperative to mitigate the harmful effects
of malicious Deepfakes and safeguard societal well-being. Deepfakes can be
maliciously used to create false and damaging content, leading to defamation and
misinformation. Addressing this misuse is crucial to prevent harm and protect the
integrity of information and media [52], [53], [54]. Deep fakes can be misused for
spreading disinformation, manipulating public opinion, and creating fake news.
Individuals and organizations may use deep fake technology to deceive, defame, or
manipulate others for personal or political gain. The potential harm, discord, and
erosion of trust in media and information are significant concerns. Deepfakes can be
misused for spreading fake news, creating revenge pornography, and manipulating
images or videos for malicious purposes. The ability to deceive through fake content
poses a significant threat, highlighting the importance of addressing the negative
implications of Deepfake technologies. Establishing comprehensive legal
frameworks and ethical guidelines is crucial to prevent misuse and protect
individuals and society from the harmful effects of Deepfakes [52], [53], Misuse and
Malicious Applications of Deepfakes: Deepfakes can be used for spreading
misinformation, defamation, and manipulation. Malicious actors may use Deepfakes

45
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
to deceive, intimidate, or blackmail individuals. The misuse of Deepfakes poses
significant threats to privacy, security, and the integrity of information in various
contexts. The misuse of Deepfake technology poses significant risks, including
identity theft, the spread of misinformation, and reputational damage. Malicious
actors can exploit Deepfakes to create deceptive content that can be used for
fraudulent purposes or to manipulate public perception. Detecting and countering
Deepfakes is essential to prevent their misuse and protect individuals and
organizations from potential harm [3], [6].
D. Legal and Regulatory Frameworks Addressing Deepfake Concerns
The collection of studies and research underscores the pressing need for legal and
regulatory frameworks to address the multifaceted challenges posed by Deepfake
technology [1], [2], [3], [4], [5], [6], [7], [8], [9], [10], [11], [12], [13], [14], [15],
[16], [17],[52],[53],[54]. These frameworks are essential in combating the malicious
creation and dissemination of Deepfakes, holding accountable those responsible for
their production, and safeguarding individuals and society from the detrimental
effects they can inflict. Efforts toward establishing comprehensive legal measures
encompass criminalizing the creation and distribution of Deepfakes without consent,
imposing penalties on offenders, and developing guidelines for their detection and
mitigation [1], [3], [4], [6], [7], [8], [11], [12], [13], [15], [16], [17]. Collaboration
among governments, regulatory bodies, technology experts, and law enforcement
agencies is imperative to develop effective strategies and ensure international
cooperation in combating the multifarious threats posed by Deepfakes [1], [2], [7],
[16]. Furthermore, the scope of these efforts extends beyond mere prevention,
encompassing the protection of individuals' fundamental rights to privacy, consent,
and digital self-representation [2], [8], [10], [15]. Legislative endeavors also
emphasize the importance of transparency, ethical use of generative modeling, and
the implementation of robust detection tools to mitigate the deleterious impacts of
46
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
Deepfake technology [10], [11], [13], [14], [15], [16]. As the legal and regulatory
landscape surrounding Deepfake technology continues to evolve, the development
of comprehensive and adaptive regulations remains paramount to effectively
distinguish between authentic and manipulated content, uphold individuals' rights,
and ensure responsible use across diverse sectors [17]. Legal frameworks are crucial
in addressing Deepfake concerns by establishing guidelines for their creation,
distribution, and detection, thereby mitigating negative impacts and safeguarding
individuals and society [52], [53], [54]. These frameworks aim to provide clear
regulations and penalties to deter the misuse of Deepfakes while protecting
individuals' rights. However, the legal framework for Deepfakes currently lags
behind the rapid advancement of the technology, creating challenges in addressing
concerns related to manipulation and misuse. Developing a comprehensive legal
framework is crucial to enable Deepfake recognition software to effectively
distinguish between fake and authentic content. Alongside legal regulations, ethical
considerations are also essential to ensure the responsible use of Deepfake
technologies and mitigate potential risks associated with their misuse [54]. As
Deepfake technology advances, legal and regulatory frameworks are evolving to
address the challenges posed by its misuse. Issues related to authentication,
verification, and the admissibility of Deepfake evidence in legal proceedings are
being actively considered. Initiatives and collaborations between organizations,
technology companies, and policymakers aim to develop effective detection methods
and establish guidelines within legal boundaries to combat the negative impacts of
Deepfakes [3]. Developing legal and regulatory frameworks is essential to address
the challenges posed by Deepfake technology. Regulations can help deter malicious
use, protect individuals from harm, and hold accountable those responsible for
creating and disseminating Deepfakes. Establishing clear guidelines and

47
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
consequences for Deepfake misuse can contribute to mitigating the negative impact
on individuals and society [6].
(2) Deep Learning Approaches for Deep Fake Detection
A. Current and proposed techniques
Study [22] proposed the FST-Matching Deepfake detection model which aims to
identify highly-compressed altered videos by separating irrelevant features and
disentangling source/target-irrelevant representations from visual concepts. The
model achieved success in detecting manipulated compressed videos with a high
accuracy and AUC values. Similarly, study [30] has also introduced the FST-
Matching Deepfake Detection Model to enhance detection performance by analyzing
artifact-relevant features and image matching. The model improved forgery
detection on compressed videos, showcasing an average AUC of 97.0% across
various scenarios, which is considered a high rate. Another study [57] introduced
two complementary face recognition networks to obtain identity cues for the face
and its context by leveraging deep neural networks for face identification. These
networks are designed to focus on specific facial regions, specifically the segmented
face and its surrounding context. Results show that the proposed method achieves
the best AUC scores on all benchmarks and exhibits improved generalization
abilities compared to baseline methods with an AUC of 0.7555 for FSGAN
manipulation method and 0.9262 for 3DMM-based swap. One study [23] conducted
an experiment on preserving picture authenticity through the use of watermarking.
The process involved two steps: first, embedding watermarks into facial features
using a neural network with an encoder and decoder to make images face-swap
sensitive, and second, verifying the watermark's presence to determine authenticity
of the image. This method showed effectiveness with an average detection accuracy
exceeding 80%. Another study [29] introduced the FaceGuard framework, utilizing

48
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
deep-learning-based watermarking to embed semi-fragile watermarks into real face
images. FaceGuard was trained then tested on datasets containing real and Deepfake
images, demonstrating high effectiveness in detecting Deepfakes with an accuracy
rate of 99.5% overall by embedding watermarks into real face images proactively.
Study [25] introduced a detection method called Disruptive Technique (DDPM) to
address data poisoning in training data by implementing purification steps using a
denoising diffusion probabilistic model. Experimental validation showed an
enhanced detection accuracy and robustness in identifying Deepfakes even when the
dataset is poisoned, achieving accuracy rates ranging from 11.24% to 45.72%
compared to traditional methods, even in scenarios that have 100% poisoned data.
Moreover, study [40] presented NoiseScope, a blind detection method for Deepfakes
that does not require prior knowledge of generative models or access to fake images.
NoiseScope utilizes deep neural networks to analyze unique patterns from generative
models when creating fake images. It achieved over 90% F1 score in detecting fake
images across diverse datasets and generative models. NoiseScope demonstrated
effectiveness in various scenarios, including compressed images and different post-
processing techniques, showcasing resilience against countermeasures like
fingerprint spoofing attacks and attempts to evade detection by adapting GAN
models.
Study [34] introduced Pair-wise Self-Consistency Learning (PCL) to extract source
features from Deepfake images by assessing consistency within image patches. The
study also introduced the Inconsistency Image Generator (I2G) for generating forged
images with annotated manipulated regions, achieving high performance with AUC
scores of 99.11% to 99.98%. Also, study [36] presented the Common Fake Feature
Network (CFFN), a framework for detecting fake faces and general images generated
by GANs. CFFN utilized pairwise learning as well, cross-layer feature
representations, and contrastive loss to capture discriminative features across
49
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
different GANs. Experimental evaluations on CelebA and ILSVRC12 datasets
demonstrated CFFN's superior performance in fake face and general image detection,
with precision and recall rates of 0.936 to 0.930.
Study [41] introduces a CNN-based image forgery detection system that focuses on
variations in image compression to detect different types of image forgeries. Trained
on the CASIA 2.0 database, the model achieves high accuracy in distinguishing
genuine from tampered images, outperforming existing methods with an accuracy of
92.23% on test images. Another study [45] presents a method for detecting Deepfake
images using DL techniques, combining Error Level Analysis and CNNs, by
utilizing pre-trained CNN models like Alex Net and Shuffle Net, the study achieves
accuracies of 86.1% to 88.2% with SVM and KNN classifiers, demonstrating the
efficacy of DL in efficiently detecting Deepfakes. Study [48] introduces an
evolutionary learning algorithm for automatic creation of CNN architectures for
Deepfake detection. By utilizing genetic algorithms to generate diverse CNN
structures and optimizing parameters, the model achieves high accuracies of 98.45%
to 99.75% on different datasets, surpassing existing architectures and demonstrating
effectiveness in detecting manipulations without extensive preprocessing. Also,
study [50] presents a method for detecting Deepfake-forged content using a
Separable CNN and image segmentation techniques. Trained on FaceForensics++
and tested on DeepFaceLab and StyleGAN images, the model achieves high
accuracy and AUC values, outperforming state-of-the-art methods, and
demonstrating effectiveness in detecting Deepfake content.
Study [24] proposed two methods, DAG-FDD and DAW-FDD, for Deepfake
detection to enhance fairness. DAG-FDD targets Deepfakes without demographic
annotations, while DAW-FDD focuses on considering annotations, using conditional
Value-at-Risk (CVaR) and demographic factors like ethnicity and age to reduce bias
and disparity between groups. Another study [27] introduced the Deepfake Disrupter
50
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
algorithm to detect and disrupt fake images by adding imperceptible perturbations,
showing improvements in disrupting Deepfake methods compared to baseline
techniques, increasing F1-score by 10% to 20% and achieving high success rates in
detecting real and perturbed inputs. Study [37] suggested a deep learning-based
approach for Deepfake detection using transfer learning techniques like Xception,
NAS-Net, Mobile Net, and VGG16, achieving a high accuracy rate of 94% by
analyzing facial attributes to identify anomalies indicative of Deepfake
manipulation. And study [38] investigated the use of Discrete Cosine Transform
(DCT) in Deepfake image recognition using GANs, showing that DCT-transformed
images were linearly separable, enabling a simple linear classifier to achieve 100%
accuracy and outperform classifiers trained on raw pixels, with greater resistance to
image perturbations except for noise
Study [42] introduces the DeepfakeStack model for detecting manipulated videos
and images using deep ensemble learning techniques. By combining multiple base-
learners pretrained on ImageNet weights, the model achieves significant
performance improvement, with precision, recall, and F1-score values close to 1.0.
The DeepfakeStackClassifier (DFC) model attains high accuracy and AUROC
scores, surpassing individual deep learning models in detecting Deepfakes. This
approach demonstrates effectiveness in accurately identifying manipulated
multimedia content, providing a robust basis for real-time Deepfake detection
systems. Another study [43] presents a novel multi-attentional Deepfake detection
framework that efficiently captures local features crucial for distinguishing between
real and fake faces. The model integrates an Attention Module, texture enhancement
block, and Bilinear Attention Pooling to address subtle discrepancies in Deepfake
videos. Evaluation results show high accuracy on FaceForensics++ and Celeb-DF
datasets, highlighting the framework's robustness and accuracy in detecting
Deepfake content across diverse datasets. Study [44] introduces a method for
51
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
detecting Deepfake images generated by various GANs using the Expectation
Maximization (EM) algorithm. By extracting features from images produced by
GAN architectures and employing classification with K-NN, SVM, and LDA
classifiers, the model achieves impressive accuracy rates in differentiating between
authentic and generated images. It effectively distinguishes Deepfakes from different
GAN architectures based on their convolutional traces, showcasing its success in
detecting manipulated content.
The study [46] introduces a Deepfake detection method that utilizes the Vision
Transformer (ViT) model, fine-tuned on a balanced dataset of real and synthetic
images. The ViT model, with patch embedding and self-attention mechanisms,
achieves exceptional accuracy rates, reaching 100% in identifying synthetic images
from StyleGAN and Snapchat filters. When tested on a combined dataset of real and
synthetic images, the model consistently achieves high accuracy rates ranging from
99.66% to 100%, showcasing its robustness in Deepfake detection. In contrast, the
study [47] presents ADD, an attention-based digital video authentication system for
detecting Deepfakes. ADD focuses on facial regions in videos, extracting
discriminative features using attention maps and data augmentation techniques.
Evaluated on challenging datasets like Celeb-DF (V2) and WildDeepfake, ADD
significantly enhances detection accuracy rates compared to baseline models. For
instance, on Celeb-DF (V2), the ResNet architecture achieves a 98.37% detection
accuracy rate with ADD, representing a substantial improvement. Similarly, on
WildDeepfake, the Exception architecture attains an 80.13% accuracy rate with
ADD, surpassing previous state-of-the-art methods. While in study [56] ResNet-
Swish-Dense54 model was designed to effectively capture and analyze visual
features in images, results proved improved detection accuracy and robustness
against adversarial attacks across different datasets and manipulation types with an
accuracy of 99.26%.
52
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
Furthermore, the Enhanced Model for Fake Image Detection (EMFID) proposed in
the study [49] offers a comprehensive approach for identifying forged digital images.
EMFID integrates image pre-processing, histogram-based and Discrete Wavelet
Transform (DWT)-based feature extraction, and Convolutional Neural Network
(CNN) classification. The model achieves impressive performance metrics on
benchmark datasets like CASIA v1.0 and CelebA, demonstrating high sensitivity,
specificity, precision, accuracy, and F-Measure. These results highlighted the
robustness and efficacy of EMFID in distinguishing manipulated images from
authentic ones. Study [51] introduces CD-Net, a novel framework for face forgery
detection, utilizing deep learning techniques. CD-Net comprises two main
components: DICM and IDM. DICM captures communal features across multiple
frames, enhancing stability in detecting forgery patterns, while IDM adaptively
adjusts discriminative centers based on individual instance features, improving
detection accuracy. Extensive experiments on datasets like FF++, DFDC, and Celeb-
DF v2 to show the effectiveness of CD-Net, achieving an AUC of 0.952 on FF++
and significant improvements in detection performance and stability metrics with
DICM and IDM. While study [55] investigated and compared the effectiveness of
handcrafted features (SIFT, HoG) and deep features (Xception, CNN+RNN) in
detecting Deepfake videos. It was found that handcrafted features performance is
poor and may not be suitable for detecting Deepfakes due to their limitations in
capturing facial details. On the other side, deep learning methods showed high
performance in Deepfake detection, especially Xception which achieved nearly
perfect results with an accuracy of 98.77% in original Deepfake test. As shown in
table 2, we gathered all detection techniques and algorithms among the synthesized
studies.

53
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
 Table 2. Deepfake Detection Techniques for Each Study

B. Evaluation Metrics for Deepfake Detection

The evaluation of various studies on Deepfake detection techniques involves a
comprehensive range of metrics to assess performance and reliability of detection

54
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
models. As shown in table 3, we gathered all evaluation metrics used among the
studies.
Table 3. Evaluation Metrics for Each Study

55
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
 C. Benchmark Datasets for Deep Fake Detection
Several studies have contributed to the advancement of detecting fake images and
videos using machine learning tools and innovative techniques. As shown in table 4,
we gathered all used datasets among studies.
Table 4. Datasets for Each Study

56
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
 (3) Challenges and Limitations of Deep Fake Detection Techniques
A. Machine Learning Generated Threats
A study [18] highlighted machine learning-generated threats, including Adversarial
Perturbation Attacks, in which imperceptible perturbations are added to manipulate
input data. Frequency Domain Manipulation involves eliminating manipulation
traces in the frequency domain of Deepfake content to bypass detection methods.
Image Filtering Techniques apply filters or transformations to fake images.
Additionally, concerns about the robustness and security of machine learning
systems have been raised, including vulnerabilities to adversarial attacks or
manipulation, as mentioned in Study [60]. These challenges underscore the
importance of addressing ethical, technical, and societal considerations in the
development and deployment of machine learning technologies to mitigate potential
risks and ensure responsible use.
Another study introduced Fakepolisher [19], aiming to enhance Deepfake
evasiveness through shallow reconstruction techniques, altering pixel values and
color gradients to confuse detection mechanisms. Black-box attacks by Denoising
Diffusion Models (DDMs) [20] deceive detection systems through conditional image
synthesis and guided post-processing. Adversarial attacks [21] manipulate Deepfake
content to evade detection by exploiting vulnerabilities in detection models. Various
attack scenarios [26] include White-Box Attacks, where adversaries with complete
model knowledge craft imperceptible modifications to bypass detectors.
Transferable Attacks allow adversarial examples to deceive different detection
models. Universal Adversarial Perturbations are highly transferable perturbations
added to any input data. A trace removal attack method, TR-Net [28], enhances
Deepfake spuriousness and evades detection by forensic detectors.

57
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
The impact of adversarial examples on deceiving machine learning models and
gradient-based attacks on CNN-based Deepfake detection systems are discussed
[32]. Adversarial perturbations designed to fool one model can deceive other unseen
CNN-based detection methods. Adversarial attacks [33] demonstrate the ability to
bypassing Deepfake detectors by adversarially modifying fake videos. The emerging
threat landscape of machine learning-generated content [36] focuses on fake images
produced by advanced Generative Adversarial Networks (GANs), posing challenges
in image forgery detection.
Concerns from machine learning-generated threats, particularly Deepfake images,
have escalated [38]. The ability to create convincing fake images raises
apprehensions about their misuse in spreading misinformation, identity theft, and
privacy violations. Machine Learning Generated Threats (MLGTs) [40] encompass
risks associated with generative models like GANs, facilitating various malicious
activities. Addressing these challenges is crucial to mitigate the detrimental effects
of machine learning-generated threats.
One notable challenge mentioned in Study [58] is the potential for biased or unfair
outcomes produced by machine learning algorithms. These biases can arise from the
data used to train the algorithms, leading to discriminatory or inaccurate results,
particularly in sensitive domains such as healthcare or criminal justice. Another
challenge highlighted is the issue of interpretability and explainability of machine
learning models, as discussed in Study [59]. Complex algorithms, such as deep
learning neural networks, may produce highly accurate predictions but lack
transparency in how they arrive at those decisions, making it difficult for humans to
understand or trust the outcomes.

58
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
 B. Other Practical Challenges
Study [18] illustrates that current detection techniques face limitations in requiring
extensive datasets for training, leading to resource-intensive demands and a lack of
universal applicability [19]. The sophistication of Deepfake generating models
continuously enhances the realism of fake images, making them visually convincing
and harder to detect, thereby exposing vulnerabilities in existing detection systems
[20]. Detection methods struggle with poor generalization abilities across different
Deepfake types, impacting detection performance and susceptibility to transfer
attacks [21]. While effective on familiar datasets, detection models encounter
difficulties when faced with new data containing novel Deepfakes, emphasizing the
challenge of generalizing detection systems to unseen datasets [22]. Additionally,
computational overhead and the need for further development to adapt to various
Deepfake types beyond facial features present challenges in Deepfake detection [23].
Limited datasets for training pose a challenge in ensuring the accuracy and
effectiveness of detection models, especially with the rapid evolution of Deepfake
techniques and adversarial attacks [24]. Data poisoning, affecting the integrity of
training datasets, leading to decreased accuracy in detection, emphasizing the need
to enhance generalization capabilities for real-world applicability [25]. Assessing the
robustness of adversarial examples in compressed videos and evaluating
transferability among different Deepfake detectors are key challenges in the field
[26]. The presence of subtle visual artifacts and synthesis signals in Deepfake content
complicates the differentiation between authentic and manipulated media, posing
challenges for detection processes [27]. The complexity of trace removal
mechanisms, aimed at enhancing the authenticity of manipulated content while
evading detection, presents a significant challenge in Deepfake detection [28].
Robustness to post-processing operations, such as compression and resizing, is
crucial for maintaining detection accuracy amidst various image distortions [29].
59
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
Deepfake detection models face performance drops when tested on compressed
images or videos, highlighting the challenge of maintaining accuracy under
compression and understanding artifact features learned by these models [30].
In [31], the rapid advancements in generative AI pose a challenge in keeping pace
with evolving Deepfake techniques, potentially eroding public trust in media
authenticity. Accessibility to Deepfake creation tools by a wide user base further
complicates detection efforts, while biases in training data can propagate harmful
stereotypes through generated content. Adversarial attacks, transferability across
models, neural network complexity, and temporal consistency are key challenges
outlined in [32], emphasizing the need for robust detection mechanisms capable of
adapting to diverse datasets and sophisticated manipulation techniques.
Adversarial examples overriding detection mechanisms, challenges in classifying
compressed videos, and exploiting temporal dependencies in manipulated content
are highlighted in [33], underscoring the complexity introduced by adversarial
attacks and compression artifacts. Training Pair-wise Self-Consistency Learning
(PCL) for Deepfake detection, as discussed in [34], requires detailed annotations and
diverse training data, posing computational and annotation challenges. Issues related
to dataset diversity, model robustness, generalization, and computational complexity
are addressed in [35], emphasizing the need for fine-tuning models to minimize false
positives and negatives while ensuring scalability for real-time applications.
The diverse characteristics of fake images generated by various GANs, lack of
transparency in technical details, and challenges in dataset curation are discussed in
[36], highlighting the importance of robust learning strategies for effective detection.
The complexity of Deepfake generation processes, sophistication of techniques, and
demand for accurate detection methods in digital forensics and cybersecurity are
emphasized in [37], indicating the necessity of addressing computational complexity

60
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
and dataset handling challenges. Vulnerability to adversarial examples, limitations
in mitigating artifacts, and the need for ongoing research and development to
enhance resilience are key challenges outlined in [38], emphasizing the continuous
effort that is required to improve Deepfake detection models' effectiveness and
reliability.
Study [39] faces challenges with substantial computational resources needed for
training on large datasets and adapting to various Deepfake manipulations and image
quality variations. Another study [40] focuses on enhancing NoiseScope's resilience
against countermeasures like JPEG compression and denoising attacks. Practical
challenges in Deepfake detection models [41] include optimizing performance across
diverse computational resources and addressing hardware limitations. Challenges in
a different model [42] include sensitivity to training data quality and diversity,
robustness against emerging manipulation techniques, and deployment in real-world
scenarios. Another study [43] highlights challenges in distinguishing subtle
differences between real and fake faces and variability in textural patterns.
Challenges in a different model [44] involve computational resources,
interpretability of features, generalization across Deepfake generation techniques,
and integration into forensic workflows. Common challenges discussed in a study
[45] include adapting to evolving Deepfake techniques, robustness against
adversarial attacks, computational complexities, and dataset diversity.
Practical challenges in Deepfake detection using the Vision Transformer (ViT)
architecture [46] include high computational demands and the need for expertise in
machine learning. Another study [48] faces issues with computational complexity,
training time, and generalization to unseen data. The Enhanced Model for Fake
Image Detection (EMFID) [49] grapples with computational complexity, dataset
biases, and generalization across diverse image types. Challenges in a different study
[50] involve data quality, model interpretability, and scalability in real-world
61
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
deployment. Addressing inconsistencies in forgery patterns is a challenge for a model
[51], requiring understanding variations in fake images and refining discriminative
centers.
in [55] accuracy of Deepfake detection models dropped dramatically when the
training sets and test sets did not match during the experiment. Furthermore, a study
[56] faced lack of robustness when introducing the model of unseen cases and
adversarial attacks, in addition to lack of the explainability feature, which is essential
for forensic analysts. Moreover, study [57] faces difficulties in acquiring high-
quality and diverse datasets for training, in addition to difficulties in ensuring that
detection techniques are used responsibly and do not infringe on privacy or civil
liberties.
3. Analysis & Results
According to table 5 in which we have concluded fifty-one studies, which answer
each research question, we specified all papers that were interested in ethical
implications of deep fake, used and proposed detection techniques, and challenges
related to development of detection techniques.
Table 5. Numbers of Relevant Studies for Each Research Question

62
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
 Figure 1. Main Categories Presence Among Synthesized Studies

As shown in figure 1, papers have been shown according to their relevance to the
main aspects: (1) Ethical implications of Deepfake (2) used and proposed techniques
to detect Deepfake (3) Challenges of Deepfake detection development. The figure
illustrates that Twenty studies discussed ethical implications (33.33%), while
twenty-nine studies discussed detection techniques (56.86%), and thirty-two studies
discussed detection techniques development challenges (62.74%), noting that most
papers which explored detection techniques discussed challenges and limitations as
well, indicating transparency and providing a strong foundation for those who are
willing to continue the study. While some other papers with a literature review nature
discussed challenges as well, without proposing a detection technique or conducting
any experimental validation, justifying why the last category was covered the most
within synthesized papers.

63
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
 (1) Summary of Ethical Implications
According to table 6, we derived subcategories from the main category (1) The
ethical implications of Deepfake, which included Ethical Considerations in Deepfake
Technology, Social Impact on Individuals and Society, Misuse and Malicious
Applications Deepfake, and Legal and Regulatory Frameworks Addressing
Deepfake Concerns.
Table 6. Numbers of Relevant Studies for Each Sub-Category in Ethical Implications and Social
Impact of Deepfake Image Generation

Figure 2. Ethical Implications Sub-Categories Presence Among Studies

64
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
As shown in figure 2, the twenty synthesized studies related to the first category (1)
Ethical implications of Deepfake, had four sub-categories which were covered in all
the twenty studies, which are: Legal and regulatory framework, ethical
considerations, misuse and malicious applications, and social impact on individuals
and society. Indicating that any study discussing ethical implications necessarily
covered all sub-subjects, and providing us insights into the interconnection of these
sub-subjects.

Figure 3. Summary of Ethical Implications and Social Impact of Deepfake

As shown in figure 3 below which summarizes all ethical implications, Deepfake

technology presents ethical challenges, including concerns about consent, privacy,
and the potential harm to individuals and society. It raises issues such as identity
theft, non-consensual use for explicit content, and the erosion of public trust. The
social impact involves the spread of misinformation, potential harm to personal and
professional lives, and challenges in distinguishing real from fake content.
Deepfakes can be misused for malicious purposes, posing threats to national security
and individual well-being. Addressing these concerns requires legal and regulatory

65
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
frameworks to criminalize malicious creation, hold creators accountable, and
encourage international cooperation.
(2) Summary of detection approaches
A. Current and suggested detection techniques
Synthesis of the studies [34]-[60] illustrated a variety of detection techniques, some
of them are currently applicable while the others are newly proposed as
enhancements in this field, including forty-five different detection techniques and
algorithms as shown in figure 4.

Figure 4. Detection Techniques Repetitiveness Among Studies

As shown in figure 4, results of our analysis demonstrate the percentage of
repetitiveness of Detection Techniques Among Studies, it shows that the most used
technique within deepfake detection field is Convolutional Neural Networks (CNN),
while the least used techniques are DC-GAN, Convolutional Traces and Expectation-
66
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
Maximization, RESNET-18, EfficientNet Architecture, DAG-FDD, DAW-FDD,
DDPM, EfficientNet B7, EfficientNet B3, 3D CNN, Deepfake Disrupter, EN-B7
Selim, XN WM Team WM, EN-B3 WM Team WM, Sequence-Based Model,
VGGNet, CFFN, Mobile Net, NAS-Net, VGG16, NoiseScope, DeepfakeStack,
Expectation Maximization, Error Level Analysis, Vision Transformer, ADD
framework, EMFID, Discrete Wavelet Transform, CD-Net framework, CVM
classification, Data Augmentation, PCA, unsupervised contrastive learning, ResNet-
Swish-Dense54, and Face Recognition Model.
This analysis provides valuable insights into the prevailing detection techniques
employed within the Deepfake research domain. It becomes evident that certain
techniques are favored over others, largely due to their demonstrated effectiveness
and applicability in identifying and mitigating the proliferation of Deepfake content.
This preference may stem from various factors, including the robustness of the
technique, its scalability to different types of Deepfake media, and its ability to adapt
to evolving manipulation methods. By understanding the predominant use of specific
detection techniques, researchers and practitioners can gain a clearer understanding
of the current landscape and focus their efforts on refining and advancing these
methods to address emerging challenges in Deepfake detection.
B. Summary of Evaluation Metrics
Synthesis of the studies [34]-[60] illustrated a significant use of accuracy measures
to evaluate results, including forty-one different measurement metrics as shown in
figure 5.

67
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
 Figure 5. Evaluation Metrics Repetitiveness Among the Studies
As shown in figure 5, results of our analysis demonstrate the percentage of
repetitiveness of evaluation metrics, it shows that the most used metrics within
Deepfake detection field is Accuracy (ACC), while the least used metrics are Cross-
Validation, Perceptual Image Patch Similarity (PIPS), Cosine Similarity Index
Measure (CSIM), Stability Analysis, Metric Q, Transferability Rate, L∞, H.264
Quantization Factor, Average Precision, Geometric Mean, Sensitivity, Mean
Absolute Error (MAE), Root Mean Square Error (RMSE), K-NN, Support Vector
Machine (SVM), LDA, T-SNE, Proportion of Unstable Predictions (PUP), and
Correction Rate (CR).
Moreover, it can be noticed also that some studies used evaluation metrics more than
the others, as the highest number of metrics used in one study is eight evaluation
metrics, while the lowest number of used metrics is one only, some studies did not
use any metrics due to its literature review nature. Consequently, the usage of
68
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
evaluation metrics among a study aligns proportionally to its certainty, reliability and
credibility, the more metrics used the better the results.
C. Summary of Datasets
Synthesis of the studies [34]-[60] illustrated the use of many different datasets to
train detection models using deep learning techniques, including forty-four different
datasets as shown in figure 6.

Figure 6. Datasets Repetitiveness Among Studies

As shown in figure 6, results of our analysis illustrate the percentage of repetitiveness
of many different datasets among the synthesized studies, moreover, it illustrates that
the most widely used dataset for models training is FaceForensics++ (FF++), while
the least used datasets are DF, WDF, VoxCeleb2, ForgeryNet, C40, CelebA-HQ,
All-in-One-DF, Trump Cage, Nicolas Cage, CD2, DFDC-P, Flickr, ILSVRC12,
100K-Faces, DFFD, CASIA-WebFace, StyleGAN-Face1, StyleGAN-Face2,
PGGAN-Face, CycleGAN-Zebra, CASIA 2.0, RFFD, HHF, CASIA v1.0, UADF,

69
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
DFTMIT, Real and Fake Face-Detection, Deeper Forensics, UADFV, and
STARGANK.
This analysis gives us insights about the mostly used datasets in researches within
Deepfake field, as some datasets are preferable more than others mostly due to its
large-scale as it contains a numerous number of images and videos, as well as the
variety of characteristics it provide, making them a good option to train models and
to enhance its capabilities effectively. Moreover, it can be noted that some studies
used a number of datasets more than others, as the highest number of datasets used
in one study is six, while the lowest number of used datasets is one only, some studies
did not use any datasets due to its literature review nature. Furthermore, it can be
said that the more used datasets the better the detection model abilities, performance
and accuracy, in addition to that the number of datasets among a study aligns directly
to its certainty, and inversely with the level of bias.
(3) Summary of challenges of Deepfake detection development
As shown in figure 7 below which summarizes all challenges and limitations of
Deepfake detection techniques among synthesized studies, challenges were divided
into machine learning generated threats, which refer to the advanced techniques built
by machine learning specifically to complicate the detection process, including
adversarial attacks, transferable attacks, trace removal attacks, FakePolisher and
others. While other challenges focused on other obstacles related to the detection
process development itself, including generalization abilities, intensive
computational resources, the need for continuous evolution and adapting, the need
for numerous datasets for training and others.

70
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
 Figure 7. Summary of Challenges and Limitations of Deepfake Detection Techniques

4. Conclusion & Discussion

- Summary of Key Findings
In conclusion, our systematic review provided useful insights into Deepfake
technology ethical implications and impact on society and media, detection
techniques development and related challenges.
1- Ethical Implications of Deepfake: our review has highlighted many several
ethical concerns that surround Deepfake technology, which include issues related to
deception, manipulation and the decrease of authenticity in media content. Main
themes like consent, privacy, as well as the impact of public trust have emerged as
critical considerations along with the emergence of Deepfake. Moreover, non-
consensual use especially in some sensitive areas raises significant concerns about
individuals reputation harm.

71
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
2- Detection techniques for Deepfake: this review has explored and identified a
wide range of techniques that are used or proposed for Deepfake content detection,
leveraging different deep learning methods such as watermarking, FST-Matching
and DAG-FDD. Moreover, evaluation metrics played an important role for verifying
reliability of the results and assessing effectiveness of detection models, which
included Accuracy, AUC, FPR, TPR, precision and many others. Furthermore,
several datasets were utilized for the training and testing of detection models, which
involves FF++, CelebA, FFHQ, DFDC and others.
3- Challenges Associated with Deepfake Detection Development: This review
covered practical constraints faced during the exploration of Deepfake detection
techniques, such as the need for extensive high-quality training and testing datasets,
the need for intensive computational resources, and ensuring responsible use of
detection techniques without violating privacy or civil liberties. These challenges
reflect the nature of developing Deepfake detection strategies and emphasize the
need to overcome difficulties in order to advance the efficacy and reliability of
detection techniques, to mitigate the potential harm posed by this rapidly evolving
technology.
- Conclusion
Deepfake technology presents significant ethical implications, social impact, and
challenges that necessitate robust detection techniques and regulatory frameworks.
The synthesis of relevant studies highlights the importance of addressing ethical
concerns, enhancing detection methods, and overcoming related challenges to
mitigate the risk of Deepfake content. Efforts to address issues related to the
emergence of Deepfake are crucial to ensure the reliability, transparency, and
ethical use of synthetic media in the digital age.

72
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
 - Recommendations for Future Research
While our review provided valuable insights into this field, several gaps and areas
for further research have been identified, which involves enhancing the robustness
and scalability of Deepfake detection techniques to keep pace with evolving
Deepfake generation methods. Moreover, it is important to investigate the integration
of explainable AI techniques to improve the interpretability and transparency of
Deepfake detection models. Furthermore, researchers must explore the development
of real-time Deepfake detection systems to address the challenges of detecting
rapidly evolving Deepfake content. We also suggest conducting studies on the impact
of Deepfake technology on various sectors such as politics, journalism, and
entertainment to understand its broader implications. Finally, we recommend
collaboration across disciplines to develop comprehensive frameworks.
- Implications for Industry and Policy
Industry stakeholders need to invest in robust Deepfake detection technologies to
safeguard their platforms and users from malicious activities. In addition to that,
policymakers should prioritize the development of comprehensive legal frameworks
to regulate the creation, distribution, and malicious use of Deepfake content.
Furthermore, collaboration between industry, policymakers, and technology experts
is crucial to establish guidelines and policies that address the ethical, legal, and
societal implications of Deepfake technology. Finally, it is important to implement
measures to enhance transparency, accountability, and trust in digital media
platforms.
5. Other Information
In regard to additional information, this study was not officially published, therefore,
it does not have a registration number. The study was not financially supported. The

73
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
authors declare no lack of conflicts of interest related to this study. Furthermore, no
applicable Institutional Review Board (IRB) statement is provided, and the same for
informed consent statements. Finally, there is no data availability statement as it is
not applicable to this study.
References
[1] Heidari, A., Jafari Navimipour, N., Dag, H., & Unal, M. (2023). Deepfake detection using deep
learning methods: A systematic and comprehensive review. Wiley Interdisciplinary Reviews: Data
Mining and Knowledge Discovery, e1520.
[2] Hancock, J. T., & Bailenson, J. N. (2021). The social impact of Deepfakes. Cyberpsychology,
behavior, and social networking, 24(3), 149-152.
[3] Arshed, M. A., Mumtaz, S., Ibrahim, M., Dewi, C., Tanveer, M., & Ahmed, S. (2024).
Multiclass AI-Generated Deepfake Face Detection Using Patch-Wise Deep Learning Model.
Computers, 13(1), 31.
[4] Nightingale, S. J., & Wade, K. A. (2022). Identifying and minimising the impact of fake visual
media: Current and future directions. Memory, Mind & Media, 1, e15.
[5] Hoque, M. A., Ferdous, M. S., Khan, M., & Tarkoma, S. (2021). Real, forged or deep fake?
enabling the ground truth on the internet. IEEE Access, 9, 160471-160484.
[6] Ramachandran, V., Hardebolle, C., Kotluk, N., Ebrahimi, T., Riedl, R., & Jermann, P. (2023).
A multimodal measurement of the impact of Deepfakes on the ethical reasoning and affective
reactions of students.
[7] Al-Khazraji, S. H., Saleh, H. H., KHALID, A. I., & MISHKHAL, I. A. (2023). Impact of
Deepfake Technology on Social Media: Detection, Misinformation and Societal Implications. The
Eurasia Proceedings of Science Technology Engineering and Mathematics, 23, 429-441.
[8] De Ruiter, A. (2021). The distinct wrong of Deepfakes. Philosophy & Technology, 34(4), 1311-
1332.
[9] Jabiyev, B., Onaolapo, J., Stringhini, G., & Kirda, E. (2021, October). e-Game of FAME:
Automatic Detection of FAke MEmes. In TTO (pp. 1-11).

74
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
[10] Li, M., & Wan, Y. (2023). Norms or fun? The influence of ethical concerns and perceived
enjoyment on the regulation of Deepfake information. Internet Research.
[11] De Ruiter, A. (2021). The distinct wrong of Deepfakes. Philosophy & Technology, 34(4),
1311-1332.
[12] Groh, M., Epstein, Z., Firestone, C., & Picard, R. (2022). Deepfake detection by human
crowds, machines, and machine-informed crowds. Proceedings of the National Academy of
Sciences, 119(1), e2110013119.
[13] Ali, S., DiPaola, D., Lee, I., Sindato, V., Kim, G., Blumofe, R., & Breazeal, C. (2021).
Children as creators, thinkers and citizens in an AI-driven future. Computers and Education:
Artificial Intelligence, 2, 100040.
[14] Godulla, A., Hoffmann, C. P., & Seibert, D. (2021). Dealing with Deepfakes–an
interdisciplinary examination of the state of research and implications for communication studies.
SCM Studies in Communication and Media, 10(1), 72-96.
[15] Widder, D. G., Nafus, D., Dabbish, L., & Herbsleb, J. (2022, June). Limits and possibilities
for “Ethical AI” in open source: A study of Deepfakes. In Proceedings of the 2022 ACM
Conference on Fairness, Accountability, and Transparency (pp. 2035-2046).
[16] Wazid, M., Mishra, A. K., Mohd, N., & Das, A. K. (2024). A Secure Deepfake Mitigation
Framework: Architecture, Issues, Challenges, and Societal Impact. Cyber Security and
Applications, 100040.
[17] Neethirajan, S. (2021). Is seeing still believing? Leveraging Deepfake technology for livestock
farming. Frontiers in Veterinary Science, 8, 740253.
[18] Guarnera, L., Giudice, O., Guarnera, F., Ortis, A., Puglisi, G., Paratore, A., ... & Battiato, S.
(2022). The face Deepfake detection challenge. Journal of Imaging, 8(10), 263.
[19] Masood, M., Nawaz, M., Malik, K. M., Javed, A., Irtaza, A., & Malik, H. (2023). Deepfakes
generation and detection: State-of-the-art, open challenges, countermeasures, and way forward.
Applied intelligence, 53(4), 3974-4026.

75
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
[20] Ivanovska, M., & Struc, V. (2024). On the Vulnerability of Deepfake Detectors to Attacks
Generated by Denoising Diffusion Models. In Proceedings of the IEEE/CVF Winter Conference
on Applications of Computer Vision (pp. 1051-1060).
[21] Wang, L., Meng, X., Li, D., Zhang, X., Ji, S., & Guo, S. (2023). DeepfakeR: A Unified
Evaluation Platform for Facial Deepfake and Detection Models. ACM Transactions on Privacy and
Security.
[22] Dong, S., Wang, J., Liang, J., Fan, H., & Ji, R. (2022, October). Explaining Deepfake detection
by analysing image matching. In European Conference on Computer Vision (pp. 18-35). Cham:
Springer Nature Switzerland.
[23] Zhao, Y., Liu, B., Ding, M., Liu, B., Zhu, T., & Yu, X. (2023). Proactive Deepfake defence
via identity watermarking. In Proceedings of the IEEE/CVF winter conference on applications of
computer vision (pp. 4602-4611)
[24] Ju, Y., Hu, S., Jia, S., Chen, G. H., & Lyu, S. (2024). Improving Fairness in Deepfake
Detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer
Vision (pp. 4655-4665).
[25] Park, J., Park, L. H., Ahn, H. E., & Kwon, T. (2024). Coexistence of Deepfake Defenses:
Addressing the Poisoning Challenge. IEEE Access.
[26] Hussain, S., Neekhara, P., Dolhansky, B., Bitton, J., Ferrer, C. C., McAuley, J., & Koushanfar,
F. (2022). Exposing vulnerabilities of Deepfake detection systems with robust attacks. Digital
Threats: Research and Practice (DTRAP), 3(3), 1-23.
[27] Wang, X., Huang, J., Ma, S., Nepal, S., & Xu, C. (2022). Deepfake disrupter: The detector of
Deepfake is my friend. In Proceedings of the IEEE/CVF Conference on Computer Vision and
Pattern Recognition (pp. 14920-14929).
[28] Liu, C., Chen, H., Zhu, T., Zhang, J., & Zhou, W. (2023). Making Deepfakes more spurious:
evading deep face forgery detection via trace removal attack. IEEE Transactions on Dependable
and Secure Computing.
[29] Yang, Y., Liang, C., He, H., Cao, X., & Gong, N. Z. (2021). Faceguard: Proactive Deepfake
detection. arXiv preprint arXiv:2109.05673.

76
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
[30] Dong, S., Wang, J., Liang, J., Fan, H., & Ji, R. (2022, October). Explaining Deepfake detection
by analysing image matching. In European Conference on Computer Vision (pp. 18-35). Cham:
Springer Nature Switzerland.xa
[31] George, A. S., & George, A. H. (2023). Deepfakes: The Evolution of Hyper realistic Media
Manipulation. Partners Universal Innovative Research Publication, 1(2), 58-74.
[32] Neekhara, P., Dolhansky, B., Bitton, J., & Ferrer, C. C. (2021). Adversarial threats to Deepfake
detection: A practical perspective. In Proceedings of the IEEE/CVF conference on computer vision
and pattern recognition (pp. 923-932).
[33] Hussain, S., Neekhara, P., Jere, M., Koushanfar, F., & McAuley, J. (2021). Adversarial
Deepfakes: Evaluating vulnerability of Deepfake detectors to adversarial examples. In Proceedings
of the IEEE/CVF winter conference on applications of computer vision (pp. 3348-3357).
[34] Zhao, T., Xu, X., Xu, M., Ding, H., Xiong, Y., & Xia, W. (2021). Learning self-consistency
for Deepfake detection. In Proceedings of the IEEE/CVF international conference on computer
vision (pp. 15023-15033).
[35] Shad, H. S., Rizvee, M. M., Roza, N. T., Hoq, S. M., Monirujjaman Khan, M., Singh, A., ...
& Bourouis, S. (2021). Comparative analysis of Deepfake image detection method using
convolutional neural network. Computational Intelligence and Neuroscience, 2021.
[36] Hsu, C. C., Zhuang, Y. X., & Lee, C. Y. (2020). Deep fake image detection based on pairwise
learning. Applied Sciences, 10(1), 370.
[37] Raza, A., Munir, K., & Almutairi, M. (2022). A novel deep learning approach for Deepfake
image detection. Applied Sciences, 12(19), 9820.
[38] Frank, J., Eisenhofer, T., Schönherr, L., Fischer, A., Kolossa, D., & Holz, T. (2020,
November). Leveraging frequency analysis for deep fake image recognition. In International
conference on machine learning (pp. 3247-3258). PMLR.
[39] Suganthi, S. T., et al. "Deep learning model for deep fake face recognition and detection."
PeerJ Computer Science 8 (2022): e881.

77
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
[40] Pu, J., Mangaokar, N., Wang, B., Reddy, C. K., & Viswanath, B. (2020, December).
Noisescope: Detecting Deepfake images in a blind setting. In Annual computer security
applications conference (pp. 913-927).
[41] Ali, S. S., Ganapathi, I. I., Vu, N. S., Ali, S. D., Saxena, N., & Werghi, N. (2022). Image
forgery detection using deep learning by recompressing images. Electronics, 11(3), 403.
[42] Rana, M. S., & Sung, A. H. (2020, August). Deepfakestack: A deep ensemble-based learning
technique for Deepfake detection. In 2020 7th IEEE international conference on cyber security and
cloud computing (CSCloud)/2020 6th IEEE international conference on edge computing and
scalable cloud (EdgeCom) (pp. 70-75). IEEE.
[43] Zhao, H., Zhou, W., Chen, D., Wei, T., Zhang, W., & Yu, N. (2021). Multi-attentional
Deepfake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern
recognition (pp. 2185-2194).
[44] Guarnera, L., Giudice, O., & Battiato, S. (2020). Deepfake detection by analyzing
convolutional traces. In Proceedings of the IEEE/CVF conference on computer vision and pattern
recognition workshops (pp. 666-667).
[45] Rafique, R., Nawaz, M., Kibriya, H., & Masood, M. (2021, November). Deepfake detection
using error level analysis and deep learning. In 2021 4th International Conference on Computing
& Information Sciences (ICCIS) (pp. 1-4). IEEE.
[46] Arshed, M. A., Alwadain, A., Faizan Ali, R., Mumtaz, S., Ibrahim, M., & Muneer, A. (2023).
Unmasking Deception: Empowering Deepfake Detection with Vision Transformer Network.
Mathematics, 11(17), 3710.
[47] khormali, A., & Yuan, J. S. (2021). Add: Attention-based Deepfake detection approach. Big
Data and Cognitive Computing, 5(4), 49.
[48] Jasim, R. M., & Atia, T. S. (2023). An evolutionary-convolutional neural network for fake
image detection. Indonesian Journal of Electrical Engineering and Computer Science, 29(3), 1657-
1667.

78
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
[49] Sabitha, R., Aruna, A., Karthik, S., & Shanthini, J. (2021). Enhanced model for fake image
detection (EMFID) using convolutional neural networks with histogram and wavelet based feature
extractions. Pattern Recognition Letters, 152, 195-201.
[50] Yu, C. M., Chang, C. T., & Ti, Y. W. (2019). Detecting Deepfake-forged contents with
separable convolutional neural network and image segmentation. arXiv preprint
arXiv:1912.12184.
[51] Song, L., Fang, Z., Li, X., Dong, X., Jin, Z., Chen, Y., & Lyu, S. (2022, October). Adaptive
face forgery detection in cross domain. In European Conference on Computer Vision (pp. 467-
484). Cham: Springer Nature Switzerland.
[52] Ali, S., DiPaola, D., Lee, I., Hong, J., & Breazeal, C. (2021, May). Exploring generative
models with middle school students. In Proceedings of the 2021 CHI conference on human factors
in computing systems (pp. 1-13).
[53] Allein, L., Moens, M. F., & Perrotta, D. (2023). Preventing profiling for ethical fake news
detection. Information Processing & Management, 60(2), 103206.
[54] De Ruiter, A. (2021). The distinct wrong of Deepfakes. Philosophy & Technology, 34(4),
1311-1332.
[55] Xu, Y., & Yayilgan, S. Y. (2022). When Handcrafted Features and Deep Features Meet
Mismatched Training and Test Sets for Deepfake Detection. arXiv preprint arXiv:2209.13289.
[56] Nawaz, M., Javed, A., & Irtaza, A. (2023). ResNet-Swish-Dense54: a deep learning approach
for deepfakes detection. The Visual Computer, 39(12), 6323-6344.
[57] Nirkin, Y., Wolf, L., Keller, Y., & Hassner, T. (2020). Deepfake detection based on the
discrepancy between the face and its context. arXiv preprint arXiv:2008.12262.
[58] Taeb, M., & Chi, H. (2022). Comparison of deepfake detection techniques through deep
learning. Journal of Cybersecurity and Privacy, 2(1), 89-106.
[59] Fung, S., Lu, X., Zhang, C., & Li, C. T. (2021, July). Deepfakeucl: Deepfake detection via
unsupervised contrastive learning. In 2021 international joint conference on neural networks
(IJCNN) (pp. 1-8). IEEE.

79
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361
[60] Patel, Y., Tanwar, S., Bhattacharya, P., Gupta, R., Alsuwian, T., Davidson, I. E., & Mazibuko,
T. F. (2023). An improved dense cnn architecture for deepfake image detection. IEEE Access, 11,
22081-22095.

80
International Journal of Computers and Informatics, London Vol (3), No (8), 2024
https://doi.org/10.59992/IJCI.2024.v3n8p3 E-ISSN 2976-9361