Identity Consistency For Enhanced Deep Fake

Detection In Video Content Using Deep

Learning

Pothuganti Priyanka Rehana Shaik

Dept. of Computer Science and Engineering Dept. of Computer Science and Engineering
MLR Institue Of Technology MLR Institue Of Technology
Hyderabad, India. Hyderabad,India.
[email protected] [email protected]

Samala Sneha Guguloth Sagar

Dept. of Computer Science and Engineering Dept. of Computer Science and Engineering
MLR Institute Of Technology MLR Institue Of Technology
Hyderabad,India. Hyderabad,India.
[email protected] [email protected]

Abstract—Deep Fake Videos are a form of Deep Learning, Face swaping , 3DMM
synthetic media in which Artificial Genrative Network, Identity-based
intelligence techniques are used to Swap features, Facial Reenactment.
faces, manipulate voices, or change
identities in a video to make it appear I. INTRODUCTION
authentic. Detecting deepfake videos is
challenging because most current Deep fake Video Detection via Data
algorithms are designed to identify specific Analysis focuses on detecting manipulated
types of fakes and often fail to generalize videos using deep learning techniques,
across different manipulation techniques, analyzing identity-based features like facial
such as face swapping or facial patterns and voice cues to identify fakes. It
reenactment. This project introduces ID- aims to address the growing concern of
Reveal, an advanced deep fake detection deepfake misuse in digital media. The study
system that utilizes identity analysis mainly concentrated on using Temporal
through the temporal ID Network and the Identity analysis, which is used to track
3DMM Generative Network to enhance unique patterns over time, and 3D facial
accuracy and robustness. feature modeling for precise identity
Keywords—Deep fake, Deep Learning, recognition. Behavioral analysis of unique
Temporal and ID – Reveal, 3D Facial body movements or gestures. This deep
Feature Modeling, Behavioral analysis, learning-based project involves collecting
 and preprocessing real and synthetic video specially AEI-Net and HEAR-Net were
data and designing and training for temporal leveraged to integrate face information.
pattern detection. This project aims to In a comprehensive survey conducted in
address the growing concern of deepfake 2021, the authors highlighted various
misuse in digital media by providing a detection techniques while revealing
solution to reliably detect deepfake content. significant limitations regarding
Tempral ID: refers to a unique identifier generalization across different datasets and
associated with data that changes or evolves manipulation types. These limitations hinder
over time. Temporal IDs are valuable for the effectiveness of existing models in real-
understanding and managing data in world scenarios, which demand adaptability
contexts where time is a critical factor. to diverse inputs [2]. To overcome these
3D Facial Feature Modeling: refers to the challenges, the ID-Reveal system [3] was
process of creating three-dimensional developed, leveraging prior biometric
representations of facial structures, often characteristics to improve detection
used in fields such as computer graphics, accuracy. By focusing on individual-specific
animation, virtual reality, and deep learning facial features, ID-Reveal effectively
for tasks like facial recognition and deepfake identifies facial manipulations, although its
detection. performance may decline with unseen
manipulation methods due to reliance on
II. LITRATURE SURVEY high-quality training data. Furthermore, an
analysis of different convolutional neural
This section briefly presents the works done network (CNN) architectures [4] illustrated
on deepfake detection using deep learning that certain models can significantly
with identity reveal and 3DMM. enhance detection accuracy. However, the
FeatureTransfer: Unsupervised domain rapid evolution of deepfake generation
adaptation for cross-domain deepfake techniques poses a challenge, as many
detection Chen, B., & Tan, S. (2021) [1] existing detection methods struggle to keep
This research concentrates on a technique pace with these advancements. Meng and
based on unsupervised domain adaption, it Xiao [5] emphasized the importance of
solves many Over fitting problems, Unlike utilizing temporal information in videos,
the end-to-end adversarial training method employing Long Short-Term Memory
NANN, Feature Transfer exploits a two- (LSTM) networks for enhanced detection
stage adversarial training pipeline. In capabilities. Their work pointed to the
Feature Transfer, a two-stage Deep fake critical role of analyzing specific facial
detection method based on unsupervised regions for authenticity assessments, while
domain adaptation, is proposed. The feature also noting limitations related to dataset size
vectors extracted from CNN are used for and quality. A systematic review in 2024
adversarial transfer learning in BP_DANN, categorized deepfake detection methods by
which contributes to better performance. application type and performance,
A Survey on Deepfake Video Detection highlighting the need for scalable and real-
Peipeng Yu, Zhihua Xia & Jianwei Fei time solutions. It acknowledged that existing
(2021) [21] Like previous face-swapping literature might not encompass all emerging
studies only limited information from target methods, emphasizing the urgency for
images to synthesize faces, FaceShifter continual updates in detection strategies [6].
generates high-fidelity swapped faces by The integration of Generative Adversarial
performing comprehensive integration Networks (GANs) into detection systems
 revealed substantial advancements; the study grapples with challenges related to
however, challenges in model robustness the choice of hyperparameters [25] and class
and adaptability to evolving manipulation imbalance.
techniques were identified as critical barriers
[7]. Moreover, current detection methods III. METHODOLOGY
often struggle with achieving real-time
performance due to high computational The project involves a comprehensive process of
demands and the sophistication of deepfake data collection and preprocessing, where both
real and synthetic video datasets will be
generation tools [8]. A recurring theme
gathered. These datasets will then be
across the literature is the pressing need for meticulously curated to ensure that they
diverse datasets that capture various encompass a wide variety of scenarios,
manipulation techniques, as many existing identities, and manipulation techniques. Once
models lack generalization capabilities [9]. the data is prepared, deep learning models will
Recent studies highlight that most methods be designed and trained to optimize the detection
are limited in their application to specific of temporal patterns and identity mismatches.
datasets and are not necessarily effective in Leveraging advanced techniques such as
broader contexts. Consequently, researchers Convolutional Neural Networks (CNNs) for
stress the need for innovative methodologies spatial feature extraction and Long short-term
that can adapt to the evolving landscape of memory (LSTM) networks for temporal
analysis, the system aims to deliver a
deepfake technology, ensuring both
sophisticated solution for identifying
accuracy and reliability in diverse contexts manipulated media. By integrating various data
[10]. The study [11] into detection systems analysis techniques and deep learning
revealed substantial advancements but methodologies, this project aspires to contribute
pointed out challenges in model robustness significantly to the fight against deepfake
and adaptability to evolving manipulation proliferation, ensuring a higher level of
techniques. Furthermore, current detection authenticity and trust in digital media platforms.
methods often struggle with achieving real-
time performance due to high computational
demands and the sophistication of deepfake
generation tools. A recurring theme in the
literature is the need for diverse datasets that
capture various manipulation techniques, as
many existing models lack generalization
capabilities. Ultimately, these studies
highlight the pressing need for innovative Deep Fake Video Detection
methodologies capable of adapting to the The Deepfake videos are increasingly
evolving landscape of deepfake technology, harmful to personal privacy and social
ensuring both accuracy and reliability in security. Various methods have been
diverse contexts. However, the approach proposed to detect manipulated videos.
demands fine-tuning and hyperparameter Early attempts mainly focused on
tuning, real-time inference, and is reliant on inconsistent features caused by the face
preprocessing. Mahmon and Yaacob [15] synthesis process while current detection
provided insights on classifying the satellite methods mostly target at fundamental
images using ANNs and Maximum features. As shown in Table 1, these
Likelihood Algorithm. Despite an accuracy methods fall into five categories based on
of 89.3% and a kappa coefficient of 0.820, the features they use. To begin with,
detection based on general neural networks movements and expressions.
is commonly used in literature, where
deepfake detection task is considered as
regular classification tasks. Temporal
consistency features are also exploited in
detection tasks. Recently proposed
approaches focus on more fundamental
features, where camera fingerprint and
biological signal‐based schemes show great
potential in detection tasks. In the following
sections, we will review the detection
methods mentioned below:
ID-Reveal: it extracts the identity of the Deep fake detection techniques often
person being impersonated, either struggle with generalization, especially
intentionally or unintentionally, within the when faced with unseen manipulation
generated content. methods. The performance of these methods
largely depends on the diversity and quality
of the training datasets. If the training data
doesn't include a wide range of manipulation
techniques or variations, the model may fail
to accurately identify or classify new deep
fakes that employ different methods.
Additionally, as deep fake technology
continues to evolve, new and more
sophisticated manipulation techniques can
emerge, further complicating the ability of
existing detection models to keep pace
Temporal identity: this identity is related to
how consistently a synthesized face or a
person maintains their identity over time in
video sequences, Temporal id is crucial
because deepfakes need to retain realistic
consistency across frames so that viewers do
not notice the fluctuations.
3D Facial Feature Modeling: Also known as
3D Morphable Model(3DMM) it’s a
powerful tool used in deepfakes to create
realistic and detailed 3D representations of
human faces, it helps in mapping and
modifying access into three dimensions
allowing detection of complex head

From deep learing we usually use :-

 Feedforward Neural Networks: pictures and other little snippets. The ethical
Information moves in one direction, from implications of DeepFake technology
input to output. necessitate ongoing research into effective
 Convolutional Neural Networks detection strategies, emphasizing the
(CNNs): Primarily used for image data, they importance of real-time applications in
automatically detect features and patterns. safeguarding information integrity in digital
 Recurrent Neural Networks (RNNs): media.
Suitable for sequential data, like time series
or text, where the output depends on VI. REFERENCES
previous inputs.
 Transformers: Widely used in natural [1] M. S. Rana, M. N. Nobi, B. Murali and
language processing, they utilize attention A. H. Sung, "Deepfake Detection: A
mechanisms to weigh the importance of Systematic Literature Review," in IEEE
different input parts. Access, vol. 10, pp. 25494-25513, 2022, doi:
These have helped a lot with face detection 10.1109/ACCESS.2022.3154404.
and probability generation if the picture [2] D. Pan, L. Sun, R. Wang, X. Zhang and
might be a real or a deepfake. R. O. Sinnott, "Deepfake Detection through
IV. CONCLUSION Deep Learning," 2020 IEEE/ACM
International Conference on Big Data
We explored the applications of identity Computing, Applications and Technologies
consistency as a pivotal factor in enhancing (BDCAT), Leicester, UK, 2020, pp. 134
the detection of deepfakes in video content 143, doi:
through deep learning methodologies. Our 10.1109/BDCAT50828.2020.00001.
method uses ID-reveal, 3DMM, temporal ID [3] A. Mary and A. Edison, "Deepfake
to detect any inconsistencies in facial Detection using deep learning techniques: A
features, expressions, and movements across Literature Review," 2023 International
one frame to another, we demonstrated a Conference on Control, Communication and
significant improvement in the model’s Computing (ICCC), Thiruvananthapuram,
accuracy and robustness against various India, 2023, pp. 1-6, doi:
manipulation techniques. 10.1109/ICCC57789.2023.10164881.
The result indicates that identity consistency [4] A. Malik, M. Kuribayashi, S. M.
serves as a reliable indicator of authenticity Abdullahi and A. N. Khan, "DeepFake
allowing our model to differentiate between Detection for Human Face Images and
real and altered content, This gives a rapid Videos: A Survey," in IEEE Access, vol. 10,
evolution to the field of deepfake pp. 18757-18775, 2022, doi:
technologies. By incorporating of identity 10.1109/ACCESS.2022.3151186
features we improve detection rates and the [5] S. A. H. O. Farag, W. I. El-Horbaty, and S. Z.
model’s ability to generalize across unseen M. Badawy - "A Systematic Review of Deepfake
manipulation methods. Detection Techniques" in Journal of Computer
Virology and Hacking Techniques, 2023. DOI:
V. FUTURE SCOPE 10.1007/s11416-023-00459-1.
[6] M. Z. Khan, M. A. L. Farooq, and K. R.
Future work should further focus on refining Munir - "Deepfake Detection Using Eye Blinking
our approach by incorporating contextual Dynamics," in Journal of Ambient Intelligence
information and exploring multi-modal data and Humanized Computing, vol. 13, no. 7, pp.
sources, we could also include the use of
 3479-3491, 2022. doi: 10.1007/s12652-021-
03457-8.
[7] S. B. Ravi and K. P. S. R. V. D. D. S. K. Prasad -
"DeepFake Detection Techniques: A Review and
Comparative Analysis" in Future Generation
Computer Systems, vol. 139, pp. 33-45, 2023.
DOI: 10.1016/j.future.2023.06.028.

[8] A. A. Akinwande, A. Omisore, M. S. Aloba,

and S. O. Ogunyemi - "Deepfake Detection: A
Review of Deep Learning Approaches" in IEEE
Access, vol. 10, pp. 18743-18758, 2022. DOI:
10.1109/ACCESS.2022.3150194.

[9] K. G. M. Z. Ali and A. M. M. Shalabi -

"Deepfake Detection using Convolutional
Neural Networks: A Review" in Journal of King
Saud University - Computer and Information
Sciences, 2023. DOI:
10.1016/j.jksuci.2023.05.008.

[10] C D. Cozzolino, A. Rössler, J. Thies, M.

Nießner, and L. Verdoliva - "ID-Reveal: Identity-
aware DeepFake Video Detection," in
Proceedings of the IEEE/CVF International
Conference on Computer Vision (ICCV), 2021,
pp. 15108-15117. doi:
10.1109/ICCV48922.2021.01508.

[11] Y. R. Kim, J. J. Shapiro, and T. E. G. H. Zeid -

"Detecting Deepfakes with Identity
Consistency," in IEEE Transactions on
Information Forensics and Security, vol. 18, pp.
1234-1247, 2023. doi:
10.1109/TIFS.2023.3245630.