VISVESVARAYA TECHNOLOGICAL UNIVERSITY

Jnana Sangama, Belagavi, Karnataka – 590018

GOVERNMENT ENGINEERING COLLEGE , KUSHALNAGAR – 571234

Department of Computer Science and Engineering
Project Presentation on
“DEEPFAKE DETECTION USING PYTHON”
Presented by
NAYANA C S 4GL22CS030
VINUTHA S M 4GL22CS055
DARSHAN DEVAIAH K S 4GL23CS404
GOPAL KRISHNA M 4GL23CS407
Under the Guidance of
Dr. Ranganatha S
 CONTENTS
 Abstract
 Introduction
 Literature Survey
 Objectives
 Existing System and Disadvantage
 Proposed System and Advantage
 System Requirements :
Hardware Requirements
Software Requirements
 System Design :
High level Diagram
Low level Diagram
 Implementation
 Conclusion
 References
 ABSTRACT

The rise of deepfake technology, which uses artificial intelligence to generate

hyper-realistic fake videos, presents significant challenges to digital media
security, privacy, and the spread of misinformation.
These manipulated videos can convincingly mimic real individuals, making
them difficult to distinguish with the naked eye.
As deepfakes become more accessible and sophisticated, detecting them has
become a crucial task to ensure the integrity of digital information.
Python libraries such as OpenCV, TensorFlow, and Scikit learn are employed
to handle image preprocessing, deep learning model development, and
classification tasks.
 Introduction

 Deep fake technology, which uses artificial intelligence to create realistic fake
videos, has become a growing concern due to its potential misuse in spreading
misinformation, manipulating public opinion, and causing reputational damage.
These manipulated videos can convincingly alter faces, voices, and actions,
making it difficult to distinguish them from real content.
As deep fakes become more sophisticated, the need for reliable detection
methods becomes increasingly urgent.
Detecting deepfakes manually is impractical, especially with large volumes of
content.
 Literature survey
Title Author Methodology Limitations
“Deepfake Acchardee This paper provides a • Overreliance on Deep Learning :
video detection: p Kaur et comprehensive survey on The study emphasizes the dominance of
Challenges and al.[2] deepfake video detection, deep learning-based methods in detecting
opportunities “ focusing on the challenges and deepfakes,
opportunities in the field. It • Dataset Quality Issues :
classifies the various challenges It critically evaluates deepfake datasets,
encountered in detecting emphasizing the necessity for high-quality
deepfake videos, including issues datasets to improve detection methods.
related to unbalanced datasets,
inadequate labeled training data
“Deepfake Yogesh. It offers an in-depth review of • Lack of Practical Implementation:
generation and Patel et al. various machine learning (ML) It mostly remains theoretical and doesn’t
detection: Case [3] and deep learning (DL) provide hands-on experimental results or
study and techniques employed in testing of detection models.
challenges” synthesizing deepfake content, • Limited Dataset Analysis: It discusses
with a particular focus on popular datasets but doesn’t use or analyze
generative adversarial networks new or customized data, which limits its
(GANs). practical value for testing or developing
real-world detection tools.
 The paper by Khalifa et al.
proposes a CNN-based The main limitations of this paper is
“Convolution deepfake detection method that it primarily focuses on CNNs with
al neural that integrates diverse Gabor Gabor filters, which may not
network filters—linear, elliptical, and generalize well to newer or more
based on Ahmed H. circular—into the network. complex deepfake generation
Khalifa, et These filters are adaptively techniques. Additionally, the proposed
diverse al.[1] learned during training to model might require substantial
gabor filters enhance feature extraction, computational resources due to the
for deepfake and the model uses a dual- large receptive fields and multiple
recognition” scale architecture to improve filter types, limiting its real-time
recognition accuracy across applicability
various datasets.
The paper by Masood et al.
Deepfakes provides a comprehensive The main limitations of this paper is
generation review of deepfake generation its lack of original experimental
and and detection techniques, validation and limited exploration of
detection: focusing on both audio and real-world applications for deepfake
state-of-the- Momina visual content. It categorizes detection. Additionally, while it
Masood,
art, open et al.[4] deepfake manipulation addresses both visual and audio
challenges, methods, discusses current deepfakes, the focus is more on visual
countermeas machine learning-based content, leaving audio deepfake
detection approaches, and detection less explored.
ures, and
evaluates public datasets and
way forward,
“Unmasking Fakhar The methodology involves Limited Scope of Audio-Visual
deepfakes:A Abbas, et conducting a comprehensive Deepfakes: The paper primarily focuses on
systematic review al.[5] literature analysis to explore key visual deepfakes, with less emphasis on
of deepfake detection and generation methods, audio-visual deepfakes, which are
detection and frameworks, algorithms, and tools increasingly prevalent in real-world
generation for identifying deepfakes across scenarios
techniques using audio, images, and videos. Lack of Original Experimental
artificial Validation: The study does not present new
intelligence,” experimental results or validate the
discussed techniques through original
experiments
.
“Deepfake Md Shohel The authors systematically • Narrow Scope: The review focuses
detection: A Rana, et al. reviewed 112 articles published exclusively on image and video
systematic [6] between 2018 and 2020, manipulation, excluding audio and text-
literature review,” categorizing the detection based deepfakes, which limits its
techniques into four groups: deep comprehensiveness.
learning-based methods, classical • Lack of Unified Evaluation Metrics:
machine learning approaches, The absence of standardized evaluation
statistical techniques, and metrics across studies makes it
blockchain-based solutions. challenging to compare the effectiveness
of different detection methods.
 In this paper, the authors created two • Limited Dataset Diversity: The study
distinct datasets containing Knee focuses on two specific datasets—Knee
“A new approach Osteoarthritis X-ray and lung CT Osteoarthritis X-rays and lung CT scans—
for effective scan images, labeling each instance potentially limiting the generalizability of
Mehmet
as real or fake. They applied data the findings to other medical imaging
medical deepfake Karakose, et
pre-processing and augmentation domains
detection in al.[7]
techniques to standardize and vary • Model Generalization Challenges: Deep
medical images” the data. The detection capabilities learning models trained on specific datasets
of several YOLO were evaluated on may struggle to generalize across diverse
these datasets. types of deepfake content
• Dataset Constraints: The effectiveness of
In this paper, The authors categorize detection methods can be influenced by the
existing deepfake creation methods quality and diversity of datasets used. If the
“Deepfake into five major groups and analyze reviewed studies relied on limited or non-
detection for Asad detection approaches based on their representative datasets, it could affect the
results, performance, methodology, generalizability of the findings.
human face Malik,et al. and detection type. They also
images and [8] • Technological Advancements: Given the
examine trends in available deepfake rapid evolution of deepfake generation
videos: A survey” datasets and discuss challenges techniques, the survey's findings may
related to deepfake creation and become outdated if newer methods were not
detection. considered.
 In this paper, they proposed
“Deepfake framework was trained on datasets
detection using • High Computational Requirements: The
such as FaceForensics++ and the
convolutional Ahmed proposed technique necessitates substantial
Deep Fake Detection Challenge
Hatem computational resources for both training
vision Soudy,et al.
(DFDC) dataset. Performance
and inference, which may limit its
transformers and metrics including accuracy,
[9] applicability in resource-constrained
convolutional precision, recall, and F1-score
environments.
neural networks” were utilized to evaluate the
model's effectiveness.
This paper proposes a deep
“Deepfake learning approach using FastText
detection on social embeddings and Convolutional • Dataset Limitation: It relies on the
media: Leveraging Neural Networks (CNNs) to TweepFake dataset, which may not capture
deep learning and Saima detect machine-generated tweets. the full range of machine-generated content
They preprocess tweets from the across all social media platforms.
fasttext Sadiq,et al.
TweepFake dataset, convert text • Model Generalization: The model's
embeddings for [10] into semantic vectors with performance may vary when applied to other
identifying FastText, and classify content forms of social media content beyond tweets,
machine-generated using CNNs, demonstrating strong limiting its broader applicability.
tweets” performance compared to other
models.
 The methodology of this paper
involves preprocessing video
data by extracting and The main limitations of this paper include
preserving facial frames, which limited dataset diversity, which may affect
“A comparative
Nishika are then input into deep model generalization, and a lack of real-
study: deepfake
Khatri,et al. learning-based models to detect time performance evaluation. Additionally,
detection using
[11] deepfakes. The approach the study focuses primarily on accuracy
deep-learning,”
compares different deep without deeply analyzing robustness
learning techniques for against advanced deepfake techniques.
classifying content as real or
fake.
The methodology this paper
introduces a dual-level adapter The limitations of this paper include
module—spatial and temporal potential computational overhead due to
“Deepfake-
—that integrates with existing the dual-level adapter integration and
adapter: Dual-level
Rui Shao,et deep learning models to limited testing across diverse deepfake
adapter for
al. [12] enhance deepfake detection. datasets. The method may also face
deepfake
This adapter effectively challenges in adapting to unseen
detection,”
captures both frame-level and manipulation techniques without fine-
sequence-level inconsistencies tuning.
for improved accuracy.
 Objectives

• Detect fake images, videos, and audio.

• Build trust in online media.

• Automate the detection process.

• Create a fast and accurate detection system.

• Adapt to new deep fake techniques.

 Existing System

• Existing systems for deep fake detection using machine learning primarily rely on
various deep learning architectures tailored to analyze the unique characteristics
of synthetic content.
• Convolutional Neural Networks are the most commonly used models for
detecting deep fakes in images and videos, as they are effective at identifying
spatial inconsistencies like unusual lighting, facial movements, or unnatural
details.
• For videos, Recurrent Neural Networks and LSTMs are employed to capture
temporal patterns, allowing the model to detect inconsistencies over multiple
frames, such as unnatural facial expressions or lip-syncing issues.
 Disadvantages

• High Computational Cost

• Data Dependence
• Overfitting
• Continuous Updates Required
• Data Privacy Concerns
 Proposed System

•A proposed system for deep fake detection using machine learning could
combine multiple advanced techniques to improve detection accuracy across
various types of media (images, videos, and audio).
• The system would begin by using Convolutional Neural Networks for initial
analysis of images and video frames, detecting subtle visual artifacts such as
abnormal lighting, inconsistent skin textures, or unnatural facial expressions
that are common in deep fake content.
• For videos, the system could integrate 3D-CNNs to consider the temporal aspect
of the video, identifying inconsistencies in facial movements or lip-sync issues
across multiple frames.
 Advantages

• High Accuracy
• Scalability
• Adaptability
• Real-Time Detection
• Fraud Prevention
 Hardware Requirements

• Processor–Intel(R) Core(TM) i3-5005U CPU @ 2.00GHz 2.00 GHz

(64-bit OS).

• Memory– 16GB RAM (Higher specs are recommended for high

performance)

• Monitor
• Ethernet connection and Wi Fi
 Software Requirements

• Python 3.8.1
• Windows 10
• IDE/Jupiter Notebook
 System Design :
High level Diagram :
 Low level diagram
Sequence Diagram :
Core Modules for Deep Fake Detection System :

1. Data Collection Module

2. Media Upload Module
3. Preprocessing Module
4. Feature Extraction Module
5. Deep Fake Detection Module
6. Post Processing Module
7. Evaluation or Feedback Module

•
 Implementation
• Prepare Data: Collect images/videos of both real and deep fake content.
Clean and organize this data.

• Create the Model : Use a deep learning model to detect patterns in the data
that show the difference between real and fake content.

• Train the Model: Feed the model your data . Let the model learn the
differences so it can make predictions in the future.

• Test and Evaluate: See how accurate the model is by testing it with new,
unseen data. Measure its performance using accuracy, precision, and recall.

• Deploy: Once it’s trained and working well, use it on new data. You can
integrate it into a website or app to detect deep fakes in real-time.
 Conclusion

• Deepfake detection using Python offers a powerful and flexible approach

to combating the growing threat of manipulated digital media.
• By leveraging Python’s extensive libraries like TensorFlow,Keras, and
OpenCV, developers can build robust detection systems that utilize
advanced machine learning and deep learning techniques, such as
Convolutional Neural Networks, to identify subtle inconsistencies in
videos.
• This enables the development of systems capable of distinguishing
between real and fake content efficiently.
• Overall, Python remains a leading language for developing deepfake
detection tools, offering both the flexibility and power needed for this
critical task.
 References
[1] A. H. Khalifa, N. A. Zaher, A. S. Abdallah, and M. W. Fakhr, “Convolutional neural
network based on diverse gabor filters for deepfake recognition,” IEEE Access, vol. 10,
pp. 22678–22686, 2022.
[2] A. Kaur, A. N. Hoshyar, V. Saikrishna, S. Firmin, and F. Xia, “Deepfake video
detection: Challenges and opportunities,” Artificial Intelligence Review, vol. 57, no. 159,
pp. 1–47, 2024.
[3] M. S. Rana, M. N. Nobi, B. Murali, and A. H. Sung, “Deepfake detection: A
systematic literature review,” IEEE Access, vol. 10, pp. 25494–25510, 2022.
[4] P. Edwards, J.-C. Nebel, D. Greenhill, and X. Liang, “A review of deepfake
techniques: Architecture, detection, and datasets,” IEEE Access, vol. 12, pp. 154718–
154733, 2024.
[5] A. Malik, M. Kuribayashi, S. M. Abdullahi, and A. N. Khan, “Deepfake detection for
human face images and videos: A survey,” IEEE Access, vol. 10, pp. 18757–18777, 2022.
[6] B. M. MD SHOHEL RANA MOHAMMAD NUR NOBI3 and A. H. SUNG2, “Deepfake
detection: A systematic literature review,” IEEE Access, vol. 10, pp. 25494–25513, 2022.
[7] M. Karak¨ose, H. Yeti¸s, and M. C ¸e¸cen, “A new approach for effective medical
deepfake detection in medical images,” IEEE Access, vol. 12, pp. 52205–52214, 2024.
[8] A. Malik, M. Kuribayashi, S. M. Abdullahi, and A. N. Khan, “Deepfake detection for
human face images and videos: A survey,” IEEE Access, vol. 10, pp. 18757–18777, 2022.
[9] A. H. Soudy, O. Sayed, H. Tag-Elser, R. Ragab, S. Mohsen, T. Mostafa, A. A. Abohany,
and S. O. Slim, “Deepfake detection using convolutional vision transformers and
convolutional neural networks,” Springer, vol. 36, no. 31, p. 19759–19775, 2024.
[10] S. Sadiq, T. Aljrees, and S. Ullah, “Deepfake detection on social media: Leveraging
deep learning and fasttext embeddings for identifying machine-generated tweets,” IEEE
Access, vol. 11, pp. 95008 95021, 2023.
[11] N. Khatri, V. Borar, and R. Garg, “A comparative study: deepfake detection using deep-
learning,” in 2023 13th International Conference on Cloud Computing, Data Science &
Engineering (Confluence). IEEE, 2023, pp. 1–5.
[12] R. Shao, T. Wu, L. Nie, and Z. Liu, “Deepfake-adapter: Dual-level adapter for deepfake
detection,” International Journal of Computer Vision, pp. 1–16, 2025.
THANK YOU