VISVESVARAYA TECHNOLOGICAL UNIVERSITY

“Jnana Sangama", Belagavi - 590018

A project synopsis on

“Deep Fake Video Detection Using Deep

Learning”
Submitted in Partial Fulfilment of the requirement for the Fifth Semester

Bachelor of Engineering in
Computer Science and
Design Submitted by:

CHAITANYA N -1SJ23CG010
KRUPA C -1SJ23CG025
NITHYA S -1SJ23CG036
POORNASHREE C -1SJ23CG038

Under the guidance of

SANTOSH KUMAR M
Assistant Professor
Dept. CSD, SJCIT

SJC INSTITUTE OF TECHNOLOGY

DEPARTMENT OF COMPUTER SCIENCE AND DESIGN
CHICKABALLAPUR -562101
 2025 -2026
ABSTRACT

The rapid advancement of deep learning techniques has led to the proliferation of deepfake
videos, which manipulate facial expressions, voices, and movements with high realism.
While such technology has legitimate applications in entertainment and education, it also
poses significant threats, including misinformation, identity theft, and political manipulation.
This study explores the application of deep learning models for the automated detection of
deepfake videos. We propose a hybrid approach combining Convolutional Neural Networks
(CNNs) and Recurrent Neural Networks (RNNs) to analyze both spatial and temporal
inconsistencies in video frames. Preprocessing techniques such as face extraction and frame
sampling are employed to enhance detection accuracy. The model is trained and evaluated
using benchmark datasets like FaceForensics++ and DFDC. Experimental results
demonstrate high accuracy in distinguishing real from fake content, even under compression
and noise. Our findings indicate that deep learning offers a promising solution for real-time
deepfake detection, contributing to digital media integrity and online trust. Future work will
explore model robustness against adversarial attacks and cross-dataset generalization.
 INTRODUCTION

In recent years, the development of deep learning and generative models has led to the
emergence of highly realistic manipulated media known as deepfakes. These videos use
techniques such as Generative Adversarial Networks (GANs) and autoencoders to
synthetically alter a person's appearance, facial expressions, or voice in a way that is often
indistinguishable from authentic footage. While deepfake technology holds potential in
entertainment, film, and gaming, its misuse raises serious ethical, social, and security
concerns. From spreading misinformation and creating fake news to damaging reputations
and enabling fraud, the impact of deepfakes can be far-reaching.
Traditional methods for detecting video tampering often fail to keep up with the
sophistication of modern generative models. As a result, there is a growing demand for
intelligent, automated systems capable of detecting deepfakes with high accuracy and
reliability. Deep learning, particularly Convolutional Neural Networks (CNNs) and Recurrent
Neural Networks (RNNs), has emerged as a powerful tool for this purpose. CNNs are
effective in capturing spatial inconsistencies and artifacts within video frames, while RNNs
or Long Short-Term Memory (LSTM) networks can analyze temporal patterns across frames
to identify subtle anomalies in motion and expression.
This paper investigates the effectiveness of deep learning techniques in detecting deepfake
videos. We focus on extracting facial regions, preprocessing video frames, and training deep
neural networks using large-scale datasets such as FaceForensics++ and the DeepFake
Detection Challenge (DFDC). The objective is to develop a robust detection model capable of
generalizing across different types of deepfakes and varying video qualities. By addressing
this challenge, we aim to contribute to the broader goal of ensuring media authenticity and
protecting users from the harmful effects of synthetic media.
 LITERATURE REVIEW
The increasing sophistication of deepfake generation techniques has sparked significant
research interest in the development of reliable detection methods. Early approaches relied
heavily on hand-crafted features and forensic techniques, such as detecting inconsistencies in
lighting, shadows, and eye blinking patterns. However, these traditional methods often fail
when faced with high-quality and adversarially optimized deepfakes.
With the rise of deep learning, more robust and scalable solutions have emerged. Afchar et al.
(2018) proposed MesoNet, a shallow CNN model that captures mesoscopic features for
deepfake detection, demonstrating promising performance with low computational cost.
Similarly, Rossler et al. (2019) introduced FaceForensics++, a large-scale dataset used to
train and benchmark deepfake detection models. Their study highlighted the effectiveness of
deep CNNs in identifying facial artifacts and inconsistencies in synthetic videos.
Other researchers have explored temporal features for improved performance. Sabir et al.
(2019) used a CNN-LSTM architecture to capture both spatial and temporal inconsistencies,
which improved detection accuracy on dynamic and compressed videos. In addition, Zhou et
al. (2021) proposed Two-Branch Recurrent Networks, separating spatial and motion cues
to detect manipulation in video sequences.
Recent studies have also investigated attention mechanisms and transformer-based models to
enhance detection capabilities. Works like XceptionNet, based on depthwise separable
convolutions, have shown superior performance on datasets like DFDC and Celeb-DF.
Despite progress, challenges such as generalization across different deepfake techniques,
robustness to video compression, and adversarial attacks remain open areas of research.
Overall, the literature indicates that deep learning offers powerful tools for deepfake
detection, but ongoing advancements in generative models demand continuous innovation
and evaluation of detection techniques.
 PROBLEM STATEMENT

The advancement of deep learning and generative models, particularly

Generative Adversarial Networks (GANs), has enabled the creation of
hyper-realistic synthetic videos known as deepfakes. These videos
manipulate facial expressions, movements, and even voices in a way
that is often indistinguishable from genuine recordings. While such
technology has legitimate applications in entertainment, accessibility,
and education, its malicious use poses significant threats to individual
privacy, public trust, political stability, and cybersecurity.

Deepfakes have been increasingly used to spread misinformation,

conduct fraud, and manipulate public opinion. The rapid evolution and
accessibility of deepfake creation tools make it easier for individuals with
minimal technical knowledge to generate deceptive content. This creates
a critical need for reliable, automated detection systems that can
distinguish deepfake videos from authentic ones with high accuracy and
efficiency.

Traditional video forensics methods, which rely on manual inspection or

low-level inconsistencies, are no longer sufficient to detect sophisticated
deepfakes. These methods often fail under conditions of compression,
noise, or adversarial modifications. Therefore, there is an urgent need to
develop intelligent, deep learning-based detection techniques capable of
analyzing both spatial features (e.g., facial artifacts, unnatural textures)
and temporal patterns (e.g., inconsistent movements, lip-sync issues) in
video data.
The primary challenge lies in developing a detection system that is
robust, generalizable across multiple deepfake generation techniques,
and efficient enough for real-time or large-scale deployment. This
research aims to address these challenges by leveraging deep learning
architectures such as Convolutional Neural Networks (CNNs) and
Recurrent Neural Networks (RNNs) to identify inconsistencies in
deepfake videos. The goal is to contribute to the growing body of work
on digital media forensics and provide a practical solution to combat the
growing threat of synthetic media manipulation.

OBJECTIVES

The main objective of this study is to develop an effective deep learning-

based system for detecting deepfake videos. The specific objectives are
as follows:

1. To understand and analyze the techniques used in generating

deepfake videos, particularly those based on Generative
Adversarial Networks (GANs) and autoencoders.

2. To study existing deepfake detection methods and identify

their limitations, especially in terms of accuracy, generalization,
and robustness against adversarial attacks and video compression.

3. To design and implement a deep learning model, combining

Convolutional Neural Networks (CNNs) for spatial feature
extraction and Recurrent Neural Networks (RNNs) or Long Short-
Term Memory (LSTM) networks for temporal pattern analysis in
video sequences.
 4. To collect and preprocess relevant datasets, such as
FaceForensics++, DFDC (DeepFake Detection Challenge), or
Celeb-DF, for training and evaluating the proposed model.

5. To train the deep learning model using labeled real and fake
video data, optimizing it for high accuracy and low false
positive/false negative rates.

6. To evaluate the performance of the developed model using

standard metrics such as accuracy, precision, recall, F1-score, and
AUC-ROC.

7. To test the generalization capability of the model across

different deepfake generation methods and various levels of video
quality and compression.

PROPOSED METHODOLOGY
The proposed methodology aims to develop a robust and efficient deep learning-based system
for detecting deepfake videos by analyzing both spatial and temporal inconsistencies. The
approach can be divided into the following key stages:
1. Dataset Collection and Preparation
 Datasets: Publicly available datasets such as FaceForensics++, Celeb-DF, and
DeepFake Detection Challenge (DFDC) will be used.
 Preprocessing: Video frames will be extracted at consistent intervals. Face detection
and alignment techniques (e.g., MTCNN or dlib) will be applied to crop facial regions
from each frame.
 Normalization: Frames will be resized, normalized, and converted into a suitable
format for input into deep learning models.
2. Feature Extraction
 Spatial Features: Convolutional Neural Networks (CNNs) such as XceptionNet,
ResNet, or EfficientNet will be employed to extract spatial features from each video
 frame, identifying anomalies like blending artifacts, unnatural facial textures, or
inconsistencies in lighting.
3. Temporal Analysis
 Sequence Modeling: To capture temporal inconsistencies, a Recurrent Neural
Network (RNN) or Long Short-Term Memory (LSTM) network will be used on the
sequence of spatial features extracted from consecutive frames.
 This enables detection of subtle changes in motion patterns, lip synchronization, and
facial movements that are difficult to identify in isolated frames.
4. Model Training and Validation
 The combined CNN-LSTM architecture will be trained using labeled real and fake
videos.
 Training will be optimized using loss functions like binary cross-entropy and
evaluated using metrics such as accuracy, precision, recall, and F1-score.
5. Testing and Performance Evaluation
 The model will be tested on unseen videos from different datasets to assess its
generalization and robustness under various conditions, including compressed or low-
resolution videos.

EXPECTED OUT COME

1. High Detection Accuracy
The proposed model, combining CNN and RNN/LSTM architectures, is expected to
achieve high accuracy in distinguishing real videos from deepfakes, even when the
fake videos are generated using advanced and recently developed techniques.
2. Robustness to Variations
The system should demonstrate robustness against various video conditions, including
different resolutions, compression levels, lighting variations, and face angles, making
it suitable for real-world applications.
3. Generalization Across Datasets
The model is expected to generalize well across multiple datasets (e.g.,
FaceForensics++, DFDC, Celeb-DF), showing its ability to detect deepfakes created
using different methods, not just the ones seen during training.
4. Temporal Consistency Analysis
By integrating temporal analysis (via RNN/LSTM), the system should be capable of
 detecting inconsistencies in facial movements, expressions, or lip-syncing across
video frames—an area where many single-frame detection models fall short.
5. Scalability and Practical Application
The resulting system should be efficient enough for potential real-time
implementation or integration into media platforms, content moderation systems, or
digital forensics tools.
6. Contribution to Research
The study will provide insights into the effectiveness of different deep learning
models for deepfake detection and may serve as a foundation for future improvements
or hybrid detection approaches.

Existing Disadvantage & Proposed Advantage

Existing Disadvantages
1. Limited Accuracy on High-Quality Deepfakes
Traditional detection methods and some older deep learning models struggle to detect
high-resolution, well-crafted deepfakes that contain fewer visual artifacts.
2. Poor Generalization
Many existing models are trained on a specific type of deepfake generation technique
and perform poorly when tested on videos produced using different or newer
methods.
3. Sensitivity to Compression and Noise
Detection performance significantly drops on videos that are compressed, blurred, or
altered, which is common on social media platforms.
4. Lack of Temporal Analysis
Many models analyze only individual frames and fail to capture temporal
inconsistencies such as unnatural movements, poor lip-syncing, or inconsistent
 expressions.
5. High Computational Cost
Some existing systems are computationally expensive and not optimized for real-time
or large-scale deployment.
Proposed Advantages
1. Improved Detection Accuracy
By combining Convolutional Neural Networks (CNNs) for spatial feature extraction
and Recurrent Neural Networks (RNNs)/LSTM for temporal analysis, the proposed
model aims to achieve higher accuracy in detecting both low- and high-quality
deepfakes.
2. Better Generalization
The model will be trained on diverse datasets, enhancing its ability to detect
deepfakes generated by various techniques and tools, ensuring broader applicability.
3. Robust to Compression and Quality Loss
Advanced preprocessing and model training will help the system maintain
performance even under compressed, low-resolution, or noisy video conditions.
4. Temporal Consistency Detection
The inclusion of RNN/LSTM enables the system to analyze frame-to-frame
continuity, identifying manipulation artifacts that appear only over time.
5. Optimized for Scalability
The proposed model will be designed to be computationally efficient, making it
suitable for integration into real-time applications and media verification tools.