BANGALORE INSTITUTE OF TECHNOLOGY
K.R. Road, V.V.Puram, Bengaluru-560 004
DEPARTMENT OF ARTIFICIAL INTELLIGENCE & MACHINE LEARNING
Major Project Work Synopsis
VI – SEM 2023-2024
Mini Project Group:
Sl.
USN Name Email Id Phone No Signature
No.
1BI21AI017 Divye Jain [email protected] 7463024638
1
2 1BI21AI040 Raghav Joshi [email protected] 9339157755
3
1BI21AI052 Yash Bubna
[email protected] 7260871715
2
Major Project Details:
Advanced Deepfake Detection Using Hybrid Deep Learning Models: ResNext and
Title
LSTM for Spatial-Temporal Analysis
Domain Artificial Intelligence
Location Bengaluru
For office use only:
Group ID
Guide Prof. Sahana M.
Status
Signature of the Major Project Co-Ordinator
Advanced Deepfake Detection Using Hybrid Deep Learning Models:
ResNext and LSTM for Spatial-Temporal Analysis
1. Introduction
With the rapid advancements in artificial intelligence, particularly in the domain of generative
models, deepfake technology has emerged as a significant concern in today's digital landscape. Deepfakes
refer to synthetic media in which a person's face, voice, or actions are replaced or manipulated using AI
algorithms, typically through deep learning techniques. While the underlying technology has potential
applications in entertainment, education, and content creation, it has also been misused for malicious
purposes such as spreading misinformation, committing fraud, and invading privacy.
Deepfakes have become increasingly sophisticated, making it difficult for the average person to
discern fake videos or images from real ones. The ability to seamlessly manipulate video content poses a
major threat to media integrity, public safety, and personal reputations. This has triggered a global call for
advanced detection systems that can accurately and efficiently identify deepfake content before it is used to
cause harm.
The primary challenge in detecting deepfakes stems from the fact that many of these manipulations
are subtle and go undetected by traditional video verification methods. Deepfake algorithms such as
Generative Adversarial Networks (GANs) have become adept at generating realistic human faces, which
makes detection even more difficult. As a result, the development of an AI-powered detection system has
become essential in mitigating the risks posed by deepfakes.
This project focuses on building a deep learning-based deepfake detection system that leverages
cutting-edge models like ResNext for spatial analysis and Long Short-Term Memory (LSTM) networks for
temporal sequence learning. By analyzing both the pixel-level details within individual video frames and the
temporal dynamics between consecutive frames, the system aims to effectively detect any irregularities or
manipulations that indicate the presence of deepfakes.
The proposed solution is designed to detect various types of deepfake manipulations, including facial
expressions, lip-syncing, and voice alterations, across different media formats. The integration of
Grad-CAM for model interpretability further ensures that the system is transparent and provides visual
feedback on what areas of the image the model focuses on when making predictions, improving trust in the
detection process.
As deepfake technologies evolve, so must our methods of detecting them. This project aims to be at
the forefront of these efforts, combining the power of deep learning, computer vision, and natural language
processing to create a robust system capable of protecting the authenticity of digital media.
2. Problem Statement
“To develop an efficient deep learning-based system for detecting deepfake videos, using neural network
architectures like ResNext and LSTM, with an emphasis on improving accuracy, precision, and recall.”
3. Objectives
● Develop a model using ResNext and LSTM to detect deepfake videos by analyzing both spatial
and temporal features.
● Train the model on real and deepfake video datasets, ensuring diverse data to enhance model
robustness.
● Fine-tune hyperparameters and employ data augmentation techniques to avoid overfitting.
● Implement methods like Grad-CAM for model interpretability and explainability.
● Compare the performance of the model with existing methods to ensure superior accuracy and
detection speed.
● Create a user-friendly interface or API to integrate the deepfake detection mechanism into media
applications.
