International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 12 Issue: 06 | Jun 2025
p-ISSN: 2395-0072
www.irjet.net
A Hybrid Approach For Identification Of DeepFake Videos And Images Nakka Neha Sri1, Gade Manogna2, K. Naga Shailaja3 1,2B. Tech Student, Dept. of Electronics and Computer Engineering, Sreenidhi Institute of Science and Techology,
Telangana, India
3 Assistant Professor, Dept. of Electronics and Computer Engineering, Sreenidhi Institute of Science and Technology,
Telangana, India ---------------------------------------------------------------------***--------------------------------------------------------------------capable of capturing subtle manipulation artifacts across diverse video content.
Abstract - This paper presents a hybrid deep learning
model for deepfake video detection, combining the strengths of Convolutional Neural Networks (CNNs), Graph Neural Networks (GNNs), and Transformer architectures. The system extracts video frames, applies extensive data augmentation, and leverages MobileNetV3 for efficient feature extraction. Subsequently, a GNN module constructs a graph representation of the extracted features, capturing spatial relationships, while a Transformer module models long-range dependencies. The model is trained and evaluated on the Deep Fake Detection dataset, demonstrating robust performance through comprehensive metrics including accuracy, F1 score, and AUC. The hybrid architecture effectively leverages complementary feature learning paradigms, improving detection accuracy and generalization.
This document presents a comprehensive analysis of a novel deepfake detection system that addresses these challenges by employing a hybrid neural network architecture. This approach leverages the complementary strengths of convolutional neural networks (CNNs), graph neural networks (GNNs), and transformer models, enabling the system to capture a multifaceted representation of video data. Specifically, the system utilizes MobileNetV3, a lightweight and efficient CNN, for initial feature extraction, transforming video frames into high- dimensional feature maps. Subsequently, a GNN component constructs graph representations of these feature maps and capturing spatial relationships and also structural dependencies between image regions. Finally, a transformer model processes the graph-level representations, enabling the system to model long-range dependencies and global context within the video sequence. The rationale behind this hybrid architecture stems from the recognition that deepfake manipulation often introduces subtle artifacts that manifest across different levels of representation. CNNs excel at extracting local spatial features, while GNNs are adept at capturing structural relationships and interdependencies between image regions [2]. Transformer models, on the other hand, are highly effective at modelling long- range dependencies and contextual information, enabling the system to identify subtle inconsistencies that may span multiple frames or regions of the video.
Keywords: Deepfake Detection, Graph Neural Network, Transformers, Convolutional Neural Networks, Video Analysis, MobileNetV3, Data Augmentation
1.INTRODUCTION The proliferation of deepfake technology has emerged as a significant societal challenge, posing a serious threat to information integrity and public trust. These manipulated videos, generated using sophisticated artificial intelligence techniques, can convincingly alter or fabricate visual and auditory content, making it increasingly difficult to distinguish between authentic and synthetic media. The potential for misuse is vast, ranging from political disinformation campaigns and character defamation to financial fraud and the erosion of journalistic credibility [4] .As deepfake technology becomes more accessible and refined, the urgency to develop robust detection systems intensifies. Traditional methods for deepfake detection often rely on identifying inconsistencies in facial features or temporal anomalies within video sequences[9]. However, the rapid advancements in generative adversarial networks (GANs) and other deep learning techniques have enabled the creation of highly realistic deep fakes that can evade these conventional detection strategies. Consequently, there is a growing need for more sophisticated and adaptable detection mechanisms
© 2025, IRJET
|
Impact Factor value: 8.315
In summary, this hybrid neural network approach represents a significant advancement in deepfake detection, offering a robust and adaptable solution to combat the growing threat of manipulated video content [10]. By leveraging the complementary strengths of different neural network architectures and incorporating advanced training and evaluation techniques, this system aims to contribute to the development of effective tools for safeguarding information integrity and public trust in the digital age.
|
ISO 9001:2008 Certified Journal
|
Page 725