Year: 2024
Type: article
Abstract: In the rapidly evolving landscape of deepfake technology, which poses substantial threats to privacy and security, this research conducts a comparative analysis between two advanced detection methods: The VIT image classifier employing convolutional neural networks and the Meso4 method to analyse images at mesoscopic level.Motivated by the pervasive impact of deepfakes on individuals, particularly celebrities and politicians, the study aims to discern the effectiveness of these methods in distinguishing authentic from manipulated images.The methodological approach prioritizes practical implementation, utilizing the DFDC (Deep Fake Detection Challenge by Kaggle) dataset for ViT and a combination of self made and Face2Face dataset for Meso4 model.The DFDC dataset is a massive 260 GB dataset comprising over 1 million images, both real and fake, while the combined dataset for MESO consists of approximately 400 videos made by the author and over a thousand videos from the Face2Face dataset, encompassing real and fake instances.The ViT model exhibits exceptional proficiency, achieving a 99.9% accuracy within its dataset and an 85% accuracy for subtly deceptive images.Conversely, the Meso4 model faces challenges, resulting in a 60% accuracy and displaying biases, particularly towards female subjects due to its heavy trainings on adult content.In conclusion, the VIT image processor emerges as a superior and user-friendly method, despite its substantial data footprint, showcasing superior efficacy, especially in real-world scenarios.The findings emphasize the importance of advancing robust and unbiased deep fake detection methods, acknowledging considerations such as model size in the context of application requirements.
Author: Mr.S. Govindaraj
Institution: Velalar College of Engineering and Technology
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4405983531