Deepfakes, media manipulated using deep learning techniques, pose a growing threat to the integrity of digital content. These AI-generated forgeries are becoming increasingly sophisticated, making them difficult to detect. Traditional detection methods often lag behind the rapid evolution of deepfake techniques and are hampered by the limited variety of training data, making it hard for them to generalize effectively to new types of deepfakes. This thesis introduces a novel deepfake detection approach that combines meta-learning for domain generalization (MLDG) with self blended images (SBI) to address this challenge. MLDG, inspired by meta-learning principles, aims to improve the model’s adaptability to new manipulation techniques by simulating domain shifts during training. The model learns from various source domains representing different deepfake generation methods. Additionally, SBIs, synthetic images created by blending real and manipulated faces, are incorporated to further diversify the training data and promote the learning of features that generalize across domains. This thesis focuses on detecting image-based deepfakes using the Face Forensics++ dataset, a benchmark collection of real and manipulated videos, specifically designed for deepfake detection research. The proposed method is evaluated with a leave-one-out cross-validation scheme on this dataset, where each deepfake generation technique is used as a test case while the others are used for training. The results consistently show that MLDG, when enhanced with SBIs, outperforms the standard Empirical Risk Minimization (ERM) method, demonstrating its effectiveness in generalizing to unseen manipulation techniques. The research offers a practical solution for deepfake detection, highlighting how MLDG and SBI augmentation can create more effective and adaptable detection systems. The findings emphasize the need for models that can adapt to evolving deepfake techniques to protect the integrity of digital media.
-
International Journal of Trend in Scientific Research and Development - IJTSRD having
online ISSN 2456-6470. IJTSRD is a leading Open Access, Peer-Reviewed International
Journal which provides rapid publication of your research articles and aims to promote
the theory and practice along with knowledge sharing between researchers, developers,
engineers, students, and practitioners working in and around the world in many areas
like Sciences, Technology, Innovation, Engineering, Agriculture, Management and
many more and it is recommended by all Universities, review articles and short communications
in all subjects. IJTSRD running an International Journal who are proving quality
publication of peer reviewed and refereed international journals from diverse fields
that emphasizes new research, development and their applications. IJTSRD provides
an online access to exchange your research work, technical notes & surveying results
among professionals throughout the world in e-journals. IJTSRD is a fastest growing
and dynamic professional organization. The aim of this organization is to provide
access not only to world class research resources, but through its professionals
aim to bring in a significant transformation in the real of open access journals
and online publishing.