Authors: P.Selvaraj, A Joshua Issac, Dr.S.Shanmuga, M.Bharathi
Abstract: The increasing advancement of generation of deepfake techniques – especially manipulations involving face-swapping has brought up major concerns related to integrity of online media, data privacy and societal trust. The computer generated videos, created using advanced models can easily replace an individual face with another often fool regular detection tools because changes in lighting, skin tone, facial expressions are so small and hard to notice. Although many AI-based methods have been developed to spot deep fake, most current models still struggle because they only look at single images , don't consider changes over time or require too much computing power. This research proposes a hybrid deepfake detection framework that leverages the strengths of Convolutional Neural Networks (CNNs) for robust spatial feature extraction and Vision Transformers (ViTs) for capturing temporal and contextual relationships across video frames. The CNN part looks for small changes and edits in the face, while the Vision Transformer looks at a series of frames to catch unusual expressions , movements and facial tone. Together, this combination aims to overcome the challenges posed by diverse and highly realistic face-swap techniques. The system is trained and tested on known datasets like FaceForensics++ and DFDC-Preview, providing a complete way to detect face-swap deep fake. By improving on current methods and looking at both the details in each frame and changes over time, this study helps create a stronger and more flexible deepfake detection system that can handle new and growing threats in visual content.
DOI: http://doi.org/
