DEEPFAKESHIELD: ENHANCED VIDEO AUTHENTICITY DETECTION VIA CONVOLUTIONAL VISION TRANSFORMER

Abstract

The way the deep learning algorithms have quickly evolved to be able to produce and recreate the extremely realistic videos, also known as Deepfakes, has caused considerable alarm about the misuse of such tools. Innovative techniques of deep learning are now capable of producing synthetic faces, swapping faces across people, changing facial expressions, modifying gender traits and manipulating facial features with incredible accuracy. Although virtual reality, digital content creation, and entertainment are lawful uses of these technologies, they have serious risks when they are used in bad intentions like misinformation, stealing of identities, and internet fraud. This study presents an effort to introduce a hybrid model combining EfficientNet with the Swin Transformer to detect Deepfakes effectively. EfficientNet is used to extract fine-grained spatial features, whereas the Swin Transformer uses hierarchical attention to capture long-range dependencies to classify authenticity. The proposed framework was also trained and tested on FaceForensics data, with the accuracy of 93.2, AUC of 0.94, and the loss rate of 0.28. A combination of EfficientNet convolutional representations and the Swin Transformer attention-based system proves to be at a better level of detecting manipulated content, and this points to the model being able to distinguish well between the manipulated and veritable videos.