Deepfake Detection and Analysis Using Fusion Model

Abstract

In today's digital era, deepfake technology presents both innovation and peril, with hyper-realistic synthetic media capable of widespread deception. This research delves into deepfake detection, focusing two deepfake detection models, namely, the VIT image classifier and Meso4 model. Utilizing convolutional neural networks, VIT analyzes images, while Meso4 scrutinizes at a mesoscopic level. A comparative analysis evaluates their effectiveness in discerning authentic from manipulated content. Using the Open Forensics dataset and self-generated content, VIT achieves remarkable accuracy, while Meso4 encounters challenges, such as limited generalization, task-dependent accuracy levels, etc. Therefore, in the proposed work additional features like, Eye movement error detection, Skin texture inconsistency detection and Facial feature inconsistency detection are integrated into a customized model, which results significantly in augmenting accuracy and computation speed compared to above compared models. This research work emphasizes the need for advancing unbiased deepfake detection methods, urging vigilance in safeguarding privacy and security amidst pervasive digital deception.