ATTENTION BASED DEEPFAKE DETECTION LEVERAGING LONG RANGE DEPENDENCIES

Abstract

This investigation demonstrates an improved method of employing deep learning that employs attention processes to identify long-range connections between temporal and visual data in order to identify deepfakes. The proposed attention-based design, which integrates transformer models and self-attention, more effectively learns global contextual relationships across frames than traditional convolutional approaches, which frequently struggle to identify nuanced and widely dispersed changes. The system improves its ability to identify intricate and high-quality deepfakes by concentrating on critical face features and simulating the interaction between frames. The detection accuracy, robustness, and generalization have improved across a variety of datasets as a consequence of the experiments. This demonstrates the significance of long-range dependency modeling in addressing the increasing issues that advance synthetic media generate.