A Long-Range Attention Framework for Deepfake Detection Using Deep Learning

Abstract

The complexity of the methods used to create deepfakes is the reason why digital media authentication is requiring an extended period of time. Normal methods are incapable of detecting minute variations in texture, pace, and facial expressions. This research demonstrates a novel approach to identifying deepfakes through the use of long-distance attention. Time and location data that are stored in video frames are the primary focus. The model is trained to identify discrepancies in time and space through the utilization of two modules in the proposed design. Consequently, the model is more adaptable and capable of accommodating minor modifications. The network is able to identify and classify false positives and declines by utilizing long-distance attention, which connects data at the frame level. Celeb-DF, DFDC, and Face Forensics++ are benchmark datasets that demonstrate the effectiveness of these methods in comparison to CNN and LSTM. In the long term, this technology has the potential to supplant scalable, precise deepfake detection methods.