A DEEP LEARNING APPROACH FOR DEEPFAKE DETECTION USING LONG-RANGE ATTENTION MODELS

Abstract

The intricacy of deepfake production methods is slowing digital media authentication. Traditional approaches struggle to identify tiny changes in texture, cadence, and facial emotions. This paper uses long-distance attention to recognize deepfakes in a novel way. The focus is on temporal and spatial information in video frames. The suggested design trains the model to identify space and time mismatches using two modules. This makes the model more sensitive to small changes. By adding long-distance attention, the network can correlate frame-level data and identify and classify fakes. Benchmark datasets like Celeb-DF, DFDC, and Face Forensics++ show they are more accurate than CNN and LSTM techniques. This method may replace scalable, realistic deepfake detection techniques.