Temporal Deepfake Detection using CNN with Spatio-Temporal Features

With the advancement of Artificial Intelligence (AI), facial recognition has become a crucial biometric feature. Deepfake technology leverages AI and can create hyper-realistic digitally manipulated images and videos of people appearing to say or do things that never occurred. The emergence of Generative Adversarial Networks (GANs) in 2014 has further enabled the creation of fake visual content. This technology has diverse applications, such as in the film industry, where it allows for video recreation without reshooting, creating awareness videos, restoring the voices of those who have lost them, and updating movie scenes at low cost. However, video-based manipulations pose significant challenges to detection systems. While most deepfake detectors focus on spatial anomalies in individual frames, temporal inconsistencies across frames can offer crucial clues. This paper presents a novel approach to video-based deepfake detection using Dense Swin Transformer, which leverages spatio-temporal feature extraction. Our proposed method, trained on the DFDC dataset, demonstrates improved accuracy in detecting deepfakes, achieving 98.25% accuracy with low computational cost.