An integrated spatiotemporal-based methodology for deepfake detection

Abstract Rapid advances in deep learning models have made it easier for public and crackers to generate hyper-realistic deepfake videos which faces are swapped. Such may constitute a significant threat the world if they misused blackmail figures deceive systems of face recognition. As result, distinguishing these fake from real ones has become fundamental. This paper introduces new video detection method. You Only Look Once (YOLO) detector is used detect frames. A proposed hybrid method based on proposing two different feature extraction methods applied faces. The first method, Convolution Neural Network (CNN), Histogram Oriented Gradient (HOG) second one an ameliorated XceptionNet CNN. extracted sets features merged together fed as input sequence Gated Recurrent Units (GRUs) extract spatial temporal then individuate authenticity videos. trained CelebDF-FaceForencics++ (c23) dataset evaluated CelebDF test set. experimental results analysis confirm superiority suggested over state-of-the-art methods.