Deepfake Detection Model Based on VGGFace with Head Pose Estimation Technique

Rapid developments in deepfake technology have produced hyper-realistic fake media such as images, audio, and video. Especially, the fabricated videos, which have gained a large widespread in social media sites. Because of the great harm inflicted by these videos, many researchers have increased their efforts to find a reliable method to identify and distinguish these fake videos from the actual ones. In this work, we suggested a new model to detect fake videos, based on two major techniques. First, the VGGFace model was used to extract the most important facial features combined with, second: the estimation of the head pose angle that represents the relative orientation of the human face in video frames. All these calculations are done on human faces detected and cropped from video frames, where 10, 20, and 30 frames were extracted from each video. FF++ dataset was used to train and test the model, which produced a max test accuracy of 0.885. The code was written using Python version 3.9.