Detecting Deepfake Videos in Data Scarcity Conditions by Means of Video Coding Features

The most powerful deepfake detection methods developed so far are based on deep learning, requiring that large amounts of training data representative of the specific task are available to the trainer. In this paper, we propose a feature-based method for video deepfake detection that can work in data scarcity conditions, that is, when only very few examples are available to the forensic analyst. The proposed method is based on video coding analysis and relies on a simple footprint obtained from the motion prediction modes in the video sequence. The footprint is extracted from video sequences and used to train a simple linear Support Vector Machine classifier. The effectiveness of the proposed method is validated experimentally on three different datasets, namely, a synthetic street video dataset and two datasets of Deepfake face videos.