Video forensics based on expression dynamics

Digital graphics tools are nowadays capable of rendering highly photorealistic imagery, which easily puzzle our perception of reality. This poses serious ethical and legal issues, which in turn create the need for further technologies able to ensure the trustworthiness of digital media as a true representation of reality, especially when depicting humans. In this work, we propose a novel forensic technique to tackle the problem of distinguishing computer generated (CG) from real humans in videos. It exploits the temporal information inherent of a video sequence by analyzing the spatio-temporal appearance of facial expressions in both CG and real humans. Even if rendering facial expression has reached outstanding performances, CG face appearance over time still presents some underlying mechanical properties that greatly differ from the natural muscle movements of real humans. We build an efficient classifier on a set of features describing facial dynamics and spatio-temporal changes during smiling to distinguish CG from human faces. Experimental results demonstrate the effectiveness of the proposed approach.