Predicting Heart Rate Variations of Deepfake Videos using Neural ODE

Deepfake is a technique used to manipulate videos using computer code. It involves replacing the face of a person in a video with the face of another person. The automation of video manipulation means that deepfakes are becoming more prevalent and easier to implement. This can be credited to the emergence of apps like FaceApp and FakeApp, which allow users to create their own deepfake videos using their smartphones. It has hence become essential to detect fake videos, to avoid the spread of false information. A recent study shows that the heart rate of fake videos can be used to distinguish original and fake videos. In the study presented, we obtained the heart rate of original videos and trained the state-of-the-art Neural Ordinary Differential Equations (Neural-ODE) model. We then created deepfake videos using commercial software. The average loss obtained for ten original videos is 0.010927, and ten donor videos are 0.010041. The trained Neural-ODE was able to predict the heart rate of our 10 deepfake videos generated using commercial software and 320 deepfake videos of deepfakeTIMI database. To best of our knowledge, this is the first attempt to train a Neural-ODE on original videos to predict the heart rate of fake videos.