Fighting Deepfake by Residual Noise Using Convolutional Neural Networks

In the last few years, the easy access to images and videos shared online have been continuously increased. The generative adversarial networks using deep learning leads to create very realistic deepfake videos by playing with the digital content of images and videos. The spread of such deepfake videos on social media networks urged the international community to consider seriously its danger and accordingly encouraged the researchers around the world to develop powerful deepfake detection methods. Many approaches are available in the recent literature. In this paper, the proposed approach is based on exploiting the residual noise which is the difference between original image and its denoised version. The study of residual noise has shown effectiveness in deep-fake detection with regards to its distinctive and discriminative features which can be effectively captured by convolutional neural networks with transfer learning. The performance of our approach is evaluated on two datasets: low-resolution video sequences of the FaceForensics++ and high-resolution videos from Kaggle Deepfake Detection challenge (DFDC). The obtained results show relevant accuracy in comparison with other competitive methods.