You can find here the implementation of the network architecture and the dataset used in our paper on digital forensics. It was accepted at the WIFS 2018 conference.
We present a method to automatically detect face tampering in videos. We particularly focus on two recent approaches used to generate hyper-realistic forged videos: deepfake and face2face. Traditional image forensics techniques are usually not well suited to videos due to their compression that strongly degrades the data. Thus, we follow a deep learning approach and build two networks, both with a low number of layers to focus on mesoscopic properties of the image. We evaluate those fast networks on both an existing dataset and a dataset we have constituted from online videos. Our tests demonstrate a successful detection for more than 98\% for deepfake and 95\% for face2face.
If you want to use the complete pipeline with the face extraction from the videos, you will also need the following librairies :
| Set | Size of the forged image class | Size of real image class |
|---|---|---|
| Training | 5111 | 7250 |
| Validation | 2889 | 4259 |
You can find the pretrained weight in the weights folder. The _DF extension correspond to a model trained to classify deepfake-generated images and the _F2F to Face2Face-generated images.
Darius Afchar - École des Ponts Paristech | École Normale Supérieure (France)
Vincent Nozick - Website
Afchar, D., Nozick, V., Yamagishi, J., & Echizen, I. (2018, September). MesoNet: a Compact Facial Video Forgery Detection Network. In IEEE Workshop on Information Forensics and Security, WIFS 2018.
This research was carried out while the authors stayed at the National Institute of Informatics, Japan.
