[1N3-GS-10] Vision, speech: GAN

[1N3-GS-10-02] DeepFake video detection using time-series information

〇Satoshi Kageyama¹, Masahiro Suzuki¹, Yutaka Matsuo¹ (1. The University Of Tokyo)

Keywords:Image generation, DeepFake, GAN, VAE

In recent years, an image generation method using deep learning such as GAN or VAE can generate a high-definition image that does not actually exist. Also, by applying such an image generation technique, it is possible to convert an arbitrary image. By applying these technologies and converting face images in a moving image, it is possible to generate a FAKE videos that cannot be distinguished from reality. FAKE videos generated by manipulating facial images have spread on news sites and social networks, and their use as politics and pornography has become a social issue. Therefore, it is very important to develop a method for detecting whether a Real video or generated videos. Many of the detection methods focus on the features of the face image in each frame unit in the FAKE video, but the identification of a single face image has become difficult due to the sophistication of the generation method. Therefore, our proposed method focuses on the relationship between faces in each frame in the FAKE video, and identifies them based on the time evolution information of the faces. We verify the data that was difficult to identify with the existing method, and show the validity of the proposed method.