A Method for Detecting DeepFake Videos Using Angular Information of Faces

DOI

Bibliographic Information

Other Title
  • 顔の角度情報を用いたDeepFake動画の検出手法の提案

Search this article

Description

Recently, image generation methods based on deep learning, such as GAN and VAE, have made it possible to generate high-definition images of people and landscapes that do not actually exist. Such image generation methods can also be applied to transform arbitrary images. By applying these techniques to transform the face images in a video, it is possible to generate a fake video (Deep Fake video) that is indistinguishable from reality. Fake videos generated by manipulating facial images spread through news sites and social networks, and their use for political purposes or as pornography has become a social problem. Therefore, it is important to develop a method to detect whether a video is an artificially generated FAKE video or not. Existing studies have proposed detection methods based on the assumption that generation does not fail in all frames of a video, but in actual FAKE videos, generation fails in scenes where the face is facing directly to the side. In this proposal, we focus on this problem and propose a new perspective: the use of face angle information. The proposed method improves the accuracy by up to 3.4 percent in terms of AUC.

Journal

Keywords

Details 詳細情報について

Report a problem

Back to top