3DDGD: 3D Deepfake Generation and Detection Using 3D Face Meshes

Abstract

3D face technology is revolutionizing various fields by providing superior security and realism compared with 2D methods. In biometric authentication, 3D facial features serve as unique, hard-to-forge identifiers, improving accuracy in facial recognition for border control and criminal identification. Additionally, 3D avatars enhance virtual interactions. In this study, we aimed to strengthen 3D facial biometric systems against deepfakes. Key contributions include proving the superior protection of 3D faces over 2D ones, creating a dataset of real and fake 3D faces, and developing advanced models for accurate 3D deepfake detection. We evaluated our models for generalization to other datasets and stability when changing training data. Our experiments used the mesh multi-layer perceptron model for deepfake detection along with self-attention mechanisms and the newly introduced TabTransformer model. Results indicate that 3D face meshes greatly improve security by distinguishing real faces from deepfakes. Future work will focus on enhancing detection tools and integrating geometric features with facial textures for more accurate 3D deepfake detection. The dataset and models are publicly available on GitHub, excluding licensed elements: https://github.com/hichemfelouat/3DDGD