GrDT: Towards Robust Deepfake Detection using Geometric Representation Distribution and Texture

Hongyang Xie, Hongyang He, Boyang Fu, Victor Sanchez; Proceedings of the Winter Conference on Applications of Computer Vision (WACV) Workshops, 2025, pp. 734-744

In recent years deepfake images and videos have rapidly spread across social media platforms. This poses significant threats to public privacy property and safety. Several detection methods have been proposed with the most common approaches being those based on deep learning and on the analysis of bimetric signals. However these methods generally suffer from poor generalization capabilities struggle to detect high-quality deepfakes and depend on high-resolution training data. Based on these observations we propose a detection method based on a Graph Attention Network (GAT) and biometric features referred to as GrDT. The core idea of GrDT is to identify deepfake face images by leveraging facial texture representations and the geometric relationships of key facial points. Cross-validation on the DF40 and ForGeryNet datasets shows that GrDT outperforms other methods by up to 1.8% in terms of AP and by 0.8% in terms of AUC. The code is available at https://github.com/SIPLab24/GrDT.