3D Morphable Models Meet Surface Frames for Generalizable and Robust Deepfake Detection

Giovanni Affatato, Andrea Ciamarra, Edoardo Daniele Cannas, Sara Mandelli, Benedetta Tondi, Roberto Caldelli, Paolo Bestagini

Abstract: With the rapid advancements in AI-generated imagery, particularly diffusion-based models, detecting synthetic human faces has become increasingly challenging. In this paper, we introduce a synthetic face detection framework that leverages two complementary features: (i) UV textures extracted using 3D Morphable Models (3DMM) and (ii) surface frames capturing geometric structures. These modalities are fused using both featurelevel and score-level fusion strategies to enhance generalization to unseen generators and robustness against post-processing operations. Experimental evaluations on diverse datasets demonstrate that our proposed method outperforms single-modality and CLIP-based approaches and provides improved generalization across different diffusion generative models, as well as improved robustness against common and strong processing operations.

External IDs:dblp:conf/eusipco/AffatatoCCMTCB25