A Generalizable Face Security Detection Method via Unified Texture and Semantic Feature Framework

Ke-Yue Zhang, Ruoxin Chen, Jiamu Sun, Jiangming Wang, Hao Yang, Taiping Yao, Shouhong Ding; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 3221-3229

Physical media-based presentation attacks (PAs) and digital editing-based DeepFakes (DFs) pose significant security challenges to face authentication systems. Previous research has proposed Presentation Attack Detection (PAD) and Face Forgery Detection (FFD) to address these issues separately. However, training these two models independently increases vulnerability to unknown attacks and complicates deployment. To simultaneously address both types of attacks, we propose a novel detection framework that leverages unified texture and semantic feature analysis to jointly counter these threats. Specifically, to enhance generalization to unseen and highly realistic synthetic attacks, we create synthetic counterparts closely aligned with real samples, reducing bias and promoting learning of transferable texture cues indicative of synthesis. Furthermore, to effectively counter 3D attacks, we fully leverage semantic features extracted by CLIP, a strategy that has proven robust against semantic attacks such as 3D masks. Finally, we fuse the two feature-based results to produce the final detection score. Our method achieves outstanding performance in the 6th Face Anti-Spoofing: Unified Physical-Digital Attacks Detection Challenge at ICCV 2025.