Eyes don't lie: Unmasking fake faces with ocular clues.

Abstract

The rapid evolution of generative models necessitates robust methods to distinguish synthesized faces from real ones. This work evaluates a set of ocular geometry features for detecting synthesized faces generated by both GAN-based models (StyleGAN2/3) and text-to-image diffusion models (SFHQ-T2I), under controlled (Chicago Face Database, CFD) and uncontrolled (FFHQ, CelebA-HQ) acquisition conditions. Facial landmarks were extracted using MediaPipe Face Mesh, followed by XGBoost-based classification in both pairwise and pooled settings, covering 3128 real and 3470 synthesized images. The results showed a clear contrast across generation methods. GAN-based artifacts were strongly detectable under controlled conditions (CFD vs. StyleGAN2, AUC 99.11%). Diffusion-generated faces (SFHQ-T2I) were more effectively detected when compared against the uncontrolled FFHQ and CelebA-HQ datasets (AUC 84.36-86.47%). Vertical Misalignment consistently emerged as the strongest individual feature (AUC 65.85-97.56%). However, the pooled analysis showed reduced performance (AUC 65.66-67.36%), largely due to dataset heterogeneity that reflects realistic, in-the-wild conditions as confirmed by the literature. These findings indicate that physiological cues are highly context-dependent and that detection frameworks must adapt as generative models continue to improve in visual realism. Despite reduced performance in pooled settings, the proposed approach remains highly reliable for standardized, controlled images, supporting its use as a rapid triage tool for identity verification and as a complementary component within multimodal detection systems.