Periocular biometrics survey uses eye region to spot deepfakes

A new survey paper accepted at the ECCV 2026 Workshop on AI for Multimedia Forensics & Disinformation Detection provides the first comprehensive review of periocular soft biometrics—extracting demographic attributes like gender, age, and ethnicity from the eye region. Authored by Fernando Alonso-Fernandez, Kevin Hernandez-Diaz, and Josef Bigun, the paper argues the periocular area is uniquely valuable because it remains visible even when the rest of the face is occluded by masks, scarves, or poor surveillance angles. The survey systematically catalogs publicly available datasets, traces the evolution from handcrafted descriptors to modern deep learning architectures, and benchmarks current state-of-the-art performance for gender, age, and ethnicity prediction.

The authors position this technology as a critical tool for multimedia forensics and disinformation detection. Use cases include demographic filtering in surveillance footage, age verification systems, and a novel application: spotting demographic inconsistencies in AI-generated synthetic media. For example, if a deepfake image claims a subject's age or ethnicity but the periocular region subtly violates biometric patterns, that can trigger forensic flagging. However, the paper also warns of open challenges: dataset bias, poor cross-domain generalization, fairness concerns, and the near-total absence of forensic-oriented benchmarks. The authors call for standardized evaluation protocols specifically designed for real-world forensic scenarios, where image quality is degraded and subject cooperation is zero.