Snapshot Spectral Face Anti‑Spoofing Detection Framework Based on Supervised Contrastive Learning Wins CVPR2024 Challenge

Recently, the results of the CVPR2024 Global Face Anti‑Spoofing Detection Challenge were announced. Out of 205 participating teams, the Ant Group Security Lab team achieved a 0.0000% average recognition error rate, securing first place in the snapshot spectral face anti‑spoofing track.

CVPR, organized by IEEE, is one of the three top conferences in computer vision alongside ICCV and ECCV. Face anti‑spoofing is a key research direction in this field, and this year’s global challenge was held at CVPR.

The challenge focuses on resisting face fraud in uncontrolled scenarios and promoting the development of anti‑spoofing technologies. It is now in its fifth edition, having attracted more than 1,200 teams worldwide.

This edition introduced a new snapshot spectral imaging method, encouraging the exploration of algorithms tailored to snapshot spectral images.

Compared with traditional RGB images, snapshot spectral imaging captures both spatial and spectral information in a single exposure, revealing material‑specific spectral reflections and absorptions and providing richer color details. This greatly enhances the ability of face recognition systems to detect silicone/latex masks, photos, or AI‑generated deepfakes.

However, the application of snapshot spectral imaging to face anti‑spoofing is still exploratory. The provided dataset suffers from a scarcity of training samples and a severely imbalanced distribution of real versus fake faces and various spoof types, which can lead to poor model generalization on unseen forgeries.

To address these issues, Ant Group Security Lab leveraged over a decade of experience in face anti‑spoofing and proposed an innovative "Supervised Contrastive Learning‑Based Snapshot Spectral Face Anti‑Spoofing Detection Framework." The framework incorporates a multi‑task learning strategy that adds a contrastive learning task to the conventional classification task, strengthening the aggregation of features across different spoof types and improving the learning of common spoof characteristics.

To further boost supervised contrastive learning, the team employed data resampling and intra‑class mixing strategies to rebalance the data distribution and increase sample diversity, as well as a real‑face‑oriented re‑weighting scheme to reduce the influence of identity, mask, or other irrelevant features, thereby focusing the model on distinguishing real from fake faces.

The related paper has been accepted to the CVPR2024 Face Anti‑Spoofing Workshop.

With the rapid development of AIGC, digital humans, and other technologies, face anti‑spoofing faces increasingly complex challenges, prompting the industry to explore and adopt new algorithmic research.

Ant Group has been deep‑invested in biometric security for over ten years, holding more than 50 international patents and multiple national standards, winning over ten top‑level portrait perception competitions, and obtaining certifications such as CNAS, China Academy of Information and Communications Technology deep‑fake video detection, and iBeta Level 2.

These research outcomes have been integrated into core products of companies like Omron, Huawei, and Ant Group, achieving large‑scale commercial deployment and serving hundreds of millions of users worldwide.