Reproduce “Liveness Detection in Computer Vision: Transformer-based Self-Supervised Learning for Face Anti-Spoofing” on OpenResearch

About this paper

Arman Keresh, Pakizar Shamoi

Researchers from the Kazakh-British Technical University developed a face anti-spoofing system utilizing a Vision Transformer fine-tuned with the DINO self-supervised learning framework. This approach dramatically reduced the Attack Presentation Classification Error Rate to 1.6% and the Average Classification Error Rate to 0.8%, leveraging both public and a large proprietary unlabeled dataset to accurately detect spoofing attempts.

Key insights

• Vision Transformers, with their global receptive fields, are significantly more effective at detecting sophisticated spoofing attacks than CNNs, which primarily focus on local features.
• Self-supervised learning via DINO allows for the robust training of face anti-spoofing models using vast amounts of unlabeled real-world data, circumventing the need for costly and time-consuming manual labeling.
• The integration of ViTs with DINO achieves a superior balance in preventing both false acceptance of spoofing attacks and false rejection of legitimate users, as indicated by notably reduced ACER.

Summary by alphaXiv.