Face anti-spoofing based on projective invariants

The most common security authentication systems rely on automatic face recognition, which is particularly vulnerable to various spoofing attacks. Often these attacks include attempts to deceive a system by using a photo or video recording of a legitimate user. Recent approaches to this problem are based on pure machine learning techniques that require large training datasets and generalize or scale, poorly. By contrast, we present a geometric approach for detecting spoofing attacks in face recognition based authentication systems. By locating planar regions around facial landmarks, our method distinguishes between genuine user recordings and recordings of spoofed images such as printed photos and video replays. The proposed algorithm is based on projective invariant relationships that are independent of the camera parameters and lighting conditions. Unlike previous geometric approaches, the input to our system is a stream of two RGB cameras. Comparing with methods implemented by a single RGB camera, our approach is significantly more accurate and is completely automatic, since we do not require head movements and other user interactions. While, on the other hand, our method does not employ expensive devices, such as depth or thermal cameras, and it operates both in indoor and outdoor settings.