Geopixelnet: Monocular Gaussian Curvature Maps For Real-Time Face Anti-Spoofing

Abstract

Face recognition systems are increasingly vulnerable to presentation attacks, including printed photographs and screen replay. Existing anti-spoofing methods predominantly operate on 2D texture or binary supervision, overlooking the intrinsic geometric distinction between a real 3D face and a flat spoofing artifact. This paper introduces Geo-Shield, a dual-stream deep learning framework that exploits the differential Gaussian curvature of reconstructed 3D facial manifolds for presentation attack detection. The system extracts 468 three-dimensional landmarks from a single RGB image via MediaPipe Face Mesh, constructs a Delaunay-triangulated surface, and computes per-vertex Gaussian curvature K through local quadric fitting of the shape operator. A real face exhibits non-zero curvature (K not equal to 0), whereas a flat photo or display replay produces K approximately equal to 0. The curvature map is processed by a custom manifold-stream CNN, fused with a MobileNetV3 spatial-stream backbone through a squeeze-and-excitation gated adaptive fusion module. Experiments on four benchmark datasets demonstrate that Geo-Shield achieves 0.68% Average Classification Error Rate (ACER) on the NUAA benchmark and 2.32% ACER on the challenging CelebA-Spoof dataset after 30 training epochs, at an inference latency under 4 ms per sample.