Monocular Facial Presentation–Attack–Detection: Classifying Near-Infrared Reflectance Patterns

Abstract

This paper presents a novel material spectroscopy approach to facial presentation–attack–defense (PAD). Best-in-class PAD methods typically detect artifacts in the 3D space. This paper proposes similar features can be achieved in a monocular, single-frame approach by using controlled light. A mathematical model is produced to show how live faces and their spoof counterparts have unique reflectance patterns due to geometry and albedo. A rigorous dataset is collected to evaluate this proposal: 30 diverse adults and their spoofs (paper-mask, display-replay, spandex-mask and COVID mask) under varied pose, position, and lighting for 80,000 unique frames. A panel of 13 texture classifiers are then benchmarked to verify the hypothesis. The experimental results are excellent. The material spectroscopy process enables a conventional MobileNetV3 network to achieve 0.8% average-classification-error rate, outperforming the selected state-of-the-art algorithms. This demonstrates the proposed imaging methodology generates extremely robust features.

Improving Monocular Facial Presentation–Attack–Detection Robustness with Synthetic Noise Augmentations

“…Chingovska et al alternatively use a popular texture descriptor, the local-binary-pattern (LBP) [34], noting it cannot discern liveliness for RGB cameras. We similarly concluded that LBP is sensitive to visible light variance, but robustness can be notably improved when using the illuminated near-infrared [16].…”

“…This work, however, is proven out in static lighting conditions. Others have been unsuccessfully able to reproduce the results with dynamic lighting (including our own lab [16]). Chingovska et al alternatively use a popular texture descriptor, the local-binary-pattern (LBP) [34], noting it cannot discern liveliness for RGB cameras.…”

“…All three of these are benchmarked in our prior work. While they do excellently in competition data, they actually over-fit when introducing a high-variance spoofing dataset (and are outperformed by a simple MobileNet binary classifier) [16]. This is the fundamental challenge of PAD.…”

“…Monocular PAD methods are often based on fine artifact detection methods [6,7,13]. Spoofing articles are rarely 100% perfect representations of the real face, where artifacts can be observed in textural aberrations [14], facial "correctness" [15], and light reflectance [16]. These features work well in measured laboratory conditions but can be impacted by the imaging environment.…”

“…This can be partially mitigated by picking the right imaging spectrum. Our figure juxtaposes the human perspective (RGB camera) against the near-infrared, where it is obvious that the contrast between live faces and spoofs increases due to differences in material reflectance [16]. Still, there are concerns with over-fitting.…”