[Paper Review] Specular- and Diffuse-reflection-based Face Spoofing Detection for Mobile Devices

Akinori F. Ebihara, K. Sakurai+1 more|arXiv (Cornell University)|Jul 29, 2019

Biometric Identification and Security28 references4 citations

TL;DR

This paper proposes a low-cost, high-efficiency face spoofing detection method for mobile devices using only a single visible-light camera and flash. By analyzing specular reflections from the iris and diffuse reflections from the face surface via a dual-descriptor $SpecDiff$, the method achieves state-of-the-art accuracy—outperforming even deep learning models—while running six times faster and requiring only a small training dataset of ~1,000 image pairs per class.

ABSTRACT

In light of the rising demand for biometric-authentication systems, preventing face spoofing attacks is a critical issue for the safe deployment of face recognition systems. Here, we propose an efficient face presentation attack detection (PAD) algorithm that requires minimal hardware and only a small database, making it suitable for resource-constrained devices such as mobile phones. Utilizing one monocular visible light camera, the proposed algorithm takes two facial photos, one taken with a flash, the other without a flash. The proposed $SpecDiff$ descriptor is constructed by leveraging two types of reflection: (i) specular reflections from the iris region that have a specific intensity distribution depending on liveness, and (ii) diffuse reflections from the entire face region that represents the 3D structure of a subject's face. Classifiers trained with $SpecDiff$ descriptor outperforms other flash-based PAD algorithms on both an in-house database and on publicly available NUAA, Replay-Attack, and SiW databases. Moreover, the proposed algorithm achieves statistically significantly better accuracy to that of an end-to-end, deep neural network classifier, while being approximately six-times faster execution speed. The code is publicly available at https://github.com/Akinori-F-Ebihara/SpecDiff-spoofing-detector.

Motivation & Objective

• To develop a face presentation attack detection (PAD) algorithm suitable for resource-constrained mobile devices with minimal hardware and data requirements.
• To leverage visible-light flash-based imaging to distinguish live from spoofed faces using physical reflection properties.
• To achieve high detection accuracy with low computational cost, enabling real-time deployment on mobile CPUs.
• To validate generalization across diverse databases without requiring domain-specific retraining.
• To demonstrate that a handcrafted descriptor can outperform end-to-end deep learning in speed and accuracy under constrained conditions.

Proposed method

• The method captures two facial images: one with flash and one without, within ~200 ms, using a monocular visible-light camera.
• A Speculum Descriptor ($\mathcal{D}_{\mathrm{spec}} \in \mathbb{R}^{3200}$) is extracted from a 40×40 pixel region centered on the pupil to capture specular reflections from the iris.
• A Diffusion Descriptor ($\mathcal{D}_{\mathrm{diff}} \in \mathbb{R}^{10000}$) is computed from the entire face region to represent 3D facial structure via diffuse reflections.
• Both descriptors are normalized by luminance intensities to bound values in [-1, 1], improving classifier robustness.
• The descriptors are concatenated into a unified $SpecDiff$ descriptor, which is then classified using standard models like SVM or neural networks.
• The flash is simulated using a bright all-white screen to ensure consistent illumination conditions.

Experimental results

Research questions

• RQ1 Can a flash-based PAD method achieve high accuracy using only a single visible-light camera and minimal hardware?
• RQ2 Can specular and diffuse reflection characteristics be effectively combined into a single descriptor to improve spoof detection?
• RQ3 Does a handcrafted descriptor outperform deep learning models in speed and accuracy on mobile devices with limited computational resources?
• RQ4 How well does the $SpecDiff$ descriptor generalize across diverse databases such as NUAA, Replay-Attack, and SiW?
• RQ5 Can the method maintain robust performance under varying lighting conditions and spoofing types?

Key findings

• The $SpecDiff$ descriptor achieved an sACER of 0.58% on the NUAA database using an RBF-SVM, outperforming all other flash-based PAD methods and the ResNet18 model.
• On the Replay-Attack database, the $SpecDiff$ descriptor achieved an sACER of 0.43% with RBF-SVM, significantly lower than the next best method (1.24% for Implicit3D).
• On the SiW database, the $SpecDiff$ descriptor achieved an sACER of 0.86% with RBF-SVM, outperforming all other flash-based methods and the ResNet4 model.
• The method achieved a 6× speedup over an end-to-end deep neural network (ResNet18) while maintaining statistically superior accuracy on all test databases.
• The algorithm demonstrated robustness under varying lighting: APCER dropped to 0.0087% in dark conditions and 0.27% in bright conditions on the in-house dataset.
• The method achieved near-perfect detection for photo and video spoofing subcategories, with 0.00% APCER for video on Replay-Attack and SiW databases.