Detecting Presentation Attacks from 3D Face Masks Under Multispectral Imaging

Abstract

Automated detection of sensor level spoof attacks using 3D face masks is critical to protect integrity of face recognition systems deployed for security and surveillance. This paper investigates a multispectral imaging approach to more accurately detect such presentation attacks. Real human faces and spoof face images from 3D face masks are simultaneously acquired under visible and near infrared (multispectral) illumination using two separate sensors. Ranges of convolutional neural network based configurations are investigated to improve the detection accuracy from such presentation attacks. Our experimental results indicate that near-infrared based imaging of 3D face masks offers superior performance as compared to those for the respective real/spoof face images acquired under visible illumination. Combination of simultaneously acquired presentation attack images under multispectral illumination can be used to further improve the accuracy of detecting attacks from more realistic 3D face masks.