Improving Presentation Attack Detection for ID Cards on Remote Verification Systems

TL;DR

This paper presents a new two-stage deep learning approach for detecting presentation attacks on ID cards in remote verification systems, achieving high accuracy and developing a new evaluation framework.

Contribution

It introduces an end-to-end two-stage detection method based on MobileNetV2 and a new framework PyPAD for multi-class attack detection aligned with ISO standards.

Findings

Achieved BPCER100 scores of 1.69% and 2.36% with individual models.

Combined models reduce BPCER100 to 0.92%.

Developed a large database of 190,000 ID card images for training and testing.

Abstract

In this paper, an updated two-stage, end-to-end Presentation Attack Detection method for remote biometric verification systems of ID cards, based on MobileNetV2, is presented. Several presentation attack species such as printed, display, composite (based on cropped and spliced areas), plastic (PVC), and synthetic ID card images using different capture sources are used. This proposal was developed using a database consisting of 190.000 real case Chilean ID card images with the support of a third-party company. Also, a new framework called PyPAD, used to estimate multi-class metrics compliant with the ISO/IEC 30107-3 standard was developed, and will be made available for research purposes. Our method is trained on two convolutional neural networks separately, reaching BPCER\textsubscript{100} scores on ID cards attacks of 1.69\% and 2.36\% respectively. The two-stage method using both models together can reach a BPCER\textsubscript{100} score of 0.92\%.