Low Visual Distortion and Robust Morphing Attacks Based on Partial Face Image Manipulation

Face verification is a popular way for verifying identities in access control systems. In this work, a partial face manipulation-based morphing attack (MA) is proposed to compromise the uniqueness of face templates. Different from existing research, this work changes MA from a holistic face level to component level, and only the most effective facial components (eyes and nose) are used. Therefore, a manipulated face is more similar to a bona fide one in terms of visual quality, texture, and noise characteristics. To validate the effectiveness of the proposed attack, a novel metric called actual mated morph presentation match rate (AMPMR) is proposed to evaluate MA performance under real-world conditions. With a collected dataset containing different attack types, image qualities, and manipulation parameters, the results indicate the proposed attack has better anti-detectability compared with the existing complete, splicing, and combined MAs. Moreover, it has low visual distortion and can reach a better tradeoff among facial biometrics verification, anti-detectability, and visual differences.

I. Introduction

In modern unsupervised access control systems of campuses, corporations, bus and train stations, face authentication has been widely deployed for identity verification. During authentication, an access control system reads a facial biometric reference (e.g., a template) stored in a user’s enrolment record in a database or in his/her personal access card, and then compares it with a trusted live facial image from the same subject. When a similarity score resulting from this comparison process is higher than the system’s threshold, a gate automatically opens to permit passing without human check, and it can significantly reduce users’ transaction time in the access control process. Meanwhile, to simplify the enrolment process, many systems allow users to submit their own photos when applying for access cards. If those photos comply to the predefined image standards [1], on-site data collections are not required. The unsupervised capture process of the enrolment image is convenient for ordinary users, but they also provide potential attack conditions for adversaries, such as morphing attacks (MAs). The goal of face MAs is to compromise the uniqueness of facial biometric templates with non-intrusive ways, to create a template (e.g., a manipulated face) that can match with multiple subjects. Once a manipulated facial image is maliciously injected into an enrolment record as biometric reference, multiple subjects can share and use the access card, which has negative impacts on public security.