Detecting Face Morphing Attacks with Collaborative Representation of Steerable Features

Abstract

Passports have used face characteristics to verify and establish the identity of an individual. Face images provide high accuracy in verification and also present the opportunity of verifying the identity visually against the passport face image if the need arises. Morphed image-based identity attacks are recently shown to exploit the vulnerability of passport issuance and verification systems, where two different identities are morphed into one image to match against both images. The challenge is further increased when the properties in the digital domain are lost after the process of print and scan. This work addresses such a problem of detecting the morphing of face images such that the attacks are detected even after the print and scan process. As the first contribution of this work, we extend an existing database with 693 bonafide and 1202 morphed face images with the newly added of 579 bonafide and 1315 morphed images. We further propose a new approach based on extracting textural features across scale-space and classifying them using collaborative representation. With a set of extensive experiments and benchmarking against the traditional (non-deep-learning methods) and deep-learning methods, we illustrate the applicability of the proposed approach in detecting the morphing attacks. With an obtained Bonafide Presentation Classification Error (BPCER) of $13.12\%$ at Attack Presentation Classification Error Rate (APCER) of $10\%$, the use of the proposed method can be envisioned for detecting morph attacks even after print and scan process.