Robust Algorithms For Face Morphing Attack Detection: Database, Vulnerability and Detection

Abstract

Biometrics has emerged as a promising technology for automated recognition of individuals. The stored biometric characteristics are used to recognise an individual and hence biometrics technology plays a major role in security-related applications and stands in the front-line for authentication of data subjects. Biometrics has a wide range of applications in law enforcement, surveillance, banking, border control, medical records, time and attendance tracking etc. As the inherent biometrics characteristics don’t undergo any changes, biometric technology has shown the best performance for authentication of data subjects. Though biometrics technology is promising for person authentication, attackers may employ various techniques like presentation attacks and adversarial attacks to impersonate an enrolled individual with interest to obtain unauthorised access to the system. In addition to these attacks, in relatively recent times, biometric are attacked in the facial image enrolment stage, especially in the ID related applications, by performing face morphing. Facial morphing is initially performed for entertainment purposes, but gradually it has been used to attack face recognition systems. The face morphing process combines two different facial identities to generate a single facial image with the facial representations of both identities. An attacker may use the morphed facial image to enrol it in the ID documents (like driving license, passport). Since the morphed facial image shows a high resemblance to both facial identities, the ID document can be claimed by both identities. This indicates the severity of facial morphing and the necessity of morphing attack detection mechanisms to avoid the security lapse. Hence the primary objective of this thesis is the development of face morphing attack detection techniques using hand-crafted and deep learning approaches. During this doctoral work, morphing attack detection approaches are developed for both digital and print-scan datasets. To empirically evaluate the performance of the newly generated morphing attack detection approaches, various face morphing databases are generated using landmark and deep learning-based GAN techniques. Furthermore, the vulnerability of face recognition systems to face morphing attacks with ageing co-variate is evaluated. To this extent, this doctoral thesis contributes with the novel morphing attack detection approaches.