Cancelable Iris template generation by aggregating patch level ordinal relations with its holistically extended performance and security analysis

Abstract

Nowadays, biometric-based authentication is gaining immense popularity due to the widespread usage of digital activities. Among various biometric traits, the iris is one of the most discriminative, accurate, and popularly used biometrics. However, due to its immutable nature, it is highly vulnerable to adversarial attacks if stolen and thus poses a severe security threat. Here, in this work, we propose a cancelable iris biometric authentication system that stores a transformed version of the original iris template and thus enables cancelation and re-enrolment in case if the original template is stolen. Firstly, for extracting discriminative iris features, we have proposed a novel deep architecture based on aggregation learning. This deep architecture makes use of qualitative measure (ordinal measure), unlike popularly used quantitative measures. The usage of ordinal measures in this work enables to encode distinctive iris features quite well. Later generated iris features are protected using state-of-the-art two representative cancelable biometric techniques, namely BioHashing and 2N discretized BioPhasor. Finally, in order to justify the efficacy of the proposed architecture, we have presented rigorous and holistic security analysis. To the best of our knowledge, this is the first work that has presented such an in-depth analysis of any deep network in the context of cancelable iris biometrics. Experimental results over four datasets viz. CASIA-V3 Interval, CASIA-Lamp, IITD, and IITK demonstrate the efficacy of the proposed framework in terms of security and accuracy. Further, for better network explainability, we have also performed layer-specific heatmap and feature map analysis to ascertain what exactly our novel deep architecture is learning.