A Multimodal Biometric Authentication for Smartphones

Abstract

Biometrics is seen as a viable solution to ageing password based authentication on smartphones. Fingerprint biometric is leading the biometric technology for smartphones, however, owing to its high cost, major players in mobile industry are introducing fingerprint sensors only on their flagship devices, leaving most of their other devices without a fingerprint sensor. Cameras on the other hand have been seeing a constant upgrade in sensor and supporting hardware, courtesy of ‘selfies’ on all smartphones. Face, iris and visible vasculature are three biometric traits that can be captured in visible spectrum using existing cameras on smartphone. Current biometric recognition systems on smartphones rely on a single biometric trait for faster authentication thereby increasing the probability of failure to enroll, affecting the usability of the biometric system for practical purposes. While multibiometric system mitigates this problem, computational models for multimodal biometrics recognition on smartphones have scarcely been studied. This dissertation provides a practical multimodal biometric solution for existing smartphones using iris, periocular and eye vasculature biometrics. In this work, computational methods for quality analysis and feature detection of biometric data that are suitable for deployment on smartphones have been introduced. A fast, efficient feature detection algorithm (Vascular Point Detector) for identifying interest points on images garnered from both rear and front facing camera has been developed. It was observed that the retention ratio of VPD for final similarity score calculation was at least 10% higher than state of art interest point detectors such as FAST, over various datasets. An interest point suppression algorithm based on local histograms was introduced, reducing the computational footprint of matching algorithm by at least 30%. Further, experiments are presented which successfully combine multiple samples of eye vasculature, iris and periocular biometrics obtained from a single smartphone camera sensor. Several methods are explored to test the effectiveness of multi-modal and multi algorithm fusion at various levels of biometric recognition process, with the best algorithms performing under 2 second on an IPhone 5s. It is noted that the multimodal biometric system outperforms the unimodal biometric systems in terms of both performance and failure to enroll rates.