Design, Implementation, and Comparative Evaluation of a Reduced-Feature Biometric System for Secure Authentication

Abstract

Biometric authentication systems are widely adopted in security authentication such as access control, mobile devices, bank, and Internet of Things (IoT) applications. The traditional biometric systems, however, tend to store and process large templates of features in the system, which leads to higher storage, computational and verification costs. In this research, we suggest a reduced pattern-size biometric recognition system by selecting highly discriminative features to represent a user, which is called “louder” features. The proposed approach identifies and preserves only the features that are stable for the genuine user and significantly different from other users, rather than retaining all the features extracted from a biometric trait.

The research focuses on fingerprint biometrics since fingerprint images contain measurable global ridge features and local minutiae features which can be selectively reduced. Firstly, a full-featured biometric system was developed as a baseline system. A set of user-specific feature-ranking and selection techniques were used to form compact biometric templates composed of the most informative ones. The proposed design was applied, evaluated and validated end-to-end (from pre-processing to minutiae extraction, user-specific feature selection, minutiae matching and evaluation) on real fingerprint imagery. The system was designed for eventual deployment against the Sokoto Coventry Fingerprint Dataset (SOCOFing), a publicly available corpus of more than 6,000 fingerprint images including altered samples suitable for robustness testing; the pilot results reported here were obtained on the NIST Minutiae Interoperability Exchange (MINEX) III validation imagery for reasons of data accessibility, as detailed in Section 4.2. Storage size, verification latency, ROC curves, False Acceptance Rate (FAR), False Rejection Rate (FRR), Equal Error Rate (EER), and recognition accuracy are compared between the full and reduced feature sets.

The reduced-feature template achieved a lower EER (0.386 vs. 0.460), higher ROC AUC (0.670 vs. 0.539), a 60% smaller mean template size (153.1 vs. 382.7 bytes), and 62% faster mean matching time (6.35 ms vs. 16.74 ms) than the full-feature baseline on the evaluated pilot set. These results support the hypothesis that user-specific feature selection can yield a lightweight biometric system without sacrificing — and in this pilot, while improving — authentication performance, though confirmation on the full SOCOFing corpus and against commercial-grade matchers is required before the finding can be generalized.