Enhancing Secured Voter Authentication For E-Voting System With Fusion Of Multimodal Biometric Traits Using Cryptographic Keys

Elections increasingly rely on some form of technology-assisted identity verification to prevent duplicate voting and impersonation while accelerating voter processing at the polling place. Biometric authentication is attractive for this role because, unlike a voter-ID card or a signature, it binds a claimed identity to a trait that is difficult to transfer between individuals. However, biometric voter verification raises a distinct set of engineering and ethical concerns that a purely academic recognition-accuracy study does not have to confront: the system must operate reliably for a very large and demographically diverse population, it must produce a decision an election officer can act on immediately, and because biometric traits cannot be revoked or reissued the way a password can any stored template must be protected against extraction even if the underlying database is later compromised [1, 2].

Single-modality biometric verification compounds these problems. A fingerprint-only system fails predictably for voters with worn, scarred, or occupationally eroded ridge detail, and an iris- or face-only system is sensitive to lighting, eyewear, and camera angle. Multimodal fusion, in which two or more independent biometric matchers are combined into a single decision, is well established as a way to reduce both false-accept and false-reject rates relative to either matcher alone, precisely because an attacker must simultaneously defeat two independent sensing and matching pipelines, and because a poor-quality capture in one modality can be compensated for by higher confidence in the other [2, 3]. Several recent applied studies specifically propose fingerprint-plus-iris fusion for electronic voting for this reason [4], motivated in part by the observation that fingerprint-only electoral biometric systems in the field have measurable failure rates attributable to worn or damaged ridges among manual-labor populations [4].

At the same time, most applied biometric-voting prototypes in the recent literature we surveyed (Section II) report system architecture and, where available, headline accuracy figures, but say comparatively little about what is actually stored and who is in control of the decision to proceed. Two gaps recur across this literature and motivate the present work. First, iris- or face-capturing systems intended for a supervised polling-place setting rarely address the privacy exposure of the capture step itself: a live video feed of a voter's full face, visible to whatever operator or screen-recording software has access to the terminal, is a meaningfully different privacy posture from a system that never displays or transmits more of the face than the iris region needed for matching. Second, and more subtly, systems that combine multiple biometric stages into a single automated pipeline typically advance from capture to decision without an explicit human checkpoint, which is a reasonable default for a fully self-service kiosk but is a poor fit for the officer-supervised model used in many real polling places, where an election official is present specifically to exercise judgment over edge cases.

This paper reports on a complete, working system, SecureVote Biometrics, designed around both observations. Its contributions are:

1) A two-stage fingerprint matcher operating on inked, scanned or photographed thumbprints, combining crossing-number minutiae extraction with ORB descriptor matching, together with a duplicate-enrollment check that compares every new registration against the full existing template population.

2) An iris capture pipeline in which the operator-facing preview never displays an identifiable face  the frame is heavily blurred outside a tracked iris-sized circle combined with finite-state blink-liveness detection and automatic multi-frame best-of-burst selection driven by a five-factor quality score.

3) A human-in-the-loop confirmation design in which neither registration nor authentication auto-advances on a successful capture: the system displays the captured image and its full quality/match breakdown and requires an explicit officer action confirmation on a match, or retry on a non-match before proceeding, which we show (Section VII) is absent from comparable recent biometric-voting prototypes.

4) AES-256-GCM authenticated encryption of every stored template, with raw images and camera frames never persisted at any stage.

5) An evaluation methodology and an honest account of its limits: because a labeled, multi-subject biometric dataset and physical capture hardware were not available in the development environment, we validate the algorithmic pipelines on controlled synthetic and perturbed imagery and validate the complete officer-supervised workflow through automated, browser-driven system tests, rather than reporting fabricated population-scale accuracy figures.

The remainder of this paper is organized as follows. Section II reviews recent literature on fingerprint and iris recognition, liveness detection, multimodal fusion, template protection, and biometric voting systems. Section III analyzes the resulting research gap. Section IV describes the proposed system architecture and methodology in detail, including mathematical formulations and algorithms for each stage. Section V describes the experimental design; Section VI reports results. Section VII compares the design against recent related work. Section VIII discusses limitations and implications, and Section IX concludes.

II. LITERATURE REVIEW

We organize recent (2021–2026) work into six threads directly relevant to the system described in Section IV: fingerprint minutiae extraction and enhancement, iris recognition and texture matching, liveness and best-frame selection, multimodal fusion, biometric template protection, and applied biometric voting systems, closing with work comparing classical and learned face/eye detectors, which bears directly on a substitution decision made in our own implementation.

A. Fingerprint Minutiae Extraction and Enhancement

Minutiae-based matching comparing ridge endings and bifurcations rather than raw pixel intensity remains the dominant paradigm in both classical and learned fingerprint pipelines because it tolerates non-linear skin deformation and yields a compact, comparable template. Sumalatha et al. [1] survey unimodal and multimodal fingerprint authentication broadly, covering fusion strategies, presentation-attack vectors, and template-protection schemes, and report that hybrid classical-plus-deep-learning pipelines increasingly dominate top benchmark results while classical crossing-number and orientation-field methods remain the most common choice in resource-constrained or embedded deployments the setting our system targets. Martins et al. [10] apply a crossing-number variant together with Gabor-orientation ridge enhancement in a forensic-scenario fingerprint pipeline, reporting that enhancement quality dominates downstream minutiae-extraction accuracy more than the extraction rule itself a finding that directly informed our decision to invest in a multi-stage enhancement chain (CLAHE, median filtering, an eight-orientation Gabor bank) ahead of a comparatively simple crossing-number extractor rather than a more elaborate extractor operating on weakly enhanced ridges. Situmorang [11] applies the crossing-number method specifically to phone-camera-captured prints rather than flatbed scans, reporting usable minutiae yield despite consumer-camera noise, which supports our design choice to accept either a scanner export or a phone photograph as fingerprint input.

Deep-learning minutiae extraction is increasingly common: Zhang et al. [12] propose a multi-task convolutional network for contactless fingerprint minutiae extraction that jointly predicts minutiae location, orientation, and type, outperforming classical extractors on their contactless benchmark; Abdulkarim et al. [13] go further and combine deep-learning extraction with a classical CLAHE-Gabor-crossing-number front end in a hybrid pipeline, reporting that the classical stage still measurably improves downstream accuracy even when a learned extractor is available. These results frame a genuine design trade-off rather than a solved question: learned extractors report higher accuracy on their training distributions but require GPU inference and labeled training data neither of which are guaranteed to be available in a low-resource or fully offline electoral deployment, which is the operating assumption of the system in this paper. Manchanda et al. [14] specifically evaluate CLAHE combined with Gaussian smoothing for latent (partial, degraded) fingerprint enhancement and report consistent ridge-clarity gains, corroborating our own enhancement chain design.

B. Iris Recognition and Texture-Based Matching

Classical iris recognition following Daugman's Gabor-phase IrisCode remains accurate but computationally heavier than texture-descriptor alternatives; Local Binary Patterns (LBP) are a widely used lighter-weight substitute for encoding iris texture. Taha et al. [15] evaluate LBP-based iris feature extraction across near-infrared and visible-spectrum imagery, reporting high same-sensor accuracy but a marked drop when matching across sensor types, which is consistent with our own choice to keep enrollment and verification on the same capture pathway (a standard webcam) rather than mixing sensor modalities. Because LBP alone is a purely statistical texture descriptor with no explicit keypoint geometry, several recent systems pair it with a local keypoint descriptor for a second, geometry-aware matching signal; our own pipeline follows this pattern by pairing an LBP histogram with ORB keypoint descriptors rather than relying on LBP alone.

Robust localization of the iris/pupil boundary under non-ideal, uncontrolled capture conditions as opposed to a dedicated, cooperative iris scanner is treated explicitly in Jan et al. [16], who combine Viola-Jones (Haar cascade) eye detection with geometric fallback rules for eye/iris segmentation specifically because pure circular Hough-transform localization is unreliable on non-ideal images; our own detector chain (Section IV-B) follows essentially the same pattern Haar-cascade region detection followed by Hough-circle localization with a geometric and darkest-region fallback independently arrived at under a similar offline, non-cooperative-capture constraint. Iris capture quality assessment is addressed directly in Garea-Llano et al. [17], who propose a multi-factor iris-pupil quality framework aligned with the ISO/IEC 19794-6 biometric sample-quality standard and evaluate it on real-time video, and in Fazilov et al. [18], who combine several individual iris quality parameters via principal component analysis into a single composite score. Both works validate the general strategy used in our own quality assessment (Section IV-B): decompose capture quality into several independent, individually interpretable sub-metrics and combine them into one actionable score, rather than relying on a single opaque quality proxy.

C. Liveness Detection and Best-Frame Selection

Blink-based liveness checks, using the eye-aspect-ratio (EAR) concept, remain a standard low-cost defense against static-photo presentation attacks. Demirsoy et al. [19] implement EAR-based blink liveness for an access-control terminal with an SVM classifier and embedded hardware trigger, and Chong et al. [20] apply EAR blink detection within a masked-face recognition pipeline, both reporting that a genuine open-close-open transition is difficult to reproduce with a static image, though neither claims robustness against a video-replay attack a limitation we inherit and discuss in Section VIII. Because true eyelid landmarks were not available to us (Section IV-B), our implementation substitutes a bounding-box height/width proxy for EAR computed from Haar-cascade eye detections; this is, to our knowledge, a novel adaptation rather than one reported directly in the surveyed literature, and we validate it only as an internal liveness-transition signal rather than a landmark-accurate EAR measurement.

Selecting the best frame from a short burst, rather than acting on the first available frame after a trigger, is directly addressed in Kulloli et al. [21], who defend against iris video-replay attacks by scoring candidate frames with quality metrics including Laplacian-variance sharpness and selecting the highest-scoring frame a technique essentially identical to the best-frame selection stage in our own pipeline (Section IV-B), independently justified here for capture-quality assurance rather than anti-replay defense specifically. Kharchevnikova and Savchenko et al. [22] study best-frame selection for video-based face recognition more generally, reporting that quality-driven frame selection measurably improves downstream recognition accuracy compared to naive first-frame or fixed-interval sampling, which we take as further support for applying the same principle to iris capture.

D. Multimodal Biometric Fusion

Es-Sobbahi et al. [2] systematically review 29 recent studies on multimodal biometric fusion and find score-level fusion remains the most widely deployed strategy in applied systems specifically because it requires no joint training data and degrades gracefully when one modality is temporarily unavailable, which matches our own rationale for using a simple weighted-sum score-level fusion (Section IV-D) rather than a trained fusion classifier. Kamlaskar et al. [3] fuse iris and fingerprint modalities directly, reporting improved separation between genuine and impostor score distributions relative to either modality alone — the same pairing of modalities used in this paper. Rao et al. [23] extend fusion to three modalities (face, fingerprint, iris) with an explicit privacy-preserving template-protection layer, which is closer in spirit to our own privacy-conscious design goal, though at the cost of additional enrollment burden from a third modality; we discuss this trade-off in Section VII.

E. Biometric Template Protection

Because biometric traits cannot be revoked, encrypting or otherwise irreversibly transforming stored templates is treated as a baseline requirement in recent template-protection literature rather than an optional hardening step. Imran et al. [24] combine AES encryption with a minutiae cylinder-code representation for fingerprint templates; Ayoup et al. [25] apply AES specifically to combined eye-and-fingerprint biometric storage, closely matching our own two-modality scope; and El-Shafai et al. [26] propose a genetic-algorithm-tuned cancelable-biometric encryption scheme evaluated across several biometric modalities. All three use AES as the underlying primitive but do not uniformly specify an authenticated mode; our implementation uses AES-256-GCM specifically, which provides built-in tamper detection (an authentication tag) in addition to confidentiality, so that a corrupted or tampered stored template fails decryption explicitly rather than silently returning corrupted plaintext (Section IV-E).

F. Biometric Voting Systems

Several recent applied papers propose biometric-gated electronic voting directly. Umar et al. [4] propose a bi-factor fingerprint-and-iris voting design explicitly motivated by fingerprint-only systems' field failure rate among voters with worn or damaged ridges, closely matching our own modality choice and rationale; Zamir et al. [5] describe an EVM (electronic voting machine) gated by biometric authentication linked to a national identity database; Ajao et al. [6] and Faruk et al. [7] both combine multimodal biometrics with blockchain ledgers for vote-record integrity, the latter also incorporating deep-learning-based recognition. None of these four report on what happens to the capture preview shown to a face- or iris-capturing sensor during the voting session, nor do they describe a point at which an automated pipeline stops and asks a human supervisor to confirm a borderline or failed capture before the system proceeds both are addressed directly in the system design in Section IV of this paper.

Duplicate-enrollment prevention at scale is treated separately from real-time matching in this literature. Nlend et al. [8] describe deduplication specifically for large voter-registration fingerprint databases, and Koner et al. [9] describe Bharat ABIS, a billion-scale multimodal (fingerprint, face, iris) biometric search system underlying India's Aadhaar national identity program, reporting deduplication and search operating at national population scale using deep-learning-accelerated matching. Our own duplicate-enrollment check (Section IV-A) implements the same underlying principle compare every new template against the full existing population before accepting an enrollment at a vastly smaller scale appropriate to a single-precinct or pilot deployment, using our own two-stage classical matcher rather than a learned large-scale index.

G. Detector Choice: Classical vs. Learned Face/Eye Detection

Our own iris pipeline substitutes OpenCV Haar cascades for the originally targeted MediaPipe Face Mesh landmark detector because the latter's model bundle could not be downloaded in our network-restricted development environment (Section IV-B). Recent comparative studies support this substitution as a reasonable engineering trade-off rather than an unprincipled downgrade. Adeshina et al. [29] directly compare Haar-cascade and deep-learning (ResNet-based) face detectors, reporting substantially lower true-positive detection accuracy for Haar cascade but a two-orders-of-magnitude inference-speed advantage (on the order of one second versus over two minutes per evaluation set in their benchmark), making Haar cascade the more viable choice for a CPU-only, offline, latency-sensitive terminal such as a polling-place kiosk. Elbert et al. [30] compare Haar cascade, Dlib, and MediaPipe directly for face/eye detection and report the expected accuracy ordering (MediaPipe highest, Haar cascade lowest) but confirm Haar cascade remains usable for coarse region localization when paired with a downstream refinement stage in our pipeline, the Hough-circle iris localization step which is exactly the architecture we adopt.

H. Feature Descriptor Choice

Bansal et al. [27] provide a widely cited comparative evaluation of SIFT, SURF, and ORB descriptors, reporting that ORB, while less robust to large scale change than SIFT, offers markedly faster computation and is patent-unencumbered, which motivated our use of ORB as the descriptor stage for both the fingerprint (Section IV-A) and iris (Section IV-B) pipelines. Xie et al. [28] specifically pair adaptive histogram equalization with ORB matching and report improved keypoint repeatability on contrast-degraded imagery relative to ORB alone, which supports our own decision to apply CLAHE immediately before ORB descriptor computation in both pipelines rather than computing descriptors on raw, unenhanced images.

III. RESEARCH GAP ANALYSIS

Synthesizing Section II against the design of the system in this paper, we identify four specific gaps in the recent applied biometric-voting and multimodal-biometric literature that this work directly addresses.

A. Capture-Time Privacy Is Rarely Addressed Explicitly

Biometric voting proposals [4, 5, 6, 7] describe what is matched and, in the blockchain-based designs [6, 7], how the resulting decision is recorded immutably, but none describe what an operator or a compromised terminal actually sees during a live face- or iris-capture session. A full-frame video feed of a voter's face, even if never written to disk, is itself a privacy exposure if visible to a screen recorder, a second monitor, or an over-the-shoulder observer a concern distinct from, and complementary to, at-rest template encryption [24, 25, 26]. Our design closes this gap at the capture layer itself, not only the storage layer, by rendering only a heavily blurred frame with a sharp iris-sized circular window during capture (Section IV-B), so the raw face is never displayed, transmitted, or stored at any point in the session.

B. Automated Pipelines Lack a Human-Supervision Checkpoint

Every biometric-voting system we reviewed [4, 5, 6, 7] describes an automated capture-to-decision pipeline. This is an appropriate default for an unsupervised self-service kiosk, but the officer-supervised polling-place model in which a human official is present specifically to exercise judgment is a different operating context that these designs do not distinguish. An automated system that silently proceeds past a borderline or failed capture removes exactly the judgment the officer is present to provide. This concern is not merely theoretical for the system in this paper: an earlier iteration of our own implementation exhibited precisely this failure mode (a captured iris result auto-advanced to the fused decision regardless of match outcome), which we identified and corrected during development (Section V, Section VIII). We are not aware of a recent biometric-voting paper that discusses this checkpoint design question directly, and we position our officer-confirmation workflow (Section IV-E) as a direct response to it.

C. Duplicate-Enrollment Detection Is Studied at National Scale, Not Precinct Scale

Nlend et al. [8] and Koner et al. [9] both address fingerprint or multimodal deduplication at national-database scale using infrastructure (distributed indexing, GPU-accelerated deep matching) that is unavailable to a small-precinct or pilot deployment. A precinct-scale system nonetheless needs the same underlying guarantee that a single physical fingerprint cannot be enrolled under two different identities without that infrastructure. Our duplicate-enrollment check (Section IV-A) demonstrates that the same guarantee can be provided at smaller scale with a classical two-stage matcher and a linear scan against the existing template population, appropriate to the deployment scale this paper targets.

D. Reported Evaluations Rarely Distinguish Algorithmic Validation from System-Level Validation

Most of the literature surveyed in Section II reports a single headline accuracy or error-rate figure obtained on a specific benchmark dataset. Because we did not have access to a comparable labeled, multi-subject fingerprint/iris dataset or physical capture hardware in the development environment for this project, reporting a directly comparable population-scale accuracy figure would misrepresent what was actually measured. We instead treat this as a methodological gap worth addressing directly: Section V separates algorithmic-level validation (controlled synthetic and perturbed imagery, exercising the matching and quality-scoring logic itself) from system-level validation (automated, browser-driven tests of the complete officer-supervised workflow, including the human-confirmation checkpoint from gap B above), and we report both honestly rather than conflating the two into a single number.

IV. PROPOSED METHODOLOGY

The system, SecureVote Biometrics, is implemented as a Flask web application with a SQLite persistence layer, following a three-tier architecture (Fig. 1): a browser client that captures images and drives a multi-step officer-facing wizard, a REST API server that owns all biometric processing, and a database that stores only encrypted templates and non-reversible metadata. No biometric processing occurs client-side; the browser transmits only captured frames and receives only similarity scores, quality metrics, and rendered preview images in return.

Fig. 1.  System architecture. All biometric computation and encryption occur server-side; the client transmits only captured frames and receives scores, quality metrics, and rendered previews.

Fig. 2.  Fingerprint processing pipeline. Enhancement (CLAHE, median filter, Gabor bank) precedes minutiae extraction and descriptor computation.

Fig. 3.  Iris processing pipeline. Detection and localization are detector-agnostic with respect to the downstream quality-scoring, feature-extraction, and matching stages.

Fig. 4.  Fingerprint matching scores: self, genuine (same thumb, perturbed), and impostor (different thumb) trials. Dashed lines mark the 70/85/95 decision-tier boundaries.

C. Iris Localization Accuracy

Fig. 5.  Iris template matching scores: self, genuine (same synthetic iris, blurred), and impostor (different synthetic texture) trials.

Three observations follow. First, only Rao et al. [23] among the compared systems specifies template protection explicitly comparable in intent to ours, though at the cost of a third enrolled modality (face) that our two-modality design avoids, trading a small amount of theoretical accuracy headroom [23] for lower enrollment burden consistent with the field-failure-rate motivation for the fingerprint-plus-iris pairing specifically identified in [4]. Second, none of the five compared systems describes what is displayed to an operator during live capture; our blurred-preview design (Section IV-B1) is, to our knowledge, not previously reported in this specific literature thread. Third, none describes a human-confirmation checkpoint distinguishing a supervised polling-place deployment from a self-service kiosk; our officer-confirmation workflow (Section IV-E) and the defect it was designed to prevent (Section VI-E, T-14) are, again to our knowledge, not previously discussed in this literature.

Against the algorithmic literature specifically, our fingerprint pipeline occupies a deliberately more classical position than the deep-learning minutiae extractors in [12, 13]: we trade the higher benchmark accuracy those methods report for a CPU-only, no-training-data, fully offline implementation, following the same trade-off analysis as the Haar-cascade-versus-deep-learning face-detection comparison in [29] (Section II-G), which we treat as the more relevant precedent for our operating constraints than raw accuracy leaderboards. Our iris texture representation (LBP plus ORB) is lighter-weight than a full Gabor-phase IrisCode, trading some matching precision — visible in the modest genuine/impostor margin in the iris rows of Table II relative to the wider separations typically reported for IrisCode-based systems — for substantially lower computational cost, appropriate to the same CPU-only constraint.

VIII. DISCUSSION

A. Interpretation of Results

The algorithmic results (Section VI-A, VI-D) show correct directional discrimination  genuine trials consistently outscore impostor trials on both modalities  but with genuine-trial combined scores landing in the retry rather than accepted tier for both fingerprint and iris. We read this primarily as a property of our synthetic perturbation magnitude rather than as evidence the matchers are miscalibrated: real same-session recapture variation for an inked thumbprint or a webcam iris is plausibly gentler than a 4° rotation with independent pixel noise, or a full Gaussian blur pass, applied deliberately to stress-test discrimination. The more actionable finding is architectural: in both cases, the tiered decision system (Section IV-C) routed a borderline-but-genuine capture to a retry rather than a false reject, which is the intended behavior of a four-tier design and would not have been available with a single binary threshold.

The quality-metric validation (Section VI-B) and localization-accuracy results (Section VI-C) are, we believe, the strongest evidence in this paper, precisely because they test narrower, more mechanically verifiable claims does a specific metric respond to a specific targeted degradation; how far off-center is a specific localization estimate than the harder problem of population-scale genuine/impostor separation. The system-level results are similarly narrow in scope (software correctness of the supervised workflow, not biometric accuracy) but directly support the paper's central design claim: that an officer-supervised checkpoint changes measurable system behavior, since the specific defect it fixes was itself only detectable through this kind of end-to-end, browser-driven testing rather than unit-level API testing.

B. Limitations

We state these plainly rather than deferring them to a single closing paragraph.

(1) Detector precision: the Haar-cascade-plus-Hough-circle iris detector (Section IV-B) is measurably less precise than landmark-based detection under real, uncontrolled lighting [29, 30], and we could not validate this gap directly without live capture hardware.

(2) Liveness robustness: blink-based liveness defeats static-photo presentation attacks but not video replay [19, 20], and our EAR proxy is a bounding-box approximation rather than a landmark-accurate measurement.

(3) Threshold calibration: the 95/85/70 decision tiers, the 98% duplicate threshold, and the fusion weights are principled defaults consistent with the literature reviewed in Section II, not values fitted to a labeled dataset of genuine and impostor comparisons; production deployment requires calibration against real biometric data with reported false-accept/false-reject curves.

(4) Key management: the AES-256 master key is generated once and stored locally for this reference implementation; production use requires a managed key-management service or hardware security module with rotation. (5) Scale: the duplicate-enrollment check (Section IV-A5) is a linear scan appropriate to precinct scale, not the national scale addressed in [9].

C. Implications for Deployment

The officer-confirmation design (Section IV-E) is, we argue, the right default specifically for a supervised polling-place deployment and would be the wrong default for an unsupervised self-service kiosk, where requiring a human confirmation click at every step has no one present to provide it. This is not a limitation of the confirmation design so much as a scope boundary: the system as described targets the supervised setting, and an unsupervised variant would need a different resolution to the same underlying question what happens on a borderline or failed capture that does not depend on an available human operator.

CONCLUSION

This paper presented the design, implementation, and evaluation of a two-modality fingerprint-and-iris biometric authentication system for officer-supervised voter verification. The system combines a classical two-stage fingerprint matcher, a privacy-preserving iris capture pipeline that never displays or stores more of the face than the iris region, AES-256-GCM template encryption, score-level fusion with a four-tier decision system, precinct-scale duplicate-enrollment detection, and the paper's central design contribution an explicit officer-confirmation checkpoint that replaces automatic capture-to-decision advancement with a required human action on every registration and, on authentication, offers retry rather than any automated path forward whenever a capture fails to match. Reviewing 2021–2026 literature across fingerprint and iris recognition, liveness detection, multimodal fusion, template protection, and applied biometric voting, we identified that capture-time privacy and human-supervision checkpoints are addressed rarely or not at all in comparable systems, and positioned this system's design directly against that gap.

Evaluation combined algorithmic validation on controlled synthetic and perturbed biometric imagery with system-level validation of the complete application through automated browser interaction, deliberately avoiding a fabricated population-scale accuracy figure given the absence of a labeled dataset and live capture hardware in the development environment. Results showed correct self-match saturation, correct directional genuine/impostor separation for both modalities, sub-2-pixel mean iris-localization error, correctly independent responses of all five iris quality sub-metrics to targeted degradations, and a fully passing final system-level regression suite reached only after identifying and correcting two real defects, including the auto-advance-on-non-match behavior that directly motivated the officer-confirmation redesign, reported here transparently as part of the evaluation rather than omitted.

Future work includes replacing the Haar-cascade/Hough-circle iris detector with landmark-based detection where model-download network access is available; calibrating decision thresholds against a real, labeled, multi-subject biometric dataset with reported false-accept/false-reject curves; strengthening liveness detection against video-replay attacks; integrating a managed key-management service; and extending enrollment to both eyes and additional fingers for higher-assurance deployments.

Funding

The authors received no specific funding for this research.

Data Availability

The datasets generated and/or analyzed during the current study are not publicly available due to biometric data privacy considerations. Processed experimental data and supporting information may be made available by the corresponding author upon reasonable request, subject to applicable privacy and ethical considerations.

REFERENCES

1. Sumalatha et al., “Comprehensive Review of Unimodal and Multimodal Fingerprint Biometric Authentication Systems: Fusion, Attacks, and Template Protection,” IEEE Access, 2024.
2. R. Es-Sobbahi et al., “A systematic review of multimodal biometric fusion levels and strategies,” IET Biometrics, 2025.
3. J. Kamlaskar et al., “Iris and fingerprint multimodal biometric fusion for identity authentication,” AIMS Electronics and Electrical Engineering, 2021.
4. B. Umar et al., “A Bi-Factor Biometric Authentication System for Secure Electronic Voting System,” in Proc. IEEE Nigeria 4th Int. Conf. Disruptive Technologies for Sustainable Development (NIGERCON), 2022.
5. M. Zamir et al., “Electronic Voting Machine with Biometric Authentication,” in Proc. INDIACom, 2022.
6. K. Ajao et al., “Blockchain-Based Multimodal Biometric and Deep Learning Framework for Secure Electronic Voting,” IEEE Access, 2025.
7. M. A. Faruk et al., “Blockchain and biometric based secure online voting system,” Cluster Computing, 2024.
8. C. Nlend et al., “Large-Scale Fingerprint Deduplication for Voter Registration Databases,” arXiv preprint, 2021.
9. S. Koner et al., “Bharat ABIS: Billion-Scale Multimodal Biometric Search and Deduplication,” 2026.
10. R. Martins et al., “Fingerprint Recognition in Forensic Scenarios: A Crossing-Number and Gabor-Enhancement Approach,” Sensors, 2024.
11. T. Situmorang, “Crossing Number Method for Mobile-Camera-Captured Fingerprint Biometrics,” 2022.
12. Q. Zhang et al., “A multi-task deep convolutional neural network for contactless fingerprint minutiae extraction,” Pattern Recognition, 2021.
13. A. Abdulkarim et al., “Hybrid Deep Learning and Classical Enhancement Pipeline for Fingerprint Identification,” 2025.
14. P. Manchanda et al., “CLAHE and Gaussian Smoothing for Latent Fingerprint Image Enhancement,” 2025.
15. M. Taha et al., “Local Binary Pattern-Based Iris Feature Extraction Across Near-Infrared and Visible Spectra,” 2021.
16. S. Jan et al., “Viola-Jones and Geometrical Information Based Eye/Iris Segmentation for Non-Ideal Iris Recognition,” 2021.
17. E. Garea-Llano et al., “Iris-Pupil Region Quality Assessment Framework Aligned with ISO/IEC 19794-6,” 2021.
18. Z. Fazilov et al., “PCA-Based Combination of Iris Image Quality Parameters,” 2022.
19. E. Demirsoy et al., “Eye-Aspect-Ratio-Based Blink Liveness Detection for Access Control,” 2026.
20. H. Chong et al., “Eye Aspect Ratio Blink Detection for Masked Face Recognition Liveness Verification,” 2025.
21. V. C. Kulloli et al., “Iris Liveness Detection for Video Replay Attack Using Quality Metrics and EfficientNetB7,” in Proc. IEEE Int. Conf. Data and Software Engineering (ICoDSE), 2025.
22. A. Kharchevnikova and A. Savchenko, “Best-frame selection for video-based face recognition,” PeerJ Computer Science, 2021.
23. P. Rao et al., “DeepFusion: Tri-Modal Privacy-Preserving Biometric Fusion of Face, Fingerprint, and Iris,” 2026.
24. M. Imran et al., “AES-Encrypted Minutiae Cylinder Code Template Protection for Fingerprint Biometrics,” IEEE Access, 2025.
25. H. Ayoup et al., “AES-Based Cancelable Template Generation for Combined Eye and Fingerprint Biometrics,” Intelligent Automation & Soft Computing, 2022.
26. W. El-Shafai et al., “Genetic-Algorithm-Optimized Cancelable Biometric Encryption Scheme,” IEEE Access, 2021.
27. M. Bansal et al., “A comparative analysis of SIFT, SURF, and ORB feature descriptors,” 2021.
28. L. Xie et al., “Adaptive Histogram Equalization Combined with ORB Feature Matching,” 2022.
29. S. Adeshina et al., “Custom Face Classification Model for Classroom Using Haar-Like and LBP Features with Their Performance Comparisons,” Electronics, vol. 10, no. 1, 2021.
30. K. Elbert et al., “Comparative Evaluation of Haar Cascade, Dlib, and MediaPipe for Face and Eye Detection,” 2025.