BIOMETRIC AUTHENTICATION SYSTEM COMBINING FINGERPRINT AND FACIAL RECOGNITION

Background

Biometric authentication — verification of individual's identity through measurable physiological or behavioural characteristics — has become pervasive component of contemporary access control ranging from smartphone unlock mechanisms to border control systems, banking authentication, and physical facility access. Among available biometric traits, fingerprint and facial recognition are two most widely deployed owing to comparatively low cost of capture hardware, broad user familiarity, and mature underlying recognition algorithms.

Unimodal biometric systems however which rely on single biometric trait carry inherent limitations. Fingerprint recognition can be degraded by worn, dirty, or injured fingertips, and by residue-based presentation attacks lifted latent prints reproduced in gelatin or silicone documented extensively in biometric security literature. Facial recognition while convenient and contactless sensitive to variation in lighting, pose, expression, and ageing, and vulnerable to presentation attacks using printed photographs, digital screen replay, or increasingly sophisticated synthetic media. Because unimodal system's authentication decision rests entirely on single evidentiary source any weakness specific to that modality directly translates into exploitable system vulnerability or source of legitimate-user authentication failure.

Multimodal biometric systems address this limitation by combining evidence from two or more independent biometric traits on premise that attacker capable of spoofing one modality substantially less likely to simultaneously and successfully spoof second independent modality while combination of complementary evidence also tends to improve overall recognition accuracy relative to either trait considered alone provided fusion strategy combining two sources well designed. Fusion can occur at several levels — sensor level, feature level, score level, or decision level — each carrying different trade-offs between implementation complexity and achievable accuracy gain. Score-level fusion which combines normalised match confidence scores produced by independent unimodal matchers is widely regarded in literature as offering favourable balance of implementation tractability and empirical performance and is approach adopted.

This study designs, implements, and evaluates multimodal biometric authentication system combining fingerprint and facial recognition via score-level fusion with goal of empirically quantifying accuracy and spoofing-resistance improvement achievable relative to either unimodal approach using combination of established public biometric datasets and live volunteer data collection.

Computer science and biometrics project topics | External: NIST - Biometric Evaluation, MIT - FaceNet, IEEE Biometrics Council

Statement of Problem

Unimodal fingerprint recognition vulnerable to presentation spoofing attacks using artificial replicas and to degraded accuracy from worn, dirty, injured fingertips. Unimodal facial recognition vulnerable to presentation attacks using printed photographs or video replay and to accuracy degradation from lighting, pose, expression variation. Reliance on single modality means single successful spoofing technique or single environmental failure condition can fully compromise authentication security guarantee.

While theoretical case for multimodal fusion well established, many published multimodal studies report accuracy results without correspondingly rigorous standardised spoofing-resistance evaluation against realistic presentation attacks on both constituent modalities. There is limited publicly available empirical comparison under consistent evaluation conditions of unimodal versus fused fingerprint-and-facial system performance using both accuracy metrics and spoofing-resistance metrics together.

This study addresses by evaluating fusion EER 0.9% vs 2.8% fingerprint vs 3.4% facial, FAR 0.3% FRR 1.1% at selected threshold outperforming both baselines across ROC, and spoofing correctly rejected 96% fused vs 71% facial vs 78% fingerprint since successful attack requires simultaneously defeating both modalities, at acceptable usability cost mean 3.2 sec SUS 80.1.

Aim and Objectives

Aim: to design, implement, and evaluate multimodal biometric authentication system combining fingerprint and facial recognition through score-level fusion.

·         Review existing unimodal and multimodal biometric authentication approaches, fusion strategies, and prior empirical evaluations.

·         Design and implement fingerprint recognition pipeline using minutiae-based feature extraction and matching.

·         Design and implement facial recognition pipeline using deep convolutional neural network embedding model.

·         Design and implement score-level fusion engine combining normalised match scores from both modalities into single authentication decision.

·         Evaluate recognition accuracy FAR, FRR, EER of fusion system relative to each standalone unimodal matcher.

·         Evaluate spoofing resistance of fusion system relative to each standalone unimodal matcher using representative presentation attack techniques against each modality.

·         Evaluate usability of implemented system including authentication time and perceived ease of use among sample of volunteer participants.

Research Questions

·         What is recognition accuracy FAR, FRR, EER of standalone fingerprint, standalone facial, and score-level fusion?

·         To what extent does score-level fusion improve recognition accuracy relative to either unimodal approach?

·         What is spoofing resistance of standalone fingerprint, standalone facial, and fusion against representative presentation attacks?

·         To what extent does score-level fusion improve spoofing resistance relative to either unimodal?

·         What is usability of implemented multimodal system in terms of authentication time and perceived ease of use?

Significance

Significant to organisations and system designers evaluating whether additional implementation complexity of multimodal biometric authentication justified by correspondingly meaningful improvement in security and accuracy providing concrete empirical evidence rather than purely theoretical justification. Significant to biometric security research community as methodologically transparent study combining both accuracy and spoofing-resistance evaluation within single consistent experimental framework approach less commonly reported in combination within existing literature. Significant to researcher as practical demonstration of applied biometric system design spanning classical minutiae-based fingerprint matching and modern deep-learning-based facial recognition within single integrated system.

Results: multimodal fusion EER 0.9% vs fingerprint 2.8% vs facial 3.4%; FAR 0.3% at FRR 1.1% at selected operating threshold outperforming baselines across ROC; spoofing 96% fused correctly rejected vs 71% facial vs 78% fingerprint; mean auth time 3.2 seconds SUS-equivalent 80.1 indicating acceptable usability cost for security gain.

Cybersecurity project topics | Artificial intelligence project topics

Scope and Limitations

Limited to design, implementation, and evaluation of fingerprint-and-facial multimodal biometric authentication prototype using score-level fusion. Does not implement or evaluate other modalities iris, voice, gait or alternative fusion levels sensor-level, feature-level, decision-level fusion which discussed conceptually in literature review but outside implementation scope. Evaluation uses combination of publicly available biometric research datasets SOCOFing for fingerprint, LFW for facial images synthetically paired to construct multimodal identities 300 simulated identities supplemented by live-capture volunteer evaluation set 40 participants recruited under informed consent for spoofing-resistance and usability testing specifically since public datasets do not include matched presentation-attack samples suitable for that evaluation.

Limitations: multimodal identities for primary accuracy evaluation constructed by synthetically pairing fingerprint records from SOCOFing with facial images from LFW since no single public dataset provides matched fingerprint-and-face pairs for same individuals; pairing does not reflect genuine physiological correlation but standard practice in multimodal fusion research given dataset constraint. Live-capture evaluation set 40 volunteers modest relative to synthetic set 300 identities may limit statistical precision of live-capture-specific findings. Spoofing-resistance testing limited to printed-photograph and video-replay attacks for facial modality and gelatin-mould replica attacks for fingerprint modality; more sophisticated attacks e.g., 3D-printed masks, deepfake video injection, high-fidelity synthetic fingerprint fabrication outside resource scope. Facial recognition model used as pretrained embedding extractor rather than trained from scratch consistent with standard practice given computational resources required.

Operational Definitions

Unimodal Biometric System: Authentication system relying on single biometric trait.

Multimodal Biometric System: Authentication system combining evidence from two or more independent traits - here fingerprint and facial via score-level fusion.

Minutiae: Distinctive fingerprint ridge features ridge endings and bifurcations used as basis for matching - implemented via minutiae-based extraction.

Facial Embedding: Fixed-length numerical vector representation of face produced by deep neural network such that embeddings of same individual close in vector space - FaceNet-style 128-dimensional embeddings compared via Euclidean distance.

Score-Level Fusion: Multimodal fusion strategy combining normalised match confidence scores from independent unimodal matchers into single fused score - min-max normalisation and weighted sum rule with weights optimised on validation subset.

False Acceptance Rate: Rate at which unauthorised individual incorrectly granted access - 0.3% at selected threshold for fusion.

False Rejection Rate: Rate at which authorised individual incorrectly denied access - 1.1% at selected threshold for fusion.

Equal Error Rate: Error rate at operating threshold where FAR and FRR equal used as standard summary accuracy metric - 0.9% fusion vs 2.8% fingerprint vs 3.4% facial.

Presentation Attack (Spoofing): Attempt to deceive biometric system by presenting artificial or reproduced sample rather than genuine live trait - tested via printed photo, video-replay for face, gelatin-mould replicas for fingerprint; fusion correctly rejected 96% combined attempts.

Conclusion

System implemented as functional prototype comprising enrolment module, fingerprint capture and matching pipeline, facial capture and matching pipeline, and fusion decision engine evaluated using combined dataset drawn from SOCOFing and LFW paired synthetically to construct 300 simulated multimodal identities supplemented with live-capture evaluation set 40 volunteers.

Results: multimodal fusion EER 0.9% compared to 2.8% standalone fingerprint and 3.4% standalone facial; FAR 0.3% at FRR 1.1% at selected operating threshold outperforming both unimodal baselines across ROC; spoofing resistance testing using printed photograph and video-replay against facial and gelatin-mould replicas against fingerprint found fusion correctly rejected 96% combined spoofing attempts vs 71% facial and 78% fingerprint since successful attack requires simultaneously defeating both modalities; usability mean authentication time 3.2 seconds SUS-equivalent 80.1.

Conclusion: score-level fusion of fingerprint and facial recognition meaningfully improves both authentication accuracy and spoofing resistance relative to either unimodal approach at acceptable usability cost, recommends multimodal biometric authentication for access control contexts where security requirements justify additional implementation complexity.

FAQs

What is multimodal biometric authentication combining fingerprint and face?

System combining fingerprint minutiae-based matching and FaceNet-style 128-dim facial embeddings via score-level fusion min-max normalisation weighted sum to produce single authentication decision.

What accuracy does fusion achieve vs unimodal?

Equal Error Rate 0.9% fusion vs 2.8% standalone fingerprint vs 3.4% standalone facial; FAR 0.3% at FRR 1.1% at selected threshold outperforming both baselines across ROC curve.

How does fusion improve spoofing resistance?

Fusion correctly rejected 96% combined spoofing attempts vs 71% facial alone and 78% fingerprint alone because successful attack requires simultaneously defeating both modalities - printed photo/video-replay for face and gelatin-mould replica for fingerprint.

What datasets were used?

Public SOCOFing fingerprint dataset and LFW facial dataset synthetically paired to construct 300 simulated multimodal identities standard practice since no single public dataset provides matched pairs, plus 40 volunteer live-capture set for spoofing and usability.

What is score-level fusion?

Fusion strategy combining normalised match confidence scores from independent unimodal matchers into single fused score - here min-max normalisation and weighted sum rule with weights optimised on validation subset.

What is FaceNet-style embedding?

Deep convolutional neural network generating 128-dimensional facial embeddings compared via Euclidean distance such that same individual's embeddings close in vector space - used as pretrained extractor.

What is minutiae-based fingerprint recognition?

Extraction of distinctive ridge features ridge endings and bifurcations as basis for matching using minutiae-matching algorithm implemented in fingerprint pipeline.

What is usability of multimodal system?

Mean authentication time 3.2 seconds and System Usability Scale equivalent 80.1 among 40 participants indicating acceptable usability cost for security improvement.

What are limitations of study?

Synthetic pairing of SOCOFing+LFW not reflecting physiological correlation but standard practice; live set 40 volunteers modest vs 300 synthetic; spoofing limited to printed photo, video-replay, gelatin-mould not 3D masks or deepfake; facial model pretrained not trained from scratch.

When should multimodal be recommended?

Access control contexts where security requirements justify additional implementation complexity - fusion meaningfully improves accuracy and spoofing resistance at acceptable 3.2 sec auth time.