PaperlayerView exact revision on arXiv

Source-linked AI summary

Biometrics: Trust, but Verify

Anil K. Jain, Debayan Deb, Joshua J. Engelsma

arXiv:2105.06625v2cs.CV

TL;DR

Biometric recognition still faces unresolved problems that leave its trustworthiness unsettled. This paper surveys concerns across the recognition pipeline and identifies areas requiring further work to build more trustworthy systems.

  • Problem

    Unsolved problems in performance robustness and scalability, security, explainability, and bias remain before biometric recognition can be accepted as trustworthy.

    Source: §7 · p. 14

  • Method

    The paper surveys research addressing major concerns across the biometric recognition pipeline, discusses remaining limitations, and summarizes recommended steps and research directions.

    Source: §1 · p. 3

  • Results

    The authors identify five major research areas requiring further work to establish trustworthiness in biometrics.

    Source: §7 · p. 14

  • Takeaways & Limitations

    The paper recommends rigorous, fair research and further steps toward more trustworthy biometric recognition systems.

    Source: §1 · p. 3

    Demographic bias remains unresolved, with research needed beyond face biometrics and on PAD modules while preserving overall system accuracy.

    Source: §5 · p. 12

Abstract

from arXiv · show

Over the past two decades, biometric recognition has exploded into a plethora of different applications around the globe. This proliferation can be attributed to the high levels of authentication accuracy and user convenience that biometric recognition systems afford end-users. However, in-spite of the success of biometric recognition systems, there are a number of outstanding problems and concerns pertaining to the various sub-modules of biometric recognition systems that create an element of mistrust in their use - both by the scientific community and also the public at large. Some of these problems include: i) questions related to system recognition performance, ii) security (spoof attacks, adversarial attacks, template reconstruction attacks and demographic information leakage), iii) uncertainty over the bias and fairness of the systems to all users, iv) explainability of the seemingly black-box decisions made by most recognition systems, and v) concerns over data centralization and user privacy. In this paper, we provide an overview of each of the aforementioned open-ended challenges. We survey work that has been conducted to address each of these concerns and highlight the issues requiring further attention. Finally, we provide insights into how the biometric community can address core biometric recognition systems design issues to better instill trust, fairness, and security for all.

1 INTRODUCTION

Biometric recognition offers convenient person identification but must be shown trustworthy beyond accuracy, addressing robustness, fairness, security, explainability, and privacy. This paper surveys those concerns, the work addressing them, remaining limitations, and steps toward more trustworthy systems.

1 INTRODUCTION

  • Biometrics identifies people through physical or behavioral traits, offering an alternative to passwords and tokens that can be vulnerable to theft, deception, or lack of access.The paper defines biometrics as automatic person recognition based on traits such as face, fingerprint, and iris.Source: §1 · p. 2
  • NIST search systems report FNIR = 0.001, 0.058, and 0.0059 at FPIR = 0.001 for fingerprint, face, and iris, respectively.The authors argue that accuracy and convenience alone are insufficient because systems can still err and face attacks.Source: §1 · p. 2, Table 1 · p. 2
  • Trust concerns span robustness to noise and change, demographic fairness, spoofing and other attacks, explainability, and protection of user and training-data privacy.The questions include performance over time and at scale, template theft or reconstruction, and decentralized or encrypted training and inference.Source: §1 · p. 3, Figure 2 · p. 3
  • The paper systematically surveys research addressing these concerns, identifies remaining limitations, and recommends rigorous and fair steps for building more trustworthy biometric systems.The authors frame this as verifying systems’ truth, accuracy, or validity, while noting that more work remains [25].Source: §1 · p. 3
  • Unlike surveys focused on individual topics, this paper brings the components of trustworthy biometrics together in one overview for researchers.The authors describe it as the latest and most comprehensive overview of the field’s state.Source: §1 · p. 3

2 RECOGNITION PERFORMANCE ROBUSTNESS AND SCALABILITY

Trustworthy biometric recognition requires accuracy that withstands noisy inputs and scalability to very large user populations. This section reviews robustness and scalability challenges across face, fingerprint, and iris systems.

2 RECOGNITION PERFORMANCE ROBUSTNESS AND SCALABILITY

  • Recognition rates for fingerprints, face, and iris have significantly improved over time, and these traits are widely adopted compared with other biometrics.The section focuses on face, fingerprint, and iris as the three most popular and widely accepted traits.Source: Figure 3 · p. 4, §2 · p. 4
  • Biometric pipelines span acquisition, feature extraction, and matching across enrollment and recognition, so breakdowns at any module can undermine robustness and scalability.Systems operate in authentication (1:1 verification) or search (1:N identification); representations should separate different users while remaining similar for the same user.Source: §2 · p. 4

2.1 Noisy Inputs

  • Poor image quality and adverse conditions impair recognition, while face matching also degrades with age and longer enrollment-to-query intervals.Face recognition is affected by pose, illumination, and expression; fingerprints by distortion and debris; and iris recognition by occlusion and reflections. Face matchers can deteriorate after 10–12 years, with faster decline for younger individuals, whereas iris and fingerprint performance has remained stable over large time lapses.Source: §2.1 · pp. 4–5, Figure 4 · p. 5

2.2 Training Data

  • Privacy concerns constrain large labeled biometric datasets, and underrepresented subjects pose another challenge; unlabeled data offers routes to expand training and diversity.Semi-supervised approaches use label propagation or graph-clustered pseudolabels, while heterogeneous unlabeled data can improve generalizability to challenging images.Source: §2.2 · p. 5

2.3 Scalability

  • Despite deployments at billion-person scale, few studies evaluate biometric systems at the scale of major real-world programs, where errors could cause substantial disruption.Aadhaar has over 1.3 billion enrollees and averages 35 million biometric authentications per day.Source: §2.3 · p. 5
  • Large-gallery search evaluations may overestimate scalability when synthetic fingerprints or web-scraped faces differ from real probes, motivating domain-bridging biometric synthesis.The section identifies synthesis that bridges real–synthetic domain gaps as a way to better estimate both accuracy and speed at scale, while addressing privacy concerns.Source: §2.3 · p. 6

3 SECURITY

Biometric security threats span the sensor, feature-extraction, database, and matching stages. The section reviews presentation, adversarial, and template attacks, emphasizing unresolved generalization and the need for defenses that protect the full pipeline.

3 SECURITY

  • Trustworthy biometric security must address vulnerabilities across the pipeline, especially defenses’ limited generalization to new attack types and sensors.Mega-scale galleries of real data would be preferable for evaluation, but privacy concerns and the time and cost of collecting such datasets make them difficult to obtain.Source: §3 · p. 6

3.1 Presentation Attacks

  • Fingerprint PAD accuracy can fall below 10% when tested on spoof materials excluded from training, despite near-perfect results for known materials.PAD performance can also deteriorate across datasets and sensors as sensor models, illumination, demographics, and environmental conditions change; deep-learning solutions may be impractical on resource-constrained devices. Methods targeting unseen materials and sensors remain insufficient for field deployment.Source: §3.1 · p. 7

3.2 Adversarial Attacks

  • ArcFace’s TAR on LFW falls from 99.82% to 00.17% at 0.1% FAR under AdvFaces, whose subtle perturbations can enable impersonation or privacy obfuscation without attacker participation.Adversarial faces attack the feature-extraction module and are difficult to detect because their perturbations are extremely subtle.Source: §3.2 · p. 8
  • Adversarial training, perturbation detectors, and purification have trade-offs: training can reduce real-image accuracy, defenses may fail on unseen perturbations, and attacks also threaten fingerprint and iris systems.The paper calls for improving intrinsic robustness to avoid separate detectors or purifiers; adversarial attacks have also been shown to evade fingerprint PAD and iris matchers, and can be converted into physical attacks on PAD.Source: §3.2 · p. 8

3.3 Template Attacks

  • Biometric templates can be inverted into high-fidelity images and encode demographic attributes, so they must be secured both in storage and during matching.Templates may be stolen after feature extraction from the enrollment database or during matching if they are decrypted.Source: §3.3 · pp. 8–9
  • Existing template-protection methods trade recognition accuracy for security, while homomorphic encryption protects templates in storage and during comparison but remains computationally expensive.Encrypted fingerprint search against a 100 million-entry gallery took 500 seconds, a 275× speedup over the state of the art; the unencrypted search took 10 seconds. The paper calls for minimizing trade-offs in speed, memory, accuracy, and security.Source: §3.3 · p. 9

3.4 Unifying Security Efforts

  • A single generalizable detector could screen for presentation, adversarial, and template attacks during enrollment, reducing the computational burden of securing the full pipeline.The paper recommends this approach because the attack type may not be known in advance.Source: §3.4 · p. 9

4 EXPLAINABILITY AND INTERPRETABILITY

Biometric systems need interpretable decisions so designers, operators, and courts can understand and, where needed, adjust black-box outputs. The section surveys interpretability approaches and emphasizes unresolved questions about what systems encode and how to explain decisions across the recognition pipeline.

EXPLAINABILITY AND INTERPRETABILITY

  • Interpretability can help designers and deploying agencies understand and adjust decisions, and is important when biometric evidence is used in court because deep face-recognition models often act as black boxes.Understanding why probe and gallery faces appear similar may help prevent false convictions or false acquittals.Source: §4 · p. 9
  • Interpretability methods visualize learned features [225] or attribute decisions to features [229] and image regions [231] [235]; attribution may suit biometric classification tasks such as PAD.Feature visualization can apply to feature extraction modules, whereas decision attribution is more oriented toward explaining a final classification decision.Source: §4 · pp. 9–10
  • Yin et al. linked learned face features to facial areas, while Stylianou et al. visualized salient areas but did not provide fine-grained similarity attribution, which remains unsolved.Yin et al.'s approach requires retraining, unlike the model-agnostic visualization proposed by Stylianou et al.; spatial-diversity regularization also improved robustness to occlusion.Source: §4 · p. 10
  • Face studies found deep representations more correlated with identity and age than gender and face angle, while accessories, hairstyles, colors, face shapes, and anomalies can influence recognition performance.The attribute analyses examined identity, age, gender, and yaw, and separately assessed 47 high-level attributes.Source: §4 · p. 10
  • Shi et al. represented faces as feature-space distributions whose variance expresses uncertainty, using that uncertainty both to improve recognition performance and visualize model perception of inputs.Source: §4 · p. 10
  • Further work is needed to understand what deep face representations encode, address bias and system failures, and extend interpretability to other face-recognition pipeline modules such as spoof detection.Source: §4 · p. 10
  • Fingerprint networks have been found to learn discriminative minutiae and singularity regions [55, 239], while spoof-detection systems can visualize suspected spoof regions for human review [241].Fingerprint matcher studies [49] also examined which noise sources most affect recognition decisions; a global spoofness score alone may not suffice for manual inspection.Source: Figure 10 · p. 10, §4 · p. 11

5 DEMOGRAPHIC BIAS AND FAIRNESS

Demographic bias is a trust concern because biometric systems can perform differently across groups, and both algorithm and sensor factors can contribute. The section reviews mitigation approaches and calls for continued work across modalities and system components while maintaining overall accuracy.

5 DEMOGRAPHIC BIAS AND FAIRNESS

  • Demographic bias is unequal recognition performance across groups; all 106 NIST FRVT face algorithms showed gender-, race-, and age-linked differences.Best-performing commodity AFR bias was less than 1.0% across Black Male, Black Female, White Male, and While Female groups; manual ground-truth adjudication flipped group performance levels, while top-tier systems were most accurate on Black Males.Source: §5 · p. 11
  • Studies report worse face-recognition performance for female cohorts, age-related effects in fingerprint and iris, and difficulty recognizing infants and young children.Studies also found attribute-estimation accuracy impacts based on age, gender, and race.Source: §5 · p. 11
  • Bias can also arise from sensors: fingerprint matching across sensors is challenging, brown-eyed people are more susceptible to iris sensor issues, and measured AFR effects depend on image acquisition.Near-infrared sensors are adopted instead of RGB cameras for iris recognition.Source: §5 · pp. 11–12
  • Biased systems can contribute to unjust incarceration or denial of bail or parole; mitigation should eliminate statistically significant group performance differences without compromising overall accuracy.Ideally, overall system accuracy should not be compromised.Source: §5 · p. 12
  • Bias can stem from demographically skewed training data, learning choices that underweight under-represented samples, or intrinsic recognition difficulty for some groups.Mitigations include demographic resampling [269] and cost-sensitive loss reweighting [270], though resampling discards diverse information; disentangled representations can lower overall performance, while adaptive margins and dynamic convolution and attention improve under-represented-group performance while better maintaining overall accuracy.Source: §5 · p. 12
  • Bias remains unresolved; research should address other sources and modalities such as fingerprint, preserve overall accuracy, and examine demographic bias in presentation-attack detection modules.The authors describe reducing bias as a critical step toward trustworthy biometric recognition.Source: §5 · p. 12

6 PRIVACY

Privacy is distinct from system security: it concerns confidentiality and transparent use and storage of personal data. The section reviews legal protections and privacy-preserving approaches while identifying their tradeoffs.

6 PRIVACY

  • Unlike security, which addresses attacks on biometric systems, privacy concerns confidentiality of personal identifying information and transparency about its use and storage.Source: §6 · p. 12
  • BIPA requires written notice of collected data and its purpose and retention period, plus written consent; GDPR treats biometric data as personal data and sets processing principles.The GDPR principles include lawfulness, fairness, transparency, and purpose limitation.Source: §6 · pp. 12–13
  • Privacy concerns include biometric images collected from social media without informed consent, dataset retractions, and evidence that GANs can leak information about their training data.Clearview AI was facing litigation over allegedly amassing about 3 billion face images without subjects’ permission.Source: §6 · p. 13
  • Homomorphic encryption enables inference on encrypted data with encrypted model parameters, but its computational burden limits practical model size.Source: §6 · p. 13, Figure 15 · p. 13
  • Federated learning improved CosFace TAR from 81.43% to 83.79% on IJB-A at 0.1% FAR and accuracy from 99.15% to 99.28% on LFW, but underperformed centralized training.Training images remain on participating devices; only local updates are shared with the server for aggregation. The reported setup used face images from 1,000 mobile devices.Source: §6 · pp. 13–14
  • The authors call for further work on encrypted inference and decentralized training to reduce privacy tradeoffs and advance trustworthy biometric recognition.Source: §6 · p. 14

7 CONCLUSION

Biometric recognition is increasingly capable and may surpass human accuracy under certain conditions, but unresolved challenges must be addressed for systems to be trustworthy. The paper identifies five research areas and proposes a Grand Challenge to evaluate and certify trustworthiness.

7 CONCLUSION

  • The paper identifies performance robustness and scalability, security, explainability and interpretability, biasness and fairness, and privacy as five areas requiring further work to establish trust in biometrics.It defines each problem, explains its importance, surveys existing work, and suggests further research; addressing them could make systems both accurate and trustworthy for researchers, users, and policy makers.Source: §7 · p. 14
  • Biometric recognition is well equipped for automatic person identification and can surpass human recognition performance under certain conditions, yet many unsolved problems remain before it can be accepted as trustworthy.Applications include national ID cards, border crossings, access control, and payments.Source: §7 · p. 14
  • A proposed Grand Challenge on Trustworthy Biometrics could evaluate systems across the five categories and certify those meeting quantitative thresholds, so end-users know how trustworthy a system is as well as how accurate it is.The paper suggests that a government agency such as NIST could host the challenge.Source: §7 · p. 14
Loading 2105.06625v2…