Source-linked AI summary
Demographic Bias in Biometrics: A Survey on an Emerging Challenge
P. Drozdowski, C. Rathgeb, A. Dantcheva, N. Damer, C. Busch
TL;DR
Demographic bias in biometric systems is an emerging technical and social challenge affecting automated recognition and classification. The article surveys biometric bias estimation and mitigation literature, discusses technical and social matters, and outlines open challenges. It reports that demographic factors can substantially influence biometric algorithms, with recurring lower performance for females and youngest subjects and lower classification accuracy for dark-skinned females in some recognition settings.
Problem
Biometric systems face growing concerns about systemic demographic bias, while the technical, social, and fairness issues surrounding that bias remain insufficiently explored.
Method
The article provides an overview of biometric algorithmic bias, surveys bias estimation and mitigation literature, and discusses technical, social, and future-work issues.
Results
Demographic factors can strongly influence biometric algorithms, with recurring lower performance for females and youngest subjects and lower classification accuracy for dark-skinned females in some recognition systems.
Takeaways & Limitations
Biometric bias assessment requires attention to algorithmic, dataset, application, and social-impact factors, alongside fairness, transparency, accountability, and explainability considerations.
Abstract
from arXiv · showhide
Systems incorporating biometric technologies have become ubiquitous in personal, commercial, and governmental identity management applications. Both cooperative (e.g. access control) and non-cooperative (e.g. surveillance and forensics) systems have benefited from biometrics. Such systems rely on the uniqueness of certain biological or behavioural characteristics of human beings, which enable for individuals to be reliably recognised using automated algorithms. Recently, however, there has been a wave of public and academic concerns regarding the existence of systemic bias in automated decision systems (including biometrics). Most prominently, face recognition algorithms have often been labelled as "racist" or "biased" by the media, non-governmental organisations, and researchers alike. The main contributions of this article are: (1) an overview of the topic of algorithmic bias in the context of biometrics, (2) a comprehensive survey of the existing literature on biometric bias estimation and mitigation, (3) a discussion of the pertinent technical and social matters, and (4) an outline of the remaining challenges and future work items, both from technological and social points of view.
I. INTRODUCTION
Automated decision systems increasingly assist or replace humans in complex tasks, offering efficiency and cost benefits while raising ethical and legal concerns about transparency.
- Automated algorithms support learning, problem solving, classification, prediction, and risk assessment.
- These systems can outperform humans and are therefore used to support or replace human operators.
- Automated decision systems may increase efficiency and decrease monetary costs.
- Their use has also prompted ethical and legal concerns, particularly regarding transparency.
A. Motivation
Biometric technologies are rapidly expanding across governmental and commercial identity applications, while reports of demographic bias have intensified debates about their use and limits.
- Biometric systems are widely used for border control, law enforcement, forensics, voter registration, and national identity management.
- Investment in biometric technologies is large and rapidly growing.
- Reports of demographically unfair biometric systems have fueled debate among public, governmental, academic, commercial, and advocacy stakeholders.
- Some stakeholders have demanded or considered discontinuing biometric applications until privacy protection and demographic bias mitigation are sufficient.
B. Article Contribution and Organisation
The article surveys demographic bias in biometric systems, covering estimation and mitigation approaches alongside technical, social, and future challenges.
- The article presents an overview of algorithmic bias and fairness in biometric systems.
- It surveys biometric algorithms susceptible to bias and existing approaches to bias estimation and mitigation.
- The article discusses potential social impacts of demographic bias in biometric systems.
- Remaining challenges and open issues are addressed alongside the article’s background, literature survey, and concluding summary.
II. BACKGROUND
The background introduces general algorithmic bias and biometric-system basics, then explains the article’s terminology choices for this sensitive topic.
- The background covers bias in automated decision systems generally and the basics of biometric systems.
- It also explains the nomenclature used throughout the article because of the topic’s sensitive nature.
A. Bias in Automated Decision Systems
Automated decision systems can produce biased predictions through human, data, implementation, and interaction-related factors, with potentially serious effects on affected individuals.
- Systemic bias has been identified in automated risk-assessment and welfare-distribution predictions, including discrimination against dark-skinned people.Such decisions can affect outcomes including jail, bail, parole, and welfare access.
- Human designers and operators can directly contribute to bias in automated decision systems.Semiautomatic systems illustrate this risk because human decision makers use algorithmic assistance.
- Training data can propagate bias when it is skewed, incomplete, outdated, disproportionate, or historically biased.These data problems are described as detrimental to algorithm training.
- Bias can also arise from flawed algorithm implementation and from human interactions with automated systems.The cited implementation-related examples include moral or legal norms and poor design.
B. Biometric Systems
Biometric systems automate identity or demographic-attribute recognition from distinctive biological or behavioural characteristics. Their pipeline combines sample capture, storage, feature extraction, comparison, and decision algorithms, increasingly using machine and deep learning.
- Biometrics are automated recognition of individuals based on biological and behavioural characteristics.Biometric systems can establish or verify identity or demographic attributes.
- Distinctive physiological characteristics enable individuals to be distinguished with high confidence.The paper illustrates prominent biometric characteristics with example images in Figure 1.
- A biometric system captures samples, stores biometric and personal data, extracts distinguishing features, and compares feature vectors for decisions.The described components include a capture device, database, signal-processing algorithms, and comparison and decision algorithms.
- Machine learning and deep learning have become increasingly popular and successful in biometric systems, following earlier handcrafted features and algorithms.The paper cites breakthrough facial-recognition performance and promising deep-learning results for fingerprint and iris recognition.
C. Nomenclature
The paper reports that demographic terminology in biometric research is often conflated or coarse, adopts specified ISO/IEC distinctions for its survey vocabulary, and still treats bias study as necessary because impacts can be disparate.
- The authors report the divisive historical, cultural, social, political, and legislative load of demographic terms and do not seek to redefine them.They instead describe the terminology used in current research.
- Surveyed literature often conflates “gender” with “sex” and “race” with “ethnicity,” while using fewer than ten coarse racial or ethnic categories.These trends are identified as common practices in the literature.
- The paper follows ISO/IEC distinctions that define gender socially or behaviourally, sex biologically, and ethnicity through common origin, customs, or traditions.The surveyed literature mainly uses appearance-based categorisation, leading the paper to use “race” and “sex” in accordance with the cited standards.
- Despite terminological limitations, the paper considers demographic bias and fairness study imperative because automated decisions, including biometrics, can have real and disparate impacts.
III. BIAS IN BIOMETRIC SYSTEMS
The paper distinguishes demographic effects in biometric score distributions from differences in thresholded decision outcomes. It then frames its literature survey around affected algorithms, covariates, bias estimation, and mitigation.
- Differential performance denotes differences in genuine or impostor score distributions between demographic groups.The concept is presented as analogous to demographic differences in score distributions associated with the biometric menagerie.
- Differential outcomes denote demographic differences in false-match and false-non-match rates at a specific decision threshold.
- Most surveyed studies use ad hoc methodologies based on existing metrics rather than directly using these recently introduced terms.The paper notes that Grother et al. conducted a comprehensive benchmark using the terms and notions.
- The survey covers potentially affected biometric algorithms, relevant covariates, bias estimation literature, and bias mitigation literature.These topics are organized across subsections III-A through III-D.
A. Algorithms
Biometric systems process distinctive biological samples to recognize individuals, either by verifying claimed identities or searching databases for identities.
- Biometric pipelines use algorithms to process distinctive biological samples for recognition.
- Recognition systems perform either one-to-one verification or one-to-many database searches.
Classification and estimation Referring to the process of
Biometric systems acquire and preprocess samples before recognition or classification, while demographic, subject-specific, and environmental covariates can affect algorithm effectiveness.
- Acquisition and preprocessing precede biometric recognition or classification tasks.
- Segmentation locates the region of interest and extracts biometric features from a sample.
- Quality assessment quantifies the quality of an acquired biometric sample.
- Presentation attack detection automatically determines whether a presentation interferes with biometric data capture.
- Demographic covariates include sex, age, and race, whereas subject-specific and environmental covariates include pose, accessories, illumination, occlusion, and image resolution.The article concentrates on demographic covariates rather than environmental and subject-specific factors.
C. Estimation
The survey finds recurring demographic disparities across biometric algorithms, but emphasizes that results depend on algorithms, data, experimental setups, and fairness definitions. It also identifies limited coverage, small datasets, and unresolved technical and social challenges.
- Recognition: Many recognition studies report worse false-positive and false-negative performance for female subjects.
- Recognition: Race influences performance through an apparent other-race effect linked to software-development context and presumed training data, while very young subjects are especially challenging.
- Recognition: False-negative differentials usually vary by less than a factor of 3 across benchmarked recognition algorithms.
- Limitations and open challenges: Bias estimation in identification is non-trivial because screening-database composition can introduce and propagate acquisition, historical, and societal biases.
- Classification and estimation: Classification studies report substantially lower demographic-attribute accuracy for dark-skinned female subjects in numerous commercial facial algorithms.
- Interpretation: Biases are often algorithm-specific, and large relative error increases may have negligible absolute importance for highly accurate algorithms.
- Limitations and open challenges: Comprehensive independent benchmarks using varied fairness metrics remain lacking, while many studies use relatively small datasets and demographic coverage needs more detailed intersectional analysis.
- Social and technological responses: The survey recommends careful training-data selection, transparency and independent validation, legally defined accuracy thresholds, and due diligence by vendors.