biometric-dpia

Assessing Biometric Processing Privacy

Overview

Biometric data is classified as a special category of personal data under GDPR Art. 9(1) when processed for the purpose of uniquely identifying a natural person. Processing biometric data on a large scale triggers a mandatory DPIA under Art. 35(3)(b). This skill provides a comprehensive DPIA methodology for biometric systems including facial recognition, fingerprint identification, voice recognition, iris scanning, vein pattern analysis, and behavioural biometrics (gait, typing patterns, signature dynamics).

Legal Framework for Biometric Data

GDPR Definition — Art. 4(14)

"'Biometric data' means personal data resulting from specific technical processing relating to the physical, physiological or behavioural characteristics of a natural person, which allow or confirm the unique identification of that natural person, such as facial images or dactyloscopic data."

Art. 9(1) — Prohibition on Processing Special Categories

Processing of biometric data for the purpose of uniquely identifying a natural person is prohibited unless one of the Art. 9(2) exemptions applies.

Critical distinction: Art. 9 only applies when biometric data is processed "for the purpose of uniquely identifying" a person. A photograph used for illustration purposes is not Art. 9 data; the same photograph processed through facial recognition software to identify the person is Art. 9 data.

Art. 9(2) Exemptions Applicable to Biometric Processing

Exemption	Reference	Application to Biometrics
Explicit consent	Art. 9(2)(a)	Employee consent often not freely given due to power imbalance (WP29 Opinion 2/2017). Consumer biometric consent must meet Art. 7 standards.
Employment, social security, social protection law	Art. 9(2)(b)	Member State law may authorise biometric processing in the employment context (e.g., biometric access control for high-security areas).
Vital interests	Art. 9(2)(c)	Limited to emergency situations where biometric identification is needed to protect someone's life.
Substantial public interest	Art. 9(2)(g)	Member State law basis required. May apply to law enforcement biometrics where authorised by specific legislation.
Health or social care	Art. 9(2)(h)	Biometric patient identification in healthcare settings.
Public health	Art. 9(2)(i)	Biometric contact tracing during health emergencies (subject to proportionality).
Archiving, scientific research, statistics	Art. 9(2)(j)	Biometric research (e.g., medical imaging analysis) with Art. 89(1) safeguards.

Art. 35(3)(b) — Mandatory DPIA Trigger

Processing on a large scale of special categories of data referred to in Art. 9(1), including biometric data processed for unique identification, requires a DPIA. "Large scale" factors per WP248rev.01:

• Number of data subjects (in absolute terms or as a proportion of the relevant population)
• Volume of data and/or range of data items
• Duration or permanence of the processing
• Geographic scope

EDPB Guidelines 3/2019 on Video Surveillance

Key provisions relevant to facial recognition CCTV:

• Facial recognition in public spaces generally constitutes systematic monitoring of publicly accessible areas (Art. 35(3)(c)) in addition to large-scale biometric processing (Art. 35(3)(b)).
• Facial recognition for access control is less intrusive than identification in public spaces but still requires DPIA.
• Purpose limitation: biometric templates created for access control must not be repurposed for attendance monitoring or performance management.
• Storage limitation: biometric templates should be stored on a device controlled by the data subject (e.g., access card) rather than a central database where possible.

Biometric System Types and Risk Assessment

Verification (1:1 Matching)

One-to-one comparison of a live biometric sample against a stored template for the claimed identity. Used for access control, device unlock, payment authentication.

Risk Factor	Assessment
Data subjects	Defined, enrolled individuals
Volume of data	Limited to enrolled population
Proportionality	Generally more proportionate than identification
Storage recommendation	Template stored on user's device or card (decentralised)
Art. 35(3)(b) trigger	Depends on scale of enrolled population

Identification (1:N Matching)

One-to-many comparison of a live biometric sample against a database of templates to determine identity. Used for law enforcement, border control, surveillance.

Risk Factor	Assessment
Data subjects	Potentially unlimited (all persons in the capture area)
Volume of data	Can be very large (entire population databases)
Proportionality	Highly intrusive; requires strong justification
Storage recommendation	Centralised database is typically required for 1:N matching
Art. 35(3)(b) trigger	Almost always triggered

Categorisation

Classification of individuals into groups based on biometric characteristics (age, gender, ethnicity, emotion) without uniquely identifying them. Used for analytics, targeted advertising, audience measurement.

Risk Factor	Assessment
Art. 9 applicability	May not fall under Art. 9 if not used for unique identification, but still high risk
AI Act classification	Emotion recognition in workplace/education prohibited (Art. 5(1)(f))
Discrimination risk	Categorisation by race, ethnicity, or emotion raises equality law concerns

DPIA Content for Biometric Systems

Systematic Description

For biometric systems, the Art. 35(7)(a) systematic description must include:

Element	Required Detail
Biometric modality	Facial, fingerprint, iris, voice, vein, gait, or multi-modal
Processing mode	Verification (1:1) or identification (1:N) or categorisation
Capture environment	Controlled (sensor/scanner) or uncontrolled (CCTV, ambient camera)
Template storage	Centralised database, decentralised (user device/card), or encrypted enclave
Template format	Proprietary template, ISO 19794 standard, or raw biometric data
Matching algorithm	Vendor algorithm (specify), open-source algorithm, or custom development
Accuracy metrics	False Acceptance Rate (FAR), False Rejection Rate (FRR), Equal Error Rate (EER)
Liveness detection	Anti-spoofing measures (presentation attack detection)
Fallback mechanism	Alternative identification method when biometric fails
Retention period	Template retention, raw biometric data retention, audit log retention

Necessity and Proportionality for Biometric Processing

The proportionality assessment for biometric systems must be rigorous because biometric data:

• Is permanent (unlike passwords, biometric characteristics cannot be changed if compromised)
• Is uniquely identifying (biometric data is inherently linked to the individual)
• Creates heightened risk of function creep (biometric templates can be repurposed)
• Has disproportionate impact if breached (biometric data cannot be reissued)

Proportionality Question	Assessment Standard
Is biometric processing necessary, or can a non-biometric alternative achieve the same purpose?	Badge/card access, PIN, password, or multi-factor authentication without biometrics
Is the biometric modality the least intrusive option?	Fingerprint is generally less intrusive than facial recognition; on-device verification less intrusive than centralised identification
Is the scale of biometric processing proportionate?	Processing all persons in an area (identification) is less proportionate than processing enrolled volunteers (verification)
Is the retention of biometric data minimised?	On-card template storage preferred over centralised database; raw biometric images should not be retained after template extraction

Risk Register for Biometric DPIA

Common Biometric Processing Risks

Risk ID	Risk	Likelihood	Severity	Typical Level
BIO-R1	Biometric data breach — templates or raw data exposed to unauthorised parties	Possible	Maximum	Very High
BIO-R2	Function creep — biometric data collected for access control repurposed for surveillance or attendance monitoring	Likely	Significant	High
BIO-R3	Discriminatory accuracy — facial recognition performs worse on certain demographic groups (skin colour, age, gender)	Likely	Significant	High
BIO-R4	False rejection denying legitimate access — disabled individuals, elderly, or those with skin conditions experience higher rejection rates	Possible	Significant	High
BIO-R5	Spoofing or presentation attacks — fraudulent biometric samples (photos, masks, artificial fingerprints) bypass security	Possible	Significant	High
BIO-R6	Chilling effect — knowledge of biometric surveillance alters behaviour in public or workplace spaces	Likely	Limited	High
BIO-R7	Irreversibility — unlike passwords, compromised biometric data cannot be changed or reissued	Almost certain	Maximum	Very High
BIO-R8	Third-party capture — biometric data of non-enrolled individuals incidentally captured by the system	Likely	Limited	High

Mitigation Measures for Biometric Processing

Measure	Type	Risk Addressed
On-device or on-card template storage (no centralised database)	Technical (DPbD)	BIO-R1, BIO-R7
Template protection: cancellable biometrics or biometric encryption (BioHashing, fuzzy vault)	Technical	BIO-R1, BIO-R7
Liveness detection / presentation attack detection (ISO 30107 compliance)	Technical	BIO-R5
Purpose limitation enforcement through technical access controls	Technical + Organisational	BIO-R2
Demographic accuracy testing across skin colour, age, gender groups (NIST FRVT benchmarks)	Technical	BIO-R3
Non-biometric fallback mechanism (PIN, card, helpdesk override)	Organisational	BIO-R4
Signage and transparency notices in capture areas	Organisational (Transparency)	BIO-R6, BIO-R8
Automatic deletion of raw biometric images after template extraction	Technical (Data minimisation)	BIO-R1
Regular penetration testing of biometric system	Technical (Security)	BIO-R5

Enforcement Precedents

• CNIL vs Clearview AI (2022): EUR 20 million fine for mass collection of facial images from internet for biometric identification database without consent, transparency, or DPIA.
• ICO vs Clearview AI (2022): GBP 7.5 million fine for same biometric processing; enforcement notice ordering deletion of UK residents' biometric data.
• Swedish DPA vs Skelleftea Municipality (2019): SEK 200,000 fine for school using facial recognition for student attendance tracking. Consent relied upon was not freely given due to power imbalance. Less intrusive alternatives were available.
• ICO vs Serco Leisure (2022): Enforcement notice for requiring employees to use facial recognition for time and attendance at leisure centres. No DPIA conducted; no less intrusive alternative offered.
• CNIL vs Clearview AI (2023): EUR 5.2 million additional penalty for non-compliance with the 2022 order, demonstrating the consequences of continued biometric processing in violation of enforcement.
• French Conseil d'Etat vs Presto (2020): Upheld CNIL enforcement against biometric time-and-attendance system for employees; found that fingerprint scanning for attendance was disproportionate when badge systems were available.
• Italian Garante vs P&G (Gillette) (2022): Enforcement action against automated age estimation system using facial analysis in retail stores without DPIA and without valid consent.