Position: Age Estimation Models Do Not Process Biometric Data

Technical Summary: Position: Age Estimation Models Do Not Process Biometric Data

Problem Statement
The paper addresses a critical regulatory ambiguity: whether neural networks that estimate age from facial photographs process "biometric data" under frameworks such as the GDPR (Article 4(14)), the Illinois Biometric Information Privacy Act (BIPA), and the EU AI Act. The definition of biometric data often hinges on whether the data "allows or confirms the unique identification" of a natural person. While regulators have debated whether the purpose (age estimation vs. identification) or the technical capability (whether intermediate representations contain identity information) should determine classification, no empirical guidance exists. This uncertainty creates significant compliance risks, including the need for explicit consent, statutory damages, or high-risk AI classification, potentially hindering the deployment of age verification systems for child safety.

Methodology
To resolve the question of capability empirically, the author evaluated 14 distinct neural network models across three standard face verification benchmarks:

1. Models: The study included primary subjects (age estimation models trained solely for age prediction), facial attribute models (predicting gender, emotion, etc.), a face recognition baseline (InsightFace ResNet-100 with ArcFace), and general-purpose vision models (e.g., DINOv3, ImageNet ResNet-50).
2. Benchmarks:
   • LFW (Labeled Faces in the Wild): Standard in-the-wild verification.
   • AgeDB-30: Pairs of the same individual with up to 30-year age gaps to test robustness to aging.
   • CFP-FP: Frontal-to-profile matching to test pose invariance.
3. Evaluation Protocol:
   • Layer-by-Layer Analysis: Embeddings were extracted from intermediate layers of the networks (not just the final output) to determine if identity-discriminative information emerges during inference.
   • Metrics: Performance was measured using False Non-Match Rate (FNMR) at fixed False Match Rates (FMR). Primary results were reported at FMR=1% (statistically reliable given dataset size), with FMR=0.01% reported for regulatory comparison.
   • Attention Probing: A secondary experiment attached an attention-based probe to frozen model features, training a new head on identity-labeled data (Glint360k) to determine if identity information was latent but inaccessible via simple pooling.
   • Benchmarks: Results were compared against regulatory thresholds from NIST SP 800-63-4, the EU Entry/Exit System (EES), and the FIDO Alliance.

Key Results
The empirical evidence demonstrates that age estimation models lack meaningful identification capability:

• Performance Gap: Age estimation models fall orders of magnitude short of identification thresholds. On the LFW benchmark at FMR=1%, the best-performing commercial age estimator achieved a 26.8% FNMR, while the face recognition baseline (ArcFace) achieved 0.23%. This represents a gap of two orders of magnitude.
• Regulatory Thresholds: At the regulatory operating point (FMR=0.01%), age estimation models reached FNMRs between 64% and 99%, far exceeding the <5% requirement for biometric systems.
• Layer Analysis: No intermediate layer in any age estimation model achieved verification performance close to regulatory standards. Even the "best" layers remained significantly above the 5% FNMR threshold.
• Adversarial Settings: On AgeDB-30 (large age gaps) and CFP-FP (pose variation), age estimation models performed at or near random chance (saturating at 96–100% FNMR), indicating that features optimized for age estimation are orthogonal to identity-discriminative features.
• Probing Results: Even when an attention-based probe was trained on the frozen features of age estimation models using identity labels, performance remained poor on harder benchmarks (e.g., 67–98% FNMR on AgeDB-30 and CFP-FP). This suggests the representations fundamentally lack the information required for cross-age or cross-pose identification.

Key Contributions

• Empirical Evidence: The paper provides the first systematic, empirical evaluation of whether age estimation models possess the technical capability to identify individuals, directly addressing a gap in regulatory guidance.
• Benchmarking Against Standards: It establishes a concrete baseline by comparing model performance against specific, quantitative thresholds defined by NIST, EU EES, and FIDO, moving the debate from theoretical interpretation to measurable capability.
• Differentiation of Processing: The study distinguishes between transient intermediate representations (which are discarded after inference) and stored biometric templates, providing data to support the argument that the former does not constitute biometric data under current definitions.

Significance and Claims
The paper claims that its primary contribution is evidence, not a legal ruling. It argues that because age estimation models demonstrably cannot identify individuals (failing verification benchmarks by two orders of magnitude), their intermediate representations do not meet the capability-based definition of biometric data found in GDPR Article 4(14).

The author calls for:

1. Transparency from Researchers: Model documentation (e.g., model cards) should explicitly state whether embeddings are stored or discarded and include verification benchmark results to clarify a system's capabilities.
2. Regulatory Clarity: Regulators should distinguish between "transient processing" (where representations exist only during inference and are discarded) and "template storage" (where embeddings are saved for identification). The paper argues that systems incapable of identification should not be subject to the strictest biometric data protections designed for identification systems, as over-regulation impedes child safety applications.

The paper concludes that while facial images remain personal data, the specific high-risk protections associated with biometric data (such as Article 9 explicit consent or BIPA statutory damages) should not apply to systems that empirically cannot identify individuals.