Source-linked AI summary
Dorsal Hand Images for Immersive (XR) and Privacy-preserving Age Assurance and Child Safety
Riccardo Bovo, George Loukas, Josh P. Davis
TL;DR
XR needs age assurance that can operate during immersive sessions without requiring users to expose their faces to third parties. The paper evaluates dorsal-hand images with standard neural architectures on a diverse, unconstrained dataset spanning the minor–adult boundary. The results support dorsal-hand inference as a privacy-preserving first-stage modality, while current error rates and calibration, diversity, and capture constraints limit unassisted deployment.
Problem
Face-based age assurance is impractical for in-session XR because it requires removing the headset, using a third-party app, and uploading a selfie image.
Method
The study evaluates four neural architectures on continuous age regression and binary adult/minor classification using 436 participants captured under unconstrained XR-style conditions.
Results
SwinV2-B achieves the highest low-false-acceptance discrimination, with pAUC@0.1 of 0.0587 ± 0.0018, and is selected as the recommended age-gate model.
Takeaways & Limitations
Dorsal hand morphology provides sufficient discriminative signal for minor–adult age assurance and offers a passive in-session alternative to face verification in XR.
Takeaways & Limitations
The 6% FPR for 16–17-year-olds at τ=0.923 remains below the threshold for unassisted deployment, and broader demographic and uncontrolled real-world validation is still needed.
Abstract
from arXiv · showhide
Ensuring that Extended Reality (XR) environments are age-appropriate is an important regulatory and safety challenge. However, current age assurance operates only at registration and cannot verify the age of the active user during a session. Face-based approaches, the dominant solution in social media and adult platforms, are impractical in XR, because they require removing the headset and taking a self-captured image, often on a mobile app. This both breaks immersion and introduces the privacy risk of sharing face pictures with third parties, which leaves XR platforms without a viable path to continuous, in-session and privacy-preserving age assurance. We propose the dorsal part of the hand as an alternative to the face, by exploiting the egocentric cameras that XR headsets inherently and naturally use to capture gesture interactions. To evaluate this, we collect an age- and sex-stratified, ethnodiverse dataset of 436 participants spanning the minor--adult boundary, captured under unconstrained lighting and orientation conditions. To characterise what is achievable with off-the-shelf methods at the minor--adult boundary, we evaluate standard neural network architectures for age assurance at the legally critical 18-year threshold. Analysis confirms performance is robust to skin-tone variation. On this dataset, the challenge-31 operating point achieves zero minor admission, making the system a viable first-stage filter for age assurance. These findings position dorsal hand morphometrics as an effective and more privacy-preserving biometric modality for in-session age assurance in XR.
1. Introduction
XR’s immersive use by children creates a regulatory need for in-session age assurance, but face-based verification is impractical because it interrupts sessions and requires third-party image sharing. The paper addresses this gap with dorsal-hand inference using headset cameras during normal interaction.
- Motivation: XR introduces amplified risks for children, including harassment, grooming, and exposure to harmful content, while parents cannot observe in-headset experiences.Age-assurance requirements are being introduced across multiple jurisdictions.
- Motivation: Face-based age assurance requires removing the headset, using a third-party app, and uploading a selfie image, creating an impractical out-of-band workflow.The headset’s outward-facing pass-through cameras cannot image the wearer’s face during a session.
- Research gap: Prior work leaves discrimination at the minor–adult boundary open because aging signals below the late twenties are dominated by developmental growth effects.Existing studies also rely on controlled laboratory conditions and datasets lacking age stratification and ethnodiverse representation.
- Contribution: The paper collects an ethnodiverse, age- and sex-stratified dataset of 436 participants under unconstrained conditions and evaluates neural networks for age assurance.The dataset is designed to address gaps in age coverage, demographic representation, and ecological validity.
2. Related Work
Dorsal-hand aging cues are measurable in images, but existing age-inference studies and datasets do not adequately support generalization at the minor–adult boundary. This motivates a purpose-built corpus with age labels, minor coverage, and demographic diversity under realistic capture conditions.
- Biological basis: Dorsal-hand aging cues include wrinkles, visible veins, age spots, volume loss, skin roughness, and chromophore distribution.These features are predominantly visible in standard photography when the dorsal hand fills the frame.
- Biological basis: Aging rates differ substantially across populations, making demographic diversity a prerequisite for generalizable hand-age inference.Caucasian and East Asian cohorts show different timelines for roughness and pigmentation changes.
- Prior methods: High-resolution studies associate dorsal-hand texture, wrinkle dimensions, and epidermal properties with chronological age.Reported approaches include hyperspectral texture analysis, wrinkle-based pipelines, and epidermal image processing.
- Prior methods: Existing dorsal-hand studies often treat age as an auxiliary task in identification systems, allowing subject-specific cues and excluding minor participants.This setup can inflate performance estimates for unseen-person age inference.
- Dataset gap: 96% of 11K Hands participants are aged 20–29, making the dataset useful for biometric identification but unrepresentative for age assurance.The related datasets do not provide verified minor–adult coverage with demographic diversity.
- Dataset gap: The proposed capture protocol varies lighting and hand position to produce consumer-grade conditions rather than controlled clinical photography.Sixteen captures per participant span eight lighting conditions.
3. Dataset
The study constructs a purpose-built dorsal-hand corpus from 436 participants spanning ages 10–67, with stratified demographics and unconstrained XR-style capture. It records varied lighting and orientations, removes background pixels using depth, and characterizes skin-tone diversity with ITA.
- Participants and images: The dataset is stratified by age and sex to improve coverage near the legally critical 18-year threshold.It spans the full skin-tone spectrum for fairness analysis.
- Participants and images: 436 participants aged 10–67 contributed 6,790 dorsal-hand images under unconstrained lighting and hand-orientation conditions.Participants were recruited from secondary schools and university staff and students, with age-band quotas.
- XR equipment and preprocessing: The Meta Quest 3 supplied pass-through camera frames at 1280×960 px, while per-pixel depth removed background pixels before storage.The resulting images contain no background cues for model use.
- Capture protocol: The capture session lasted approximately five minutes and used 16 captures across eight conditions combining two lighting intensities and four lighting origins.Audio-visual cues guided hand placement within the capture frustum.
- Skin-tone characterization: Skin tone was quantified with ITA from mean CIE L∗ and b∗ values in a 50 × 50 px dorsal-hand patch.The patch was extracted from a single capture per participant at the position of strongest dorsal illumination.
- Skin-tone characterization: 40.4% of participants were classified as dark skin tone, 36.0% as medium, and 23.6% as light across the usable patch cohort.The cohort included 428 participants with successfully extracted and usable patches.
- Skin-tone characterization: ITA measurements are best-effort proxies because images were captured under unconstrained lighting without color-calibration targets.They are used primarily to characterize relative skin-tone diversity within the dataset.
4. Methods and Experiment
The study evaluates four neural architectures for dorsal-hand age regression and binary adult/minor classification under participant-level cross-validation. Training incorporates augmentation for unconstrained orientation and lighting, probabilistic uncertainty modeling, and calibrated age-gate thresholds.
- Model evaluation: Four neural architectures are evaluated on continuous age regression and binary adult/minor classification, spanning capacity and efficiency trade-offs.ResNet-50, EfficientNetV2-S, Swin Transformer V2-B, and MobileNetV3-Large represent distinct design points.
- Evaluation protocol: Participant-level stratified splits prevent identity leakage and estimate generalization to unseen participants.A held-out test set is drawn first, followed by stratified five-fold cross-validation using age and skin-tone groups.
- Data processing: Spatial augmentation covers full-range rotations and flips, while brightness and contrast jitter model unconstrained XR orientations and variable lighting.Inference uses only a centre-crop resize.
- Probabilistic regression: Models minimize Gaussian NLL while jointly predicting age mean and log-variance, enabling uncertainty estimation for age-gate calibration.Log-variance is clamped for numerical stability, and inverse-frequency weighting strengthens rare-age contributions without allowing any age to dominate.
- Binary age gate: Adult probability is computed from the Gaussian output, and model-specific thresholds balance falsely accepting minors against correctly accepting adults.The threshold is selected using the false positive rate and true positive rate operating-point framework.
- Reported metrics: Age-regression performance is summarized by MAE, RMSE, and intra-user standard deviation across five held-out test folds.MAE is also reported by decade across architectures.
5. Results
Across four architectures, SwinV2-B provides the strongest safety-relevant age-gate discrimination and is selected for deployment. Its performance remains robust across skin-tone groups, while stricter thresholds and multi-capture aggregation reduce minor admission.
- Age regression: SwinV2-B achieves the lowest MAE at 5.78 ± 0.64 yr and RMSE at 7.47 ± 0.72 yr, while MobileNetV3-L performs worst across all metrics.The RMSE gap between MobileNetV3-L and SwinV2-B is significant after Holm correction at p=0.016.
- Age-gate discrimination: SwinV2-B achieves the highest pAUC@0.1 at 0.0587 ± 0.0018, the safety-critical low-false-acceptance metric.EfficientNetV2-S has the highest overall AUC at 0.922 ± 0.008, but SwinV2-B is selected based on pAUC@0.1.
- Age-gate discrimination: SwinV2-B’s pAUC@0.1 advantage over RN50 and MobileNetV3-L is +0.0123 and +0.0135, respectively, with both differences significant after Holm correction.Both comparisons report p=0.0007.
- Operating thresholds: At τ=0.960, the Challenge-31 setting reduces minor admission to zero but produces an overall adult FNR of 63%.The stricter threshold flags nearly all adults under 40.
- Skin-tone analysis: SwinV2-B’s pAUC@0.1 values are 0.0679, 0.0636, and 0.0620 for Medium, Dark, and Light skin-tone groups, with all Holm-corrected pairwise p-values above 0.82.The analysis reports no significant discrimination difference across groups within this dataset.
- Multi-capture aggregation: At τ=0.923, multi-capture aggregation reduces 16–17-year-old FPR from 6% with one capture to 0% with four, while adult pass rate remains approximately 48%.Total minor FPR also falls from 0.9% to 0%, with participant coverage remaining 100% for n ≤4.
6. Discussion
Dorsal hand morphology supports privacy-preserving, in-session age assurance for XR under unconstrained conditions. Its conservative operating point can serve as a first-stage filter, while hand-based inference offers lower practical re-identification risk than face-based verification.
- The system supports minor–adult age assurance from dorsal hand morphology under unconstrained, ecologically valid conditions.The study advances beyond prior adult-only, laboratory-based dorsal hand evaluations.
- 0% minor admission is achieved at the strict Challenge-31 operating point, supporting conservative first-stage filtering.At the recommended Challenge-28 regime, multi-capture aggregation reduces admission of 16–17-year-olds to 0% at n=4 while adult pass rate remains 48%.
- Dorsal hand inference enables passive in-session XR verification through headset pass-through cameras, unlike face-based approaches that interrupt sessions and require third-party face presentation.
- Hand images retain biometric status but carry less practical re-identification risk than faces because fewer large-scale matching infrastructures exist for hands.
7. Limitations and Future Work
The current findings support dorsal hand age assurance but leave important validation and interpretability questions. Future work should test broader real-world conditions and determine whether models rely on biologically meaningful aging cues.
- Limitations: The 6% false-positive rate for 16–17-year-olds at τ=0.923 remains below the threshold for unassisted deployment without secondary verification.
- Limitations: The Challenge-T equivalence depends on a Gaussian calibration assumption that requires validation in the 15–21 age region.
- Limitations: Performance under greater demographic diversity and uncontrolled real-world capture remains untested, while ITA provides an uncalibrated skin-tone proxy.
- Future Work: Gradient-based saliency aligned with expert-annotated regions could test whether models use biological cues rather than artefacts.Relevant cues include wrinkle density, vein prominence, age spots, volume loss, and skin roughness.
8. Conclusion
The paper establishes an age-stratified, ethnodiverse dorsal hand dataset spanning the minor–adult boundary and benchmarks RGB-based age assurance at the 18-year threshold. Its results support dorsal hand images as a privacy-preserving modality for child-safe deployment contexts.
- This work presents the first age-stratified, ethnodiverse dorsal hand study spanning the minor–adult boundary from age 10 onward.
- SwinV2-B achieves MAE 5.78 yr and pAUC@0.1 0.0587 on the minor–adult boundary benchmark.
- Including minors enables evaluation at the legally critical 18-year boundary, while ITA-based characterization spans the full Fitzpatrick spectrum.
- The probabilistic formulation and privacy properties of hand-based inference suit child-safe, data-minimising age-assurance contexts.