💻 computer science

Dorsal Hand Images for Immersive (XR) and Privacy-preserving Age Assurance and Child Safety

This paper proposes using dorsal hand images captured by XR headset cameras as a privacy-preserving and immersive alternative to facial recognition for continuous age assurance, demonstrating through a diverse dataset that this method can effectively distinguish minors from adults at the 18-year threshold with robustness to skin tone variations.

Original authors: Riccardo Bovo, George Loukas, Josh P. Davis

Published 2026-08-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Riccardo Bovo, George Loukas, Josh P. Davis

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the rapidly expanding world of virtual reality, where users don headsets to step into digital worlds, a persistent safety challenge looms: how to ensure that children are not exposed to content or interactions meant for adults. Current rules require platforms to verify a user's age, but the methods used today are often clumsy and intrusive. The standard approach relies on facial recognition, asking a user to remove their headset, take a selfie with a phone, and upload it to a third party. This process breaks the sense of presence, forcing the user out of the immersive experience, and raises serious privacy concerns by sharing sensitive facial data with external companies. Furthermore, once a user is inside a session, there is no easy way to continuously check if the person wearing the headset is still the same person who signed up, leaving a gap where minors could slip into adult-only spaces.

Researchers have long known that the skin on the back of the hand changes in predictable ways as people grow older, developing wrinkles, changing texture, and showing veins more clearly. While these signs are well-documented in medical studies, they have rarely been tested as a tool for digital age verification, especially for distinguishing between teenagers and adults. A team of researchers at the University of Greenwich has now explored whether the back of the hand, captured naturally by the cameras already built into virtual reality headsets, could serve as a reliable, privacy-friendly way to check age without ever asking a user to show their face.

The researchers set out to build a new dataset to test this idea, recognizing that previous studies had only looked at adults or used controlled laboratory settings that did not reflect real-world use. They recruited 436 participants, ranging from ten to sixty-seven years old, ensuring a balanced mix of ages, sexes, and skin tones. The goal was to capture images under conditions that mimic how a person would actually use a virtual reality device: holding their hand up freely in varying light, without special equipment to hold the hand still or standardize the lighting. Using a Meta Quest 3 headset, the team guided participants to place their hands within a specific viewing area, capturing images of the back of the hand as the user would naturally interact with the virtual world. This resulted in thousands of images that reflected the messy, variable reality of everyday use, rather than the perfect conditions of a medical clinic.

With this diverse collection of images, the team trained several different computer systems to learn the difference between a child and an adult. They tested various types of artificial intelligence models, from standard networks to more advanced ones designed to handle complex patterns, asking them to predict the age of the person in the photo. The results showed that these systems could indeed distinguish between minors and adults with a reasonable degree of accuracy. The best-performing model made an average error of about five and three-quarter years when guessing the exact age of a person. While this might seem like a large margin for guessing a specific birthday, the researchers were less interested in the exact number and more focused on a binary decision: is this person eighteen or older?

When the system was tuned to be very careful about not letting a child into an adult space, it performed with high reliability. By setting a strict threshold, the system could effectively block all minors from entering, acting as a powerful first filter. The researchers found that the system worked consistently well across different skin tones, showing no significant bias against people with darker or lighter skin. This is a crucial finding, as many age verification tools struggle to work fairly across diverse populations. The study also demonstrated that taking multiple pictures of the hand during a single session and combining the results made the system even more accurate, reducing the chance of a mistake to nearly zero for the youngest teenagers while still allowing a significant portion of adults to pass through without needing further verification.

The implications of this work extend beyond just better accuracy; they touch on the fundamental issue of privacy in digital spaces. Unlike a face, which is visible everywhere in public life and on social media, the back of a hand is rarely photographed or shared. If a hand image were to be leaked, it would be far harder for a third party to link it back to a specific individual compared to a face. This makes the hand a much safer biometric tool for protecting children, as it allows platforms to verify age without collecting the sensitive facial data that has led to numerous data breaches in the past. The researchers suggest that this method could be integrated directly into virtual reality sessions, checking age continuously and passively as a user interacts with the world, all without ever asking them to take off their headset or interrupt their experience.

While the study does not claim that this method is a perfect, standalone solution for every scenario, it establishes that the back of the hand carries enough biological information to serve as a viable, privacy-preserving tool for age assurance. The work fills a critical gap in the field by proving that these methods can work across the difficult boundary between childhood and adulthood, a range where visual signs of aging are often subtle and inconsistent. By demonstrating that standard artificial intelligence models can learn these subtle cues from images taken in uncontrolled lighting, the researchers have provided a concrete path forward for creating safer, more immersive digital environments where children can be protected without sacrificing their privacy or their sense of presence.

Technical Summary: Dorsal Hand Images for Immersive (XR) and Privacy-preserving Age Assurance

1. Problem Statement

Ensuring age-appropriate access in Extended Reality (XR) environments is a critical regulatory and safety challenge. Current age assurance mechanisms rely on face-based estimation, which is impractical for XR due to architectural constraints: XR headsets utilize outward-facing pass-through cameras that cannot image the wearer's face during a session. Consequently, face-based verification requires users to remove their headsets and capture self-images via third-party mobile apps. This process breaks immersion, limits the frequency of checks (increasing account-sharing risks), and introduces significant privacy risks by requiring the sharing of facial biometrics with third parties. Furthermore, existing dorsal hand datasets and studies focus on adult populations under controlled laboratory conditions, leaving the discrimination of minors from adults (the minor–adult boundary) uncharacterized and untested under real-world, unconstrained conditions.

2. Methodology

The authors propose leveraging the dorsal (back) of the hand as a privacy-preserving biometric modality, utilizing the RGB and depth cameras inherent to XR headsets to capture hand morphometrics during normal interaction.

Dataset Collection:

Participants: A purpose-built dataset of 436 participants (ages 10–67) was collected, stratified by age and sex to ensure balanced coverage around the legally critical 18-year threshold. The cohort includes 45.9% minors (10–17) and 54.1% adults.
Demographics: The dataset spans the full skin-tone spectrum (Fitzpatrick grades I–VII), categorized into Light (23.6%), Medium (36.0%), and Dark (40.4%) groups based on Individual Typology Angle (ITA) calculations.
Capture Conditions: Unlike prior clinical studies, images were captured under unconstrained conditions representative of consumer XR use. Participants held their hands freely within a defined capture frustum, guided by visual depth-shader feedback.
Protocol: Each participant underwent 16 captures (two hands × 4 positions × 2 lighting intensities) across 8 distinct lighting conditions, yielding 6,790 dorsal hand images.
Preprocessing: Background pixels were removed using the headset's depth stream; models receive only the hand region.

Experimental Setup:

Models: Four neural network architectures were evaluated: ResNet-50, EfficientNetV2-S, Swin Transformer V2-B, and MobileNetV3-Large.
Tasks:
1. Continuous Age Regression: Predicting chronological age.
2. Binary Age Gate: Classifying participants as "minor" (<18) or "adult" (≥18).
Training: Models were trained using participant-level stratified 5-fold cross-validation. The training objective was Gaussian Negative Log-Likelihood (NLL), outputting a mean ( $\mu$ ) and log-variance ( $\log \sigma^2$ ) to enable probabilistic uncertainty estimation. Inverse-frequency weighting was applied to address age distribution imbalances.
Decision Thresholding: A binary decision was derived from the Gaussian output ( $P(a \ge 18)$ ). Thresholds were selected to balance False Positive Rates (FPR) for minors against True Positive Rates (TPR) for adults, mapping to NIST FATE "Challenge-T" frameworks.

3. Key Results

Age Regression:

The Swin Transformer V2-B achieved the best performance with a Mean Absolute Error (MAE) of 5.78 ± 0.64 years and RMSE of 7.47 ± 0.72 years.
EfficientNetV2-S followed with an MAE of 6.01 ± 0.11 years. MobileNetV3-Large performed the worst (MAE 6.96 years).
Performance was robust across skin-tone groups, with no statistically significant differences in discrimination capability between Light, Medium, and Dark skin tones.

Age Gate (Binary Classification):

SwinV2-B was selected as the recommended model for deployment due to its superior performance in the safety-critical low-false-acceptance region (pAUC@0.1 = 0.0587).
Operating Points:
- At a threshold of $\tau=0.923$ (equivalent to NIST Challenge-28), the system achieved a 6% FPR for 16–17-year-olds and a 48% adult pass rate.
- At a stricter threshold of $\tau=0.960$ (equivalent to NIST Challenge-31), the system achieved zero minor admission (0% FPR for minors), though this increased the False Negative Rate (FNR) for adults to 63%.
Multi-Capture Aggregation: Aggregating scores from multiple captures (e.g., $n=4$ ) significantly improved performance. At $\tau=0.923$ , aggregating 4 captures reduced the 16–17-year-old FPR from 6% to 0%, while maintaining the adult pass rate at ~48%.

4. Significance and Claims

The paper claims to provide the first empirical evidence that dorsal hand RGB images can support safety-critical age assurance across the minor–adult boundary under unconstrained, ecologically valid conditions.

Privacy Preservation: The approach offers a passive, in-session alternative to face-based verification. It avoids the need for users to remove headsets or share facial biometrics with third parties, thereby reducing re-identification risks. The authors note that while hands are biometric data, they lack the extensive infrastructure of facial datasets (e.g., social media, CCTV) that facilitate linking biometric data to specific individuals.
Regulatory Viability: The probabilistic formulation maps directly to established age-assurance frameworks (NIST FATE Challenge-T). The system is positioned not as a replacement for all verification methods, but as a viable first-stage filter that can escalate borderline cases to higher-friction document-based verification or enable adaptive safety controls (e.g., muting voice chat for minors) without interrupting the user experience.
Generalizability: By including a diverse, ethnodiverse cohort of minors and adults under variable lighting, the study addresses a critical gap left by prior adult-only, lab-based research.

The authors conclude that dorsal hand morphometrics represent an effective, privacy-preserving biometric modality for XR age assurance, capable of achieving zero minor admission at strict operating points while maintaining robustness to skin-tone variation.

Technical Summary: Dorsal Hand Images for Immersive (XR) and Privacy-preserving Age Assurance

1. Problem Statement

2. Methodology

3. Key Results

4. Significance and Claims

More like this