PaperlayerView exact revision on arXiv

Source-linked AI summary

A-GE: Privacy-Preserving Age Assurance via Fine-Grained Motion

Mohammed Gharib, Giovanni Cascone, Martin Zizi

arXiv:2610.33206v1cs.CR

TL;DR

Age assurance often requires identity or sensitive-information disclosure, while self-declaration offers little evidence and credentials shift age-establishment trust to an issuer. A-GE derives an application-defined age predicate locally from fine-grained accelerometer and gyroscope motion during ordinary device interaction, without identity documents or credentials. At a 16-year minimum-age threshold, A-GE achieves up to 93.8% accuracy with a comparable F1 score across evaluated configurations.

  • Problem

    Age assurance often requires identity or sensitive-information disclosure, while self-declaration offers little evidence and credentials shift age-establishment trust to an issuer.

    Source: §1 · p. 1

  • Method

    A-GE derives an application-defined age predicate locally from fine-grained accelerometer and gyroscope motion during ordinary device interaction, without identity documents or credentials.

    Source: §1 · p. 2, §4 · p. 7

  • Results

    At a 16-year minimum-age threshold, A-GE achieves up to 93.8% accuracy with a comparable F1 score across evaluated configurations.

    Source: §1 · pp. 2–3

  • Takeaways & Limitations

    The same sensing framework supports minimum-age thresholds from 14 to 21 years and remains effective across substantial variation in IMU sampling rates.

    Source: §8 · p. 14

    A-GE assumes sensor inputs and the client execution path are not compromised, and fine-grained motion may still reveal characteristics beyond age.

    Source: §2 · p. 4

Abstract

from arXiv · show

Age assurance faces a fundamental privacy dilemma: meaningful assurance typically requires revealing identity or other sensitive information, whereas privacy-preserving alternatives often provide little assurance. Existing approaches either disclose sensitive information directly or shift the disclosure and trust requirement to an identity provider, while self-declared age provides little meaningful assurance. Thus, a fundamental gap remains between establishing an age property and revealing information beyond what is necessary to establish that property. We present \textit{A-GE} (Age-Grounded Evidence), a privacy-preserving age assurance system that derives age-related evidence from fine-grained motion captured by commodity inertial measurement units (IMUs). A-GE exploits subtle motion dynamics that we hypothesize reflect underlying neuromuscular and physiological processes during ordinary device interaction. It requires neither identity information nor government identification and operates across available sensor sampling rates, including the approximately 60~Hz limit of web-based sensing, with configurable minimum-age thresholds ranging from 14 to 21 years. We evaluate A-GE on more than 600 participants, with separate participant-disjoint cohorts for model development and final evaluation. At a 16-year minimum-age threshold, A-GE achieves up to 93.8\% accuracy with a comparable F1 score across the evaluated configurations. We further demonstrate client-side deployment on Android, iOS, and the web, and examine unintended attribute inference and residual privacy risks, characterizing both the privacy benefits and limitations of deriving age evidence from commodity motion.

1 Introduction

Age assurance often trades meaningful evidence for sensitive disclosure, while minimizing disclosure can leave age claims weak or dependent on an identity provider. A-GE investigates age evidence from ordinary device motion without identity information.

1 Introduction

  • Existing approaches trade meaningful assurance with additional disclosure against minimal-disclosure self-declaration with little evidence of truthfulness.Identity documents may disclose attributes irrelevant to an age decision.Source: §1 · pp. 1–2
  • Digital credentials reduce disclosure to relying parties but shift the initial age-establishment and trust dependency to an issuer or identity provider.A-GE asks whether age assurance can avoid both direct identity disclosure and this issuer-mediated dependency.Source: §1 · p. 2
  • A-GE derives age-related evidence from fine-grained motion during ordinary device interaction, hypothesized to reflect physiological and neuromuscular processes.It does not require a government-issued document, facial image, or identity credential.Source: §1 · pp. 2–3
  • A-GE supports configurable minimum-age thresholds from 14 to 21 years and remains effective at approximately 60 Hz, the sampling regime available to web applications.It also supports higher sampling rates on platforms that provide them.Source: §1 · pp. 2–3
  • At a 16-year threshold, A-GE achieves up to 93.8% accuracy with a comparable F1 score across evaluated configurations.The evaluation uses more than 600 contributors and a completely held-out, participant-disjoint test cohort.Source: §1 · pp. 2–3
  • A-GE is implemented for Android, iOS, and web deployment, while its privacy analysis examines unintended attribute inference and residual risks.Raw sensor measurements remain on the user’s device; the system does not assume motion measurements are inherently non-sensitive.Source: §1 · p. 3

2 Problem Statement and Adversary Model

A-GE aims to establish an application-defined age predicate while minimizing information disclosure, but its evidence is probabilistic rather than authoritative. Its privacy goals and threat model exclude compromised sensing or execution paths.

Problem Statement and Requirements

  • A-GE seeks to establish Age(u) ≥ T without recovering identity or requiring identity-derived attributes, but its probabilistic evidence is not a replacement for legally binding verification.It targets applications prioritizing useful assurance, minimal identity disclosure, and low user friction.Source: §2 · p. 3
  • During operation, raw sensor measurements remain on-device and only the age-assurance outcome is disclosed; the system should use commodity-device sensors without explicit verification.The same framework should support application-defined thresholds rather than a single age boundary.Source: §2 · p. 4

Adversary Model and Privacy Goals

  • The threat model assumes physical sensor inputs cannot be arbitrarily modified and excludes compromise of the operating system, drivers, runtime, or post-inference outcome.A-GE seeks limited-purpose age inference, while acknowledging that fine-grained motion may contain behavioral or physiological information beyond age.Source: §2 · p. 4

3 Background and Related Work

A-GE is motivated by the possibility that ordinary-device inertial motion contains age-related information, while existing age-assurance approaches differ in their disclosure and trust requirements. The paper distinguishes its use of ordinary interaction from prior motion-based approaches requiring structured locomotion.

Physiological Basis of IMU Motion

  • Fine-grained inertial measurements may reflect motor dynamics coupled to device motion, but the proposed neural-to-IMU pathway is a hypothesis, not a direct neural measurement.An exploratory six-participant IMU–EEG study found approximately 0.19 mean correlation for the examined 4–8 Hz measures; this does not establish causality.Source: §3 · p. 5

Related Work

  • Age-assurance approaches span identity- and credential-based methods, which rely on identity or a trusted issuer, and sensing-based methods, which may collect sensitive user characteristics.The categories address privacy differently rather than providing a single common disclosure model.Source: §3 · p. 5

Identity- and credential-based age assurance

  • Self-declaration provides little evidence that an age claim is correct, while credentials can reduce disclosure to relying parties but retain issuer trust and identity-based age establishment.Credential systems can also depend on issuance, storage, revocation, interoperability, and trust frameworks.Source: §3 · pp. 5–6

Sensing- and behavior-based age assurance

  • Biometric age estimation requires sensitive biometric acquisition, while behavioral profiling can entail persistent observation and depends on stable, representative behavior.Both approaches can avoid presenting identity documents, but their inference processes create distinct privacy considerations.Source: §3 · p. 6
  • Prior IMU age studies use gait or walking, whereas A-GE investigates age information in fine-grained motion during ordinary handheld-device interaction.The comparison considers evidence strength, identity disclosure, trust, configurable predicates, data minimization, raw-sensing disclosure, and user friction.Source: §3 · p. 6

4 A-GE Design

A-GE processes commodity-device inertial measurements locally through signal-quality checks, feature representation, and an age-threshold decision. It rejects unsuitable windows rather than attempting to make poor-quality signals appear valid.

4 A-GE Design

  • The pipeline acquires IMU measurements, checks signal quality, transforms accepted windows into features, and maps those features to an age-assurance outcome.Sensing, quality assurance, representation, and decision operate locally; the outcome rather than sensor measurements is exposed to the relying party.Source: §4 · p. 7

Sensing and Data Acquisition

  • A-GE uses commodity smartphone or wearable accelerometers and gyroscopes across sampling rates, including approximately 60 Hz browser sensing.The browser-available sampling regime is sufficient for effective inference without specialized native sensing access.Source: §4 · pp. 7–8

Signal Quality Assurance

  • Signal Quality Assurance rejects windows with incomplete sampling, stationary-device conditions, or excessive motion that may obscure fine-grained dynamics.It gates inference rather than aggressively filtering artifacts that could alter potentially age-relevant signal components.Source: §4 · p. 8

Age-Related Feature Representation

  • Accepted windows are represented using complementary distributional, higher-order, temporal, frequency-domain, and magnitude-based feature families.These descriptors capture signal statistics, asymmetry and tails, temporal variation, periodic structure, and aggregate motion.Source: §4 · p. 8

Age Assurance Model

  • The final A-GE decision uses XGBoost on features from an accepted sensor window, with participant-level partitioning to measure generalization to unseen participants.The model has 300 boosted trees, maximum depth 6, learning rate 0.05, and a 0.5 probability threshold.Source: §4 · p. 9

5 Experimental Methodology

A-GE is evaluated on strictly separated development and held-out cohorts, with age thresholds and class distributions used to assess generalization. The methodology also examines performance across evaluation conditions and measures auxiliary information recoverable from motion.

5 Experimental Methodology

  • The development cohort includes 273 participants and the held-out evaluation cohort includes 359; separation occurs before window segmentation and all development procedures.This prevents participant-specific characteristics shared across recordings from inflating generalization performance.Source: §5 · p. 9
  • Evaluation spans age thresholds from 14 to 21 years, and accuracy and F1 scores are closely comparable between balanced and naturally distributed held-out evaluations.Development data are balanced between classes; Table 2 reports male/female composition of the two age classes at each threshold.Source: §5 · p. 9, Table 2 · p. 10
  • Figure 2 presents age prediction error across five-year age buckets.The caption specifies the figure’s scope but does not provide error values.Source: Figure 2 · p. 10
  • Auxiliary analyses assess exact-age, gender, and handedness inference, using regression metrics for age and ROC-AUC, accuracy, and F1 for gender and handedness.These analyses characterize potential information leakage rather than intended A-GE capabilities.Source: §5 · p. 10

6 Privacy Analysis

The privacy analysis tests what fine-grained motion may reveal beyond A-GE’s intended age property. It finds measurable age and gender information, while handedness evidence is limited, and frames local processing as reducing disclosure rather than eliminating privacy risks.

Exact-age inference

  • On 273 evaluation participants, exact-age inference from six-second observations yields an 8.53-year MAE, while only 19.8% of predictions fall within three years of actual age.The results suggest some age information may leak beyond the binary threshold, but large prediction errors limit usefulness for age estimation.Source: §6 · p. 10

Gender inference

  • Six-second observations support gender inference at 72.6% accuracy, 72.2% F1, and 0.807 AUC, indicating some gender-related information in the modality.The inference remains imperfect, and raw IMU measurements are processed locally and never leave the device.Source: §6 · p. 10

Handedness inference

  • Handedness inference reaches 54.6% accuracy for dominant versus non-dominant hand and 56.8% for left versus right, providing limited leakage evidence.The left-versus-right result may reflect device orientation or phone positioning; dominant-hand inference is affected by the smaller number of left-handed participants.Source: §6 · p. 11

Privacy analysis

  • A-GE reduces what must be disclosed for age assurance, but does not claim to eliminate sensor-inference privacy risks.The privacy analysis shows both age-assurance utility and selected secondary inferences; raw inertial measurements are not transmitted to the relying party.Source: §6 · pp. 10–11

7 Evaluation

The evaluation tests A-GE across age thresholds, observation and sampling conditions, demographic groups, and client platforms. Averaging two independent 3-second observations achieves 91.1% accuracy, exceeding the reported single-window comparisons.

Age-Threshold Generality

  • At the 16-year threshold, performance is strongest; at 21 years, accuracy is approximately 77% under the fixed experimental setting.The authors caution that class composition varies by threshold and that the 16-year peak is not evidence of an intrinsic physiological optimum.Source: §7 · p. 11

Sampling-Rate Robustness

  • Across 60–500 Hz, accuracy ranges from 87.7% to 89.0% and F1 from 87.3% to 88.7%, with highest measured performance at 500 Hz.The tested setting uses the 16-year threshold and a 5-second observation window; similar qualitative behavior is reported across thresholds from 14 to 21 years.Source: §7 · pp. 11–12, Figure 4 · p. 12

Observation Duration and Aggregation

  • Performance improves substantially as observation windows grow from 1 to approximately 5 seconds, but longer windows do not necessarily improve it further.Together with the modest variation across sampling rates, this suggests performance depends more on motion’s temporal characteristics than on sample count alone.Source: §7 · p. 12, Figure 5 · p. 12
  • Incorrect classifications have mean decision margins near 0.05 across thresholds, while correct-classification margins are substantially larger.The authors interpret this as errors being associated, on average, with predictions close to the decision boundary.Source: Figure 6 · p. 12, §7 · p. 13
  • 91.1% accuracy is achieved by averaging two independent 3-second observations, compared with 83.6% for one 3-second observation and 87.7% for one 6-second observation.The comparison evaluates both the temporal extent of the evidence and how it is aggregated.Source: §7 · p. 13, Table 3 · p. 13

Demographic Performance

  • Performance varies across four gender–age subgroups, but each contains only 32 participants and the confidence intervals are relatively wide.The authors characterize the observed differences as measured variation, not definitive demographic performance gaps.Source: §7 · p. 13, Table 4 · p. 13

Deployment Overhead

  • Client-side inference has low computational overhead on the evaluated mobile and web platforms, and the best observed performance varies by threshold rather than coming from one configuration.Table 5 reports inference latency, peak memory, and CPU utilization; Table 6 summarizes the highest observed accuracy and F1 score by threshold.Source: §7 · pp. 13–14, Table 5 · p. 14, Table 6 · p. 14

8 Conclusion

A-GE derives an application-defined age predicate locally from ordinary-device inertial measurements, with findings supporting short-window inference, temporal aggregation, sampling-rate robustness, and thresholds from 14 to 21 years. The conclusion also emphasizes that motion carries information beyond age and that practical deployment still requires end-to-end security guarantees.

8 Conclusion

  • A-GE derives an application-defined age predicate locally from accelerometer and gyroscope measurements collected during ordinary interaction with commodity devices.It does not rely on identity documents, biometric images, persistent behavioral profiles, or a trusted age-credential issuer.Source: §8 · p. 14
  • The evaluation finds that short windows can provide effective assurance, aggregation can improve performance, sampling-rate variation is tolerated, and thresholds from 14 to 21 years are supported.Longer contiguous observations do not consistently improve performance.Source: §8 · p. 14
  • IMU measurements encode information beyond age, and A-GE establishes the feasibility of sensing and inference rather than end-to-end deployment security.Future work is to protect acquisition, feature extraction, inference, and decision-making while producing a cryptographically verifiable age-assurance certificate.Source: §8 · p. 14
Loading 2610.33206v1…