PaperlayerView exact revision on arXiv

Source-linked AI summary

Prayer-Gait-Auth: Smartphone IMU-based Behavioral Biometrics from Structured Islamic Prayer Movements

Hassan Hizeh, Anwar B. Alshaibani, Muhammad Mahboob Ur Rahman, Tareq Y. Al-Naffouri

arXiv:2608.26994v1eess.SP

TL;DR

The paper investigates whether participant identity can be inferred from movement execution within a fixed Islamic prayer sequence, especially across independently acquired sessions. It uses smartphone IMU recordings from 95 participants, structurally complete two-rakaah units, complementary orientation representations, and unit- and session-level biometric protocols. The results support structured prayer movement as a measurable cross-session behavioral biometric, while identifying device diversity and limited temporal scope as important boundaries.

  • Problem

    Whether participant-specific execution in a prescribed prayer sequence is distinctive and repeatable enough to support identification and verification across recording sessions remains the central biometric question.

  • Method

    The study analyzes smartphone IMU recordings from 95 participants, segments them into complete two-rakaah units, and compares rotation-invariant with Qibla-referenced representations across complementary biometric protocols.

  • Results

    Participant-discriminative motion structure supports strong identification and verification at unit level, while complete-session holdout shows that a substantial component persists across independently acquired recordings.

  • Takeaways & Limitations

    Structured prayer movement is a measurable cross-session behavioral biometric, and participant information is distributed across motion dynamics rather than concentrated in one signal or posture.

  • Takeaways & Limitations

    Session-disjoint performance is not yet equivalent to device-disjoint validation, and the data do not establish long-term biometric permanence over months or years.

Abstract

from arXiv · show

Islamic prayer is a structured movement activity that offers a distinctive setting for behavioral biometrics: all participants execute the same action sequence, so identity must be inferred from differences in execution. We collected inertial data from 95 participants using their own smartphones during nightly congregational Islamic prayer (Taraweeh). A label-aware pipeline converts long recordings into structurally complete two-rakaah behavioral samples (prayer units). This allows us to design unit-level and complete-session behavioral biometrics protocols. To address arbitrary smartphone orientation in worshippers' pockets, we study two motion representations: rotation-invariant magnitudes with gravity-relative acceleration components, and a metadata-aware Qibla-referenced canonicalization that harmonizes platform conventions, reconstructs device-to-world attitude, corrects Android magnetic north to true north via WMM2025, and expresses acceleration and angular velocity in a common Qibla-left-up frame. Under unit-level protocol, the invariant and Qibla-referenced representations reach learned pairwise Random Forest AUC/EER of 0.9932/4.07% and 0.9923/3.96%; under complete-session holdout, 0.9700/7.62% and 0.9618/7.81%. Qibla-frame directional ablations show the complete six-axis representation is strongest overall, with acceleration retaining most learned-verification performance and vertical motion the strongest single-axis cue. A Qibla-referenced SimCLR experiment further yields participant-template AUC 0.9805-0.9817 and EER 5.81-6.24% across two unit-level runs. Signal-, descriptor-, prayer-component ablation analyses show participant identity is distributed across movement dynamics rather than concentrated in one signal or posture. These results establish structured prayer movement as a measurable cross-session behavioral biometric.

I. INTRODUCTION

The paper asks whether participant-specific execution within a prescribed Islamic prayer sequence can support cross-session behavioral biometrics. It introduces matched orientation-compensation strategies and complementary evaluation protocols using real-world smartphone IMU recordings from 95 participants.

  • Motivation and study setting: The fixed ordered sequence of prayer postures controls activity content, shifting biometric analysis toward differences in execution.The biometric question concerns whether participant-specific execution is distinctive and repeatable across recording sessions.
  • Study design: The study defines a prayer unit as two consecutive rakaahs and a session as a single-night Taraweeh recording containing at least four units.Units and sessions serve as behavioral samples for unit-level and session-level analyses.
  • Study design: The paper compares rotation-invariant magnitude-plus-gravity-relative features with metadata-aware Qibla-referenced directional canonicalization.The two representations address arbitrary phone orientation while making different trade-offs between directional detail and orientation dependence.
  • Motivation and study setting: 95 participants recorded smartphone IMU data during nightly congregational Taraweeh prayer using their own phones.Phones were carried in one of the two front trouser pockets.
  • Study design: The evaluation spans unit-level and session-level protocols, including identification, verification, ablations, and supervised and self-supervised representation learning.The contribution list explicitly includes matched representation comparison and multiple biometric evaluation settings.
  • Caveat: Heterogeneous handset hardware is treated as a potential confound because device effects may correlate with participant identity.The study intentionally uses participant-carried phones rather than identical calibrated devices.

II. RELATED WORK

Prior prayer-sensing work mainly recognizes postures, transitions, activities, or rakaah counts rather than participant identity. This paper instead uses prescribed prayer as a controlled behavioral task and evaluates coordinate treatments under matched protocols.

  • Prayer sensing: Existing prayer IMU studies have primarily addressed posture and transition recognition, sequence deviations, or rakaah counting.These tasks differ from identifying or verifying the person performing the prayer.
  • Prayer sensing: Reported prayer-recognition results often use small cohorts, constrained protocols, or sensing modalities unsuitable for private worship.Examples include seven-subject evaluation and camera-based posture recognition.
  • Motion biometrics: Gait and other motion-biometrics research uses inertial signals for identity inference across devices, body locations, speeds, and days.The literature includes similarity frameworks, frequency-domain features, wrist-worn systems, and deep temporal models.
  • Positioning: This work differs by targeting participant identity in a fully prescribed activity rather than posture recognition or naturally varying activity content.The prescribed sequence controls activity-content variation by design.
  • Positioning: The study compares rotation-invariant and Qibla-referenced coordinate treatments under an exactly matched evaluation protocol.It also separates unit-level from broader session-level evaluation.

III. THE DATASET

The dataset consists of smartphone IMU recordings from congregational Taraweeh prayer, segmented into structurally complete two-rakaah units and evaluated at unit and complete-session levels. The preprocessing provides both invariant and directional motion representations.

  • Data collection: Recordings were collected during congregational Taraweeh prayer with participant-carried smartphones in front trouser pockets.Acceleration, gyroscope, gravity, and orientation streams were recorded at 100 Hz.
  • Data collection: The final cohort contained 95 participants using heterogeneous consumer smartphone hardware without study-specific cross-device calibration.This improves ecological validity but leaves device-dependent effects entangled with participant identity.
  • Segmentation: 785 of 910 candidate units passed structural quality control, and 780 feature-valid units from 195 recordings entered the principal analyses.A retained unit required two valid rakaah sequences with the expected bowing and prostration events.
  • Evaluation populations: Complete-session holdout used 132 held-out recordings and 527 query units from 32 participants with repeated independent enrollment recordings.Single-session participants could not serve as cross-session query identities.
  • Representations: The Qibla-referenced approach explicitly estimates attitude and rotates motion into a common physical frame, retaining direction while depending on orientation conventions and heading reference.The two strategies therefore address the same phone-orientation nuisance with different assumptions.
  • Representations: The rotation-invariant representation analytically preserves movement intensity and gravity-relative structure without orientation metadata.Its five signals are preserved under common proper rotations and support the principal ablation studies.

B. Qibla-Referenced IMU Representation

The Qibla-referenced representation harmonizes platform conventions, reconstructs device attitude, corrects Android’s magnetic-north reference, and expresses acceleration and angular velocity in Qibla-left-up coordinates.

  • Orientation reconstruction: The pipeline constructs a device-to-platform-world rotation from platform-specific yaw, pitch, and roll conventions.Native non-standardized iOS and Android or standardized iOS use different sign conventions before rotation construction.
  • Orientation reconstruction: Orientation samples are converted to unit quaternions and synchronized independently to sensor timestamps using spherical linear interpolation.Interpolation is not bridged across orientation gaps exceeding 0.25 s.
  • North harmonization: The platform world frame is north–west–up, but its horizontal reference differs because iOS uses true north while Android uses magnetic north.This distinction is necessary before constructing a globally consistent physical frame.
  • North harmonization: For Android recordings, WMM declination correction removes the magnetic-versus-true-north reference difference without reconstructing or calibrating the handset magnetometer.The collection-site declination was approximately 3.68° during the collection period.
  • Qibla frame: The final Qibla-left-up channels are [aQ, aL, aV, ωQ, ωL, ωV], combining directional acceleration and angular velocity.Q points toward Qibla, L points horizontally left when facing Qibla, and V points upward.

A. Feature Extraction

Complete prayer units are converted into windowed descriptor representations for biometric analysis, with separate feature dimensionalities for the invariant and Qibla-referenced routes.

  • A. Feature Extraction: Each complete unit is divided into 5-s windows with 50% overlap, and 20 descriptors are computed per channel.Descriptors cover central tendency, variability, range, distribution shape, first-difference dynamics, and spectral structure.
  • A. Feature Extraction: The rotation-invariant representation yields 100 features per window, while the Qibla-referenced representation yields 120.These dimensionalities correspond to five and six channels, respectively.
  • A. Feature Extraction: Window feature vectors are averaged across each unit’s valid windows to form one unit-level feature vector.
  • A. Feature Extraction: Identification assigns a query to one enrolled participant, whereas verification tests a proposed identity claim for a query–enrollment pair.Pairwise verification may use fixed similarity scores or a trained binary classifier.
  • A. Feature Extraction: All identification results use closed-set evaluation, so queries are assumed to belong to enrolled identities and open-set rejection is not evaluated.
  • A. Feature Extraction: Unit-level analysis treats each complete two-rakaah unit as independent, while complete-session analysis holds out whole recordings to test cross-session behavior.The protocols measure distinctiveness and cross-session persistence under different split structures.

C. Unit-Level Protocol

The unit-level protocol treats every complete two-rakaah unit as a behavioral sample and evaluates retrieval, multiclass identification, and pairwise verification with learned and fixed comparisons.

  • C. Unit-Level Protocol: Every complete unit is treated as one behavioral sample in the principal representation-comparison and ablation setting.
  • C. Unit-Level Protocol: Each eligible unit is queried once, excluded from its own gallery, and evaluated with Pearson correlation, cosine similarity, and Euclidean distance.
  • C. Unit-Level Protocol: Logistic Regression, Linear SVM, and Random Forest learn participant class boundaries under leave-one-unit-out evaluation.Identity is assigned by a learned multiclass decision function rather than nearest-sample retrieval.
  • C. Unit-Level Protocol: Pairs are labeled genuine when they belong to the same participant and impostor otherwise, with Pearson, cosine, and Euclidean scores evaluated without learning.
  • C. Unit-Level Protocol: Participant-template evaluation averages remaining enrollment units for each participant before comparing the query with one template per participant.This reduces unit-specific variation relative to single-unit gallery matching.

D. Complete-Session Holdout Protocol

Complete-session holdout defines each nightly recording as the behavioral sample and prevents query-session information from entering enrollment, training, templates, or learned pairs.

  • D. Complete-Session Holdout Protocol: Each fold holds out every unit from one recording simultaneously, and query eligibility requires an independent enrollment recording for that participant.
  • D. Complete-Session Holdout Protocol: Nearest-neighbor matching uses units from other recordings, while participant templates are formed only from non-query sessions.
  • D. Complete-Session Holdout Protocol: Learned classifiers use all non-query recordings and are applied to every unit in the held-out session.
  • D. Complete-Session Holdout Protocol: Window CNN and SimCLR are evaluated under complete-session holdout using identical session partitions and enrollment definitions after unit-level comparisons.The broader temporal-learning branch also includes full-unit CNN, BYOL, and DINO at unit level.
  • D. Complete-Session Holdout Protocol: Verification is reported with ROC AUC, while decidability summarizes genuine–impostor score separation in standardized units.The genuine and impostor statistics are represented by their respective means and standard deviations.
  • D. Complete-Session Holdout Protocol: 95% confidence intervals use participant-clustered bootstrap resampling, and learned pairwise Random Forest point estimates average three random seeds.
  • D. Complete-Session Holdout Protocol: Table II summarizes handcrafted results for both unit-level and complete-session holdout analyses across the two IMU representations.
  • D. Complete-Session Holdout Protocol: 0.9700/7.62% is the invariant route’s complete-session learned pairwise Random Forest AUC/EER, versus 0.9618/7.81% for the Qibla-referenced route.The reported comparison indicates similar or slightly stronger complete-session verification for the invariant route.

VII. WHERE IS THE IDENTITY INFORMATION?

Ablations indicate that participant identity is distributed across complementary signal, descriptor, prayer-component, and physical-direction information rather than concentrated in one cue. The complete six-channel Qibla-referenced representation remains strongest overall, while selected dynamic and vertical components retain substantial discriminative value.

  • Descriptor-Family Ablation: Frequency-domain features achieve the highest complete-session fixed-score AUC of 0.8958, showing that rhythm and temporal regularity contribute beyond average movement magnitude.
  • Prayer-Component Ablation: The prostration-phase subset uses 22.6% of retained labeled duration yet approaches the full representation on several cross-session metrics, while Sujud-only matches complete-session ML model-based accuracy.The prostration phase contains two Sujud events and the intervening Julus event.
  • Prayer-Component Ablation: Participant information is concentrated in selected dynamic prayer components rather than being proportional to observation time.
  • Directional Attribution: The complete six-channel Qibla-referenced representation remains strongest overall, indicating complementary information across direction and sensor type.No single physical axis explains the complete result.
  • Signal-Family Ablation: Acceleration preserves more complete-session ML model-based verification performance than angular velocity alone, reaching AUC/EER 0.9560/8.25% versus 0.9463/10.67%.
  • Directional Attribution: Vertical motion is the strongest single-direction representation, whereas the Q direction is weakest; the horizontal Q–L plane reaches fixed-score AUC 0.8218.

VIII. SUPERVISED AND SELF-SUPERVISED LEARNING (SSL)-BASED RESULTS

The learned-representation experiments compare supervised CNNs with self-supervised temporal encoders under unit-level and complete-session protocols. Window-based learning and participant-template aggregation improve or preserve verification structure, while Qibla-referenced SimCLR achieves strong unit-level template performance.

  • Supervised Learning: Window CNN substantially outperforms full-unit CNN, supporting local pattern learning followed by aggregation for variable-duration prayer units.
  • Self-Supervised Learning: SimCLR is strongest among rotation-invariant SSL encoders, reaching participant-template AUC/EER 0.9699/8.52% and unit-to-unit Top-1 of 91.75%.
  • Qibla-Referenced SimCLR: Qibla-referenced SimCLR reaches participant-template AUC 0.9805–0.9817 and EER 5.81–6.24% across two unit-level runs.Template Top-1 rises to 82.36–83.41%, while direct unit-to-unit Top-1 remains comparable to rotation-invariant SimCLR.
  • Complete-Session Holdout: Under complete-session holdout, SimCLR reaches unit-to-unit Top-1 of 72.03% versus 65.58% for Window CNN, while Window CNN provides stronger unit-level verification.
  • Complete-Session Holdout: Participant-template aggregation improves verification for both learned encoders under complete-session holdout.The results are consistent with averaging suppressing unit-specific variation.
  • Complete-Session Holdout: Unit-level and complete-session protocols together show that learned cross-session identity structure is not restricted to handcrafted descriptor definitions.

B. Rotation Invariance Versus Physical Canonicalization

The study finds participant-discriminative motion structure under both rotation-invariant and Qibla-referenced coordinate assumptions, while showing that directional canonicalization is useful but not uniformly superior. Ablations and limitations clarify where discrimination appears and how hardware and threat-model boundaries constrain interpretation.

  • Directional ablations: The complete six-channel Qibla representation performs best overall, with vertical motion strongest individually and acceleration more informative than angular velocity for learned verification.Participant information is distributed across complementary motion directions rather than concentrated in the Qibla axis.
  • Representation comparison: Qibla-referenced canonicalization improves selected learned and fixed-score metrics but does not uniformly improve learned pairwise verification.The matched comparison supports usefulness without establishing universal superiority over the rotation-invariant representation.
  • Verification design: Learned pairwise verification outperforms fixed similarity under both protocols, indicating that identity structure is distributed across complementary feature interactions.The learned pair exposes feature-wise differences and products alongside Pearson, cosine, and Euclidean comparisons.
  • Interpretive boundaries: The heterogeneous-phone collection prevents quantifying how much discrimination reflects human movement versus stable handset characteristics.Complete-session holdout is session-disjoint but not device-disjoint; crossed participant-device recordings are proposed for separation.
  • Threat-model boundaries: The experiments model zero-effort impostors and do not establish resistance to targeted imitation, replay, sensor injection, adversarial spoofing, or open-set rejection.Strong known-participant identification is therefore narrower than deployment-ready authentication evidence.
  • Information localization: Signal, descriptor, and prayer-component analyses localize identity information to acceleration/gravity-relative dynamics, compact temporal and spectral features, and selected bowing or prostration intervals.The strongest performance still requires complementary motion across all physical axes.

APPENDIX A HANDCRAFTED FEATURE DEFINITIONS

Appendix A defines descriptors computed from each derived signal within a 5-s window, including statistical, first-difference, and frequency-domain features. Applying the descriptor set yields 100 features for five invariant signal families and 120 for six Qibla-referenced channels.

  • Descriptor inputs: Each derived signal is represented by samples within a 5-s window, from which population statistics, quantiles, first differences, and spectral descriptors are computed.The frequency-domain descriptors remove the signal mean, use a one-sided Fourier power spectrum, and exclude the zero-frequency component.
  • Feature dimensionality: 100 features per window result from applying 20 descriptors to five invariant signal families, while six Qibla-referenced channels produce 120 features.The feature counts reflect the number of descriptors multiplied by the number of derived channels.
Loading 2608.26994v1…