PaperlayerView exact revision on arXiv

Source-linked AI summary

Measurement-Guided Conversion of Laboratory Eye-Tracking Data for VR Authentication

Md Arif Shahriar, Mehedi Hasan Raju, Oleg V. Komogortsev

arXiv:2610.30032v1cs.HC

TL;DR

Longitudinal gaze data for VR authentication are costly to collect, while lab-tracker recordings differ from lower-quality VR headset signals. NDR re-creates device effects in lab recordings while preserving eye movement; converted data match several VR signal properties and support authentication of unseen VR users.

  • Problem

    Longitudinal datasets needed to train eye-movement authenticators are slow and costly to collect, and lab-tracker signals differ in quality and device-specific noise from VR headset signals.

    Source: §1 · p. 1

  • Method

    NDR re-creates headset sampling, person-consistent calibration error, and spectrally shaped device noise, then uses a frequency-locked diffusion renderer to re-texture noise, using unpaired VR recordings from other people.

    Source: §1 · p. 3, §8 · p. 17

  • Results

    Converted recordings match GazeBaseVR in noise level, signal type, and accuracy, preserve source saccade dynamics, and are harder to distinguish from real VR data than plain lab recordings.

    Source: §7 · p. 16

  • Takeaways & Limitations

    Converted GazeBase recordings preserve source eye movement while reproducing several VR device signal properties, providing a route to use lab recordings in VR authentication research.

    Source: §7 · p. 16, §8 · p. 17

    Authentication was tested only on same-day sessions and one device pair and task, and the minimum VR data requirements for the twin and renderer remain unknown.

    Source: §7 · p. 17

Abstract

from arXiv · show

Eye-movement authentication in virtual reality (VR) needs longitudinal training data, yet most such datasets come from lab trackers. We convert lab recordings to a VR headset's signal quality while keeping their eye movement, using only unpaired VR recordings of other people. Measurements show that the headset's added noise dominates above 7.8 Hz and that its quality is consistent within a person. A parametric device model re-creates the headset's sampling, a person-consistent calibration error and spectrally shaped noise, and a frequency-locked diffusion renderer re-textures that high-frequency noise. Converted GazeBase recordings match GazeBaseVR in noise level and accuracy, keep their source's saccade dynamics, and are harder to tell from real VR data than two prior conversions. An authenticator trained only on converted data verifies unseen VR users at a 5.9% equal error rate with 45 s of gaze, close to real VR training (5.8%); with 5 s, a gap remains.

1 Introduction

Longitudinal authentication data are costly to collect, and lab recordings do not reproduce the device effects of VR headsets. The paper asks whether lab recordings can be converted to a target device’s signal quality while preserving eye movement, using unpaired VR recordings.

1 Introduction

  • Earlier conversions either add white noise while degrading each recording independently or generate VR-like signals with poor target accuracy.Neither approach both reproduces device signal characteristics and preserves source eye movement.Source: §1 · p. 2
  • The conversion question is whether lab recordings can match a target device’s signal quality while keeping their eye movement, without paired recordings of the same people.The assumed setup uses a higher-quality lab source and a lower-quality VR target.Source: §1 · p. 2
  • NDR models a VR signal as eye movement plus device effects, then re-creates camera sampling, calibration error, and spectrally shaped noise from a matched VR twin.A frequency-locked diffusion renderer re-textures high-frequency noise while leaving the signal below 7.8 Hz essentially unchanged.Source: §1 · pp. 2–3
  • Measurements identify device differences beyond standard quality measures, including noise spectrum and how quality varies within a person.The VR device’s added noise exceeds the lab tracker’s fixation spectrum from 7.8 Hz upward.Source: §1 · pp. 2–3
  • The evaluation uses 101 unseen real VR users and assesses signal quality, realism, movement preservation, prior conversions, and authentication utility.The downstream test trains an authenticator on converted recordings and tests it on unseen VR users.Source: §1 · p. 3

2 Related Work

Prior work characterizes device- and person-dependent eye-tracking quality, develops synthetic gaze and cross-device conversions, and uses diffusion models to edit signals. This paper situates conversion as a way to prepare laboratory recordings for downstream VR authentication.

2.1 Eye-tracking data quality across devices and people

  • Eye-tracking quality varies with device, person, and procedure, and its temporal noise structure includes signal type, magnitude, and spectral characteristics.The section frames quality measurement as a basis for reproducing measured differences between devices.Source: §2.1 · p. 3

2.2 Synthetic eye-movement data and cross-device conversion

  • Percentile matching degrades recordings toward target quality, while MUNIT-based conversion can lose spatial accuracy; this work instead models device effects explicitly.The cited approaches concern conversion between the same datasets [Griffith et al. 2025; Lohr et al. 2025].Source: §2.2 · p. 4

2.3 Diffusion models for editing signals

  • The renderer combines partial noising, low-frequency replacement from a reference, and frequency-shaped noise to edit existing signals rather than synthesize new gaze.The section relates these choices to SDEdit [Meng et al. 2021], ILVR, and frequency-shaped diffusion methods.Source: §2.3 · p. 4

2.4 Eye-movement biometrics as a downstream task

  • EKYT serves as the downstream authentication test; developing a new authentication method is not the paper’s goal.The authors use it to assess whether converted laboratory recordings provide useful training data.Source: §2.4 · p. 4

3 Data and the Device Gap

GazeBase and GazeBaseVR differ in several device-quality measures despite matching median saccade latency. The observed noise crossover and person-level consistency shape how the datasets are compared and how VR device profiles are assigned.

3.1 Datasets and split

  • GazeBase contributes 1,762 RAN recordings from 322 people; GazeBaseVR Set A supports development, while person-held-out Set B supplies final RAN evaluation.Set A contains 306 people and 756 RAN recordings; Set B contains 101 people and 248 RAN recordings.Source: §3.1 · pp. 4–5, Table 1 · p. 5

3.2 What separates the two devices

  • After resampling to 250 Hz, VR has stronger, whiter noise and twice the fixation spread, but fewer missing samples; median saccade latency matches.RMS-S2S is 0.156° vs. 0.029°, signal type 0.780 vs. 0.208, spectral slope α 0.635 vs. 2.273, MAD precision 0.175° vs. 0.086°, and missing samples 0.4% vs. 1.5% (VR vs. EyeLink); median saccade latency is 228 ms in both cohorts.Source: §3.2 · p. 5, Table 2

3.3 Where the added device noise takes over

  • The estimated VR-added noise crosses EyeLink’s fixation spectrum at 7.8 Hz, an empirical renderer cutoff under a comparable-fixation assumption, not a physiological boundary.The estimate is based on the difference between median spectra; the learned renderer leaves frequencies below the crossover essentially unchanged and re-textures those above it, while the parametric model adds noise at all frequencies.Source: §3.3 · pp. 5–6, Figure 2 · p. 6

3.4 Device quality is consistent within a person

  • 82% of VR noise-level variance, 73% of precision variance, and 75% of signal-type variance lie between people, supporting consistency chiefly within visits.Set A participants have 2.5 RAN recordings on average, mostly from two contiguous sessions in one visit, so these values describe within-visit consistency more than consistency over months.Source: §3.4 · pp. 6–7

4 Method

The method models a VR recording as eye movement plus device effects, then replaces the modeled effects in lab recordings while retaining their eye movement. A frequency-locked renderer refines the added noise texture, and converted recordings are evaluated for authentication on unseen VR users.

4.1 Overview: render the device, keep the eye movement

  • The pipeline retains the source eye movement while replacing modeled device effects, then evaluates converted recordings with EKYT on unseen GazeBaseVR users.The device model creates proposal u; the frequency-locked renderer produces the converted recording.Source: §4.1 · p. 7, §4.5 · p. 11

4.2 Part (a): Parametric device model

  • Camera exposure and irregular sampling are reproduced, and each GazeBase participant is assigned a VR twin using standardized participant signatures and optimal transport.Recording profiles map within-person deviations to the VR distribution, preserving participant consistency while reflecting VR session-to-session variation.Source: §4.2.1 · pp. 7, 9
  • Calibration replacement swaps the lab field for the twin’s field, while synthesized noise uses the estimated difference between VR and typical EyeLink spectra.The resulting proposal has the intended noise level and spectrum, but its noise is Gaussian and time-invariant.Source: §4.2.3 · p. 9, §4.2.4 · p. 10

4.3 Part (b): Frequency-locked diffusion renderer

  • The renderer replaces the proposal’s Gaussian, time-invariant texture with VR texture while locking its low-frequency content to the proposal.It refines overlapping windows and retains only their residuals, restoring missing-sample locations and applying the VR data-loss process afterward.Source: §4.3.1 · pp. 10–11

4.4 Part (c): The denoiser and its training

  • A denoiser trained on 3,761 usable set A recordings learns clean gaze crops from device-shaped noise, conditioned on noise level and ten texture features.At noise levels near the device level, it removes only a small fraction of the added noise.Source: §4.4.1 · p. 11

4.5 Part (d): Authentication with converted training data

  • EKYT is trained only on converted recordings and tested on real GazeBaseVR users from the unseen set B RAN task.This evaluates conversion in its intended use as training data for headset authentication.Source: §4.5 · p. 11

5 Experimental Setup

The experiments compare three GazeBase conversion conditions with real VR data on held-out Set B recordings, measuring signal quality, realism, retained eye-movement characteristics, and authentication performance.

5.1 Conditions

  • Plain used camera sampling only, Parametric used the device model, NDR added the renderer, and Real VR used GazeBaseVR set A; all were compared on Set B RAN.The conversions covered all 1,762 GazeBase RAN recordings.Source: §5.1 · p. 11

5.2 Training

  • The denoiser was trained on set A for 150,000 steps, while EKYT was separately trained per condition using four participant folds and the GazeBaseVR baseline settings.The denoiser used batches of 32 crops; four disjoint folds trained the EKYT models.Source: §5.2 · p. 12

5.3 Signal quality, realism and content

  • Signal quality was compared with Set B using medians and KS distances; realism used a six-measure nearest-neighbour test, person share used ICC(1), and content checks covered fixations and saccades.The nearest-neighbour test used five draws; saccade main-sequence comparisons used 150 recordings per set.Source: §5.3 · p. 12, Table 3 · p. 13

5.4 Authentication protocol

  • Four fold models were ensembled to verify 101 unseen Set B participants using 5- or 45-second gaze; reported metrics included EER, FRR at FAR 10^-4, rank-1, and bootstrap confidence intervals.Intervals resampled test participants, not training randomness.Source: §5.4 · p. 12

6 Results

NDR most closely matched VR noise while preserving source saccade dynamics, and its 45-second authentication performance was close to real-VR training. Prior-method comparisons and ablations identify the renderer’s gains and the need for device noise as input.

6.1 Signal quality and realism

  • NDR most closely matched VR noise (RMS-S2S 0.142° vs 0.147°; KS 0.06) and accuracy (KS 0.06), while nearest-neighbour distinguishability fell to 70.1%.Noise person share was similar (ICC 0.80 for NDR vs 0.84 for Set B), but precision remained too high (MAD 0.129° vs 0.166°) and data loss was 1.8% vs 0.3%.Source: §6.1 · p. 13

6.2 Authentication of real VR users

  • At 45 s, NDR's FRR at FAR 10^-4 was 64.4% versus 56.4% for real VR, and rank-1 identification was 80.2% versus 86.1%; FRR is exploratory.Parametric rank-1 identification was 82.2%.Source: §6.2 · p. 14
  • 5.9% EER with 45 s for NDR versus 5.8% with real-VR training; d′ was 3.21 versus 3.10, with paired difference +0.11 [−0.14, +0.35].At 5 s, all converted sets lagged real VR (NDR d′ 1.47 vs 1.86); at 45 s, NDR exceeded Parametric by +0.25 [+0.13, +0.38].Source: §6.2 · pp. 13–14, Figure 4 · p. 14

6.3 The coarse eye movement is retained

  • Saccade latency changed by at most 4 ms in 99.8% of recordings; converted main-sequence velocities tracked the source (204 vs 206°/s) but remained above slower VR saccades.Fixation counts were nearly unchanged: 96.7 after Plain and 96.8 after NDR.Source: §6.3 · p. 14, Figure 5 · p. 15

6.4 Comparison with prior conversion methods

  • PM best matched its targeted precision but added white noise; ST was overly smooth, while NDR exceeded PM in d′ at both durations and ST did no better than plain resampling.PM and ST were fitted to Set A only.Source: §6.4 · pp. 14–15, Table 5 · p. 15

6.5 Ablations

  • Removing Step-4 noise reduced d′ at both durations; random profiles hurt 5-s results, while replacing the typical spectrum did not help.The renderer can reshape device noise but cannot create it, so it needs Step-4 noise as its starting point.Source: §6.5 · p. 15, Table 6 · p. 16

7 Discussion

NDR brings converted-data authentication close to real-VR training with 45 seconds of gaze, while preserving source saccade dynamics; several signal mismatches and study boundaries remain.

7 Discussion

  • With 45 s of data, converted-only training reaches d′ 3.21 vs. 3.10 for real-VR training and a similar EER of 5.9 vs. 5.8%.With 5 s, every converted set remains below real-VR training.Source: §7 · p. 16
  • Converted gaze is too precise (MAD 0.13° vs. 0.17°), its saccades are faster than real-VR saccades, and its noise is stationary.The authors suggest missing temporal filtering may explain faster saccades; fine-scale detail and dataset population differences may contribute to the 5 s performance gap.Source: §7 · p. 16
  • The twin improves person-consistent noise and helps at 5 s, but 45 s random profiles have a lower EER (4.0 vs. 5.9%), with the difference including zero.The frequency lock preserves source eye movements but prevents correction of the low-frequency precision gap.Source: §7 · p. 16
  • Evidence is bounded by one training run per condition, unequal training-data sizes, same-day sessions, one device pair and task, and unknown minimum data requirements.The study also leaves transfer of relative device quality untested and does not quantify whether authenticators exploit the twin’s noise identity cue.Source: §7 · p. 17

8 Conclusion

NDR converts lab eye-tracking recordings toward VR headset signal quality while retaining their eye movement. Converted data match several VR signal measures and support authentication close to real-VR training at 45 seconds, though signal mismatches remain.

8 Conclusion

  • Converted GazeBase matches GazeBaseVR in noise level, signal type, and accuracy, and is harder to distinguish from real VR than resampled lab data.It also retains source saccade latency and dynamics.Source: §8 · p. 17
  • With 45 s of data, converted-only training reaches d′ 3.21 vs. 3.10 for real-VR training on unseen VR users.With 5 s, converted training remains below real-VR training (EER 22.5 vs. 16.8%).Source: §8 · p. 17
  • Converted gaze remains too precise, its saccades are faster than the VR device’s, and its noise does not change over time.NDR models device effects with a parametric model and a frequency-locked diffusion renderer that leaves the signal below f* essentially unchanged.Source: §8 · p. 17

Privacy and Ethics Statement

The study uses two public datasets whose participants consented to collection and de-identified sharing; no new data were collected.

Privacy and Ethics Statement

  • The study uses GazeBase and GazeBaseVR, public datasets distributed under the CC BY 4.0 license.Their participants consented to collection and de-identified sharing, and the study collected no new data.Source: Privacy and Ethics Statement · p. 17

Related papers

Loading 2610.30032v1…