Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Privatization of Synthetic Gaze: Attenuating State Signatures in Diffusion-Generated Eye Movements

Kamrul Hasan, Oleg V. Komogortsev

TL;DR

The paper tackles privacy concerns in gaze data by evaluating diffusion-based synthetic gaze (DiffEyeSyn) for privatization of internal states. It generates subject-specific 5-second gaze sequences conditioned on velocity representations and embeddings, then extracts 58 eye-movement features and computes Spearman correlations with self-reported fatigue and task difficulty across HSS, RAN, and TEX tasks in the GazeBase dataset. It finds that synthetic gaze preserves spatial accuracy and oculomotor dynamics (e.g., spatial errors ~3.7–4.1°, RMS 0.01, embedding cosines ~0.92–0.95) but substantially weakens or eliminates state-related correlations, indicating effective privatization of internal states. The work contributes a modular DiffEyeSyn-based pipeline and a 58-feature correlation framework, highlighting a practical path toward privacy-preserving gaze datasets while acknowledging trade-offs for tasks requiring state decoding and proposing future disentanglement strategies to balance fidelity and privacy. This has implications for gaze-enabled interfaces and benchmarking where protecting user state information is critical.

Abstract

The recent success of deep learning (DL) has enabled the generation of high-quality synthetic gaze data. However, such data also raises privacy concerns because gaze sequences can encode subjects' internal states, like fatigue, emotional load, or stress. Ideally, synthetic gaze should preserve the signal quality of real recordings and remove or attenuate state-related, privacy-sensitive attributes. Many recent DL-based generative models focus on replicating real gaze trajectories and do not explicitly consider subjective reports or the privatization of internal states. However, in this work, we consider a recent diffusion-based gaze synthesis approach and examine correlations between synthetic gaze features and subjective reports (e.g., fatigue and related self-reported states). Our result shows that these correlations are trivial, which suggests the generative approach suppresses state-related features. Moreover, synthetic gaze preserves necessary signal characteristics similar to those of real data, which supports its use for privacy-preserving gaze-based applications.

Privatization of Synthetic Gaze: Attenuating State Signatures in Diffusion-Generated Eye Movements

TL;DR

The paper tackles privacy concerns in gaze data by evaluating diffusion-based synthetic gaze (DiffEyeSyn) for privatization of internal states. It generates subject-specific 5-second gaze sequences conditioned on velocity representations and embeddings, then extracts 58 eye-movement features and computes Spearman correlations with self-reported fatigue and task difficulty across HSS, RAN, and TEX tasks in the GazeBase dataset. It finds that synthetic gaze preserves spatial accuracy and oculomotor dynamics (e.g., spatial errors ~3.7–4.1°, RMS 0.01, embedding cosines ~0.92–0.95) but substantially weakens or eliminates state-related correlations, indicating effective privatization of internal states. The work contributes a modular DiffEyeSyn-based pipeline and a 58-feature correlation framework, highlighting a practical path toward privacy-preserving gaze datasets while acknowledging trade-offs for tasks requiring state decoding and proposing future disentanglement strategies to balance fidelity and privacy. This has implications for gaze-enabled interfaces and benchmarking where protecting user state information is critical.

Abstract

The recent success of deep learning (DL) has enabled the generation of high-quality synthetic gaze data. However, such data also raises privacy concerns because gaze sequences can encode subjects' internal states, like fatigue, emotional load, or stress. Ideally, synthetic gaze should preserve the signal quality of real recordings and remove or attenuate state-related, privacy-sensitive attributes. Many recent DL-based generative models focus on replicating real gaze trajectories and do not explicitly consider subjective reports or the privatization of internal states. However, in this work, we consider a recent diffusion-based gaze synthesis approach and examine correlations between synthetic gaze features and subjective reports (e.g., fatigue and related self-reported states). Our result shows that these correlations are trivial, which suggests the generative approach suppresses state-related features. Moreover, synthetic gaze preserves necessary signal characteristics similar to those of real data, which supports its use for privacy-preserving gaze-based applications.
Paper Structure (15 sections, 4 figures)

This paper contains 15 sections, 4 figures.

Table of Contents

  1. Introduction
  2. Related Works
  3. Synthetic Gaze Generation
  4. Privatization of Gaze
  5. Implemented Architectures
  6. Synthetic Gaze Generation
  7. Subjective Correlations
  8. Experiments
  9. Dataset
  10. Implementation Details
  11. Results
  12. Synthetic Signal Quality
  13. Correlation between Eye Movement Features and Subjective Reports
  14. Discussion
  15. Conclusions

Figures (4)

  • Figure 1: Overview of the architecture for the correlation between subjective reports and objective features: The real positional gaze signal is fed through the updated DiffEyeSyn framework to generate a synthetic gaze signal, from which features are extracted using a Feature Extractor. Finally, extracted features and subjective reports are analyzed by a Correlation Analyzer to produce the final correlation output.
  • Figure 2: Spearman correlations between synthetic gaze-derived features and subjective ratings for the HSS task.
  • Figure 3: Spearman correlations between synthetic gaze-derived features and subjective ratings for the RAN task.
  • Figure 4: Spearman correlations between synthetic gaze-derived features and subjective ratings for the TEX task.