Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Real-Time Lightweight Gaze Privacy-Preservation Techniques Validated via Offline Gaze-Based Interaction Simulation

Mehedi Hasan Raju, Oleg V. Komogortsev

TL;DR

This work tackles the privacy risks of gaze data in XR by evaluating real-time, lightweight privacy-preserving techniques within an offline gaze-based interaction simulation. It introduces a three-part methodology—Data Privatization, Privacy Metric Computation, and Utility Measurement—and analyzes seven privacy methods, with Kalman-filter smoothing emerging as the most balanced approach, substantially reducing identifiability while preserving interaction quality. The study leverages the large-scale GazeBase dataset to demonstrate privacy gains without sacrificing real-time usability, highlighting the role of signal processing in privacy preservation. The findings have practical implications for designing privacy-aware gaze interfaces that maintain user experience in real-time applications like foveated rendering and gaze-driven interaction.

Abstract

This study examines the effectiveness of the real-time privacy-preserving techniques through an offline gaze-based interaction simulation framework. Those techniques aim to reduce the amount of identity-related information in eye-tracking data while improving the efficacy of the gaze-based interaction. Although some real-time gaze privatization methods were previously explored, their validation on the large dataset was not conducted. We propose a functional framework that allows to study the efficacy of real-time gaze privatization on an already collected offline dataset. The key metric used to assess the reduction of identity-related information is the identification rate, while improvements in gaze-based interactions are evaluated through signal quality during interaction. Our additional contribution is the employment of an extremely lightweight Kalman filter framework that reduces the amount of identity-related information in the gaze signal and improves gaze-based interaction performance.

Real-Time Lightweight Gaze Privacy-Preservation Techniques Validated via Offline Gaze-Based Interaction Simulation

TL;DR

This work tackles the privacy risks of gaze data in XR by evaluating real-time, lightweight privacy-preserving techniques within an offline gaze-based interaction simulation. It introduces a three-part methodology—Data Privatization, Privacy Metric Computation, and Utility Measurement—and analyzes seven privacy methods, with Kalman-filter smoothing emerging as the most balanced approach, substantially reducing identifiability while preserving interaction quality. The study leverages the large-scale GazeBase dataset to demonstrate privacy gains without sacrificing real-time usability, highlighting the role of signal processing in privacy preservation. The findings have practical implications for designing privacy-aware gaze interfaces that maintain user experience in real-time applications like foveated rendering and gaze-driven interaction.

Abstract

This study examines the effectiveness of the real-time privacy-preserving techniques through an offline gaze-based interaction simulation framework. Those techniques aim to reduce the amount of identity-related information in eye-tracking data while improving the efficacy of the gaze-based interaction. Although some real-time gaze privatization methods were previously explored, their validation on the large dataset was not conducted. We propose a functional framework that allows to study the efficacy of real-time gaze privatization on an already collected offline dataset. The key metric used to assess the reduction of identity-related information is the identification rate, while improvements in gaze-based interactions are evaluated through signal quality during interaction. Our additional contribution is the employment of an extremely lightweight Kalman filter framework that reduces the amount of identity-related information in the gaze signal and improves gaze-based interaction performance.

Paper Structure

This paper contains 33 sections, 12 equations, 2 figures, 1 table.

Table of Contents

  1. INTRODUCTION
  2. BACKGROUND AND PRIOR WORK
  3. METHODOLOGY
  4. Data Privatization
  5. Dataset
  6. Data Preprocessing
  7. Privatization Techniques
  8. Median Filter
  9. Temporal Downsampling
  10. Gaussian Noise
  11. Linear Weighted Moving Average Smoothing
  12. Targeted Noise Addition
  13. Causal Finite Impulse Response Filter
  14. Kalman Filter--Based Data Smoothing
  15. Privacy Metric Computation
  16. ...and 18 more sections

Figures (2)

  • Figure 1: Workflow Diagram of the Methodology.
  • Figure 2: Exemplar comparison of privacy preservation techniques. Each technique is applied to the same two-second segment of raw data. Downsampling reduces the number of samples while preserving the structure of the segment, which is why it ends earlier.