Abstract
With eye tracking being increasingly integrated into virtual and augmented reality (VR/AR) head-mounted displays, preserving users' privacy is an ever more important, yet under-explored, topic in the eye tracking community. We report a large-scale online survey (N=124) on privacy aspects of eye tracking that provides the first comprehensive account of with whom, for which services, and to what extent users are willing to share their gaze data. Using these insights, we design a privacy-aware VR interface that uses differential privacy, which we evaluate on a new 20-participant dataset for two privacy sensitive tasks: We show that our method can prevent user re-identification and protect gender information while maintaining high performance for gaze-based document type classification. Our results highlight the privacy challenges particular to gaze data and demonstrate that differential privacy is a potential means to address them. Thus, this paper lays important foundations for future research on privacy-aware gaze interfaces.
Introduction
As eye tracking becomes a standard feature in VR and AR head-mounted displays, we are entering a privacy minefield. Eye movements reveal far more than just where a user is looking; they leak "biometric breadcrumbs" such as gender, health status, and identity. Unlike a mouse click, eye movements are largely involuntary and difficult to "fake." Research has shown that gaze patterns can reveal a user's personality, cognitive load, and even early signs of neurological disorders like Alzheimer’s or Parkinson’s.
The challenge is a classic "utility vs. privacy" trade-off. Users want the benefits of eye tracking—such as hands-free interfaces or reading assistants—without handing over a digital blueprint of their identity to third parties. This research presents a framework using Differential Privacy (DP) to add strategic noise to gaze data, masking private user attributes while maintaining high accuracy for functional tasks.
Methodology
The researchers employed a two-pronged approach to address these privacy concerns. First, a large-scale survey of 124 users established the "social ground truth" regarding what data people are willing to share. Second, they developed a technical framework applying DP to eye-tracking features.
Instead of noising raw (x, y) gaze coordinates, which are susceptible to smoothing algorithms, the authors applied the Exponential Mechanism to 52 aggregated statistical features, such as fixation duration and blink rate. This approach was tested against two threat models: attackers without prior knowledge (guessing general attributes like gender) and attackers with prior knowledge (attempting re-identification using existing "clean" data).
The experimental evaluation utilized a new dataset of 20 participants reading three document types (Comics, Newspapers, Textbooks) while wearing a VR headset equipped with an eye tracker. This dataset served as the ground truth for both the utility task (document classification) and the privacy tasks (gender and identity inference).
Key Results and Insights
The study revealed a "Privacy Paradox": while over 80% of users would share gaze data for health-related disease detection, they are fiercely protective against commercial profiling regarding sexual orientation, race, and identity.
Technically, the authors identified a "sweet spot" for the privacy-loss parameter, $\epsilon$. At $\epsilon = 15$, the system effectively "blinded" gender-prediction AI (dropping accuracy to 54%, near random chance) while allowing document-classification AI to maintain high performance (~55% accuracy, well above the 33% chance level). In re-identification scenarios, DP was even more effective, dropping an attacker's success rate from a high baseline to just 6.4%, while document classification utility remained robust at ~70% accuracy.
Conclusion
This work demonstrates that raw gaze streams are a significant liability and that shifting to aggregated features makes them easier to protect with noise. Even when AI is trained on "clean" data, applying DP noise during testing provides substantial protection. Future work will focus on scaling these findings to larger datasets, exploring different VR contexts beyond reading, and implementing these protections in real-time, low-latency scenarios.
Summary updated