Privacy-aware eye tracking using differential privacy

Julian Steil, Inken Hagestedt, Michael Xuelin Huang, Andreas Bulling

TL;DR

Eye tracking in VR/AR headsets reveals far more than just where you are looking—it can leak sensitive "biometric breadcrumbs" like your gender, health status, and identity. This paper presents a framework using Differential Privacy (DP) to add strategic noise to gaze data, successfully masking private user attributes while maintaining high accuracy for functional tasks like document classification.

What Problem Does This Paper Address?

As eye tracking becomes a standard feature in VR and AR head-mounted displays, we are entering a privacy minefield. 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. We want the benefits of eye tracking—such as hands-free interfaces or reading assistants that know when you're struggling with a word—but we don't want to hand over a digital blueprint of our identity to every app developer or advertiser. The authors seek to bridge this gap by providing a formal mathematical guarantee of privacy that doesn't break the underlying application.

Methods

The researchers took a two-pronged approach:

  1. Large-Scale Survey: They first surveyed 124 users to understand what people are actually willing to share. This established the "social ground truth" for their privacy model.
  2. Differential Privacy (DP) Framework: The core technical contribution is the application of DP to eye-tracking features. Instead of noising raw (x, y) gaze coordinates—which can often be "cleaned" by simple smoothing algorithms—they applied the Exponential Mechanism to 52 aggregated statistical features (e.g., fixation duration, saccade velocity, blink rate).
  3. Threat Modeling: They tested their system against two types of attackers:
  4. Without Prior Knowledge: An attacker trying to guess a general attribute (like gender) from a pool of users.
  5. With Prior Knowledge: A more dangerous attacker who already has some of your "clean" data and is trying to re-identify you in a new, noised dataset.

Datasets

  • Survey Dataset: Responses from 124 participants regarding privacy preferences across various services and environments.
  • Experimental Dataset: A newly recorded dataset of 20 participants (10 male, 10 female). Participants wore an Oculus DK2 headset with a Pupil Labs eye tracker while reading three distinct document types: Comics, Newspapers, and Textbooks. This dataset provided the ground truth for both the "utility" task (document classification) and the "privacy" tasks (gender and identity inference).

Key Insights and Ideas

  • The Privacy Paradox: The survey revealed that users are surprisingly altruistic regarding health; over 80% would share gaze data to detect diseases like Parkinson's. However, they are fiercely protective against commercial profiling, specifically regarding sexual orientation, race, and identity.
  • Raw Data is a Liability: The authors argue that raw gaze streams are nearly impossible to anonymize effectively. By shifting the focus to aggregated features (statistical summaries of movement over 30-second windows), they found a representation that is both useful for machine learning and easier to protect with noise.
  • The "Sweet Spot" of Epsilon ($\epsilon$): In Differential Privacy, $\epsilon$ controls the privacy-loss. The authors identified that at $\epsilon = 15$, they could effectively "blind" a gender-prediction AI (dropping it to random chance) while still allowing a document-classification AI to perform with high accuracy.

Key Results

  • Protecting Gender: For the "Without Prior Knowledge" model, applying DP reduced gender prediction accuracy to 54% (essentially a coin flip) while maintaining document classification accuracy at ~55% (well above the 33% chance level for three categories).
  • Preventing Re-identification: In the "With Prior Knowledge" scenario, DP was even more effective. It dropped an attacker's ability to re-identify a specific user from a high baseline down to just 6.4%, while the utility task (document type) remained robust at ~70% accuracy.
  • Clean vs. Noisy Training: The study found that even if an AI is trained on "clean" (non-noised) data, applying DP noise to the test data still provides significant privacy protection, which is a vital finding for securing existing eye-tracking systems.

Limitations and Future Work

  • Scale: The experimental evaluation was limited to 20 participants. While sufficient for a proof-of-concept, larger datasets are needed to see how these $\epsilon$ values scale across thousands of users.
  • Task Specificity: The study focused exclusively on reading tasks. Eye movement behavior varies wildly between reading, gaming, and social interaction in VR; the "sensitivity" of features in these different contexts remains an open question.
  • Real-time Implementation: The current method uses 30-second windows. Future work needs to explore how to apply DP in real-time, low-latency scenarios required for active gaze-contingent rendering or immediate UI feedback.

Summary updated

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.
BibTeX
@inproceedings{steil2019privacy,
  author = {Julian Steil and Inken Hagestedt and Michael Xuelin Huang and Andreas Bulling},
  title = {{Privacy-aware eye tracking using differential privacy}},
  booktitle = {ETRA},
  year = {2019},
  url = {https://openreview.net/forum?id=BpQw1wqUFV},
}

Loading related content...

Figure