Privacy-preserving datasets of eye-tracking samples with applications in XR

Area of Science:

• Computer Science
• Human-Computer Interaction
• Cybersecurity

Background:

• Virtual and mixed-reality (XR) technologies are rapidly advancing, impacting work, education, and entertainment.
• Eye-tracking data is essential for XR interactions, avatar animation, and performance optimizations.
• However, eye-tracking data poses privacy risks, enabling user re-identification.

Purpose of the Study:

• To evaluate privacy-preserving techniques for eye-tracking data in XR.
• To compare k-anonymity, plausible deniability (PD), and differential privacy (DP) in terms of privacy and utility.
• To minimize identification rates while maintaining the performance of machine learning models.

Main Methods:

• Applied k-anonymity and plausible deniability (PD) privacy definitions to eye-tracking datasets.
• Evaluated these methods against the state-of-the-art differential privacy (DP) approach.
• Processed two VR datasets to assess re-identification rates and model performance.

Main Results:

• Both PD and DP mechanisms demonstrated practical privacy-utility trade-offs for re-identification and activity classification.
• K-anonymity showed superior performance in retaining utility for gaze prediction tasks.
• The study quantified the impact of privacy methods on machine learning model accuracy.

Conclusions:

• Plausible deniability and differential privacy offer viable solutions for securing XR eye-tracking data.
• K-anonymity is particularly effective when preserving the utility of gaze prediction is paramount.
• Balancing privacy and utility is achievable for sensitive eye-tracking datasets in XR environments.