Evaluating Eye Movement Biometrics in Virtual Reality: A Comparative Analysis of VR Headset and High-End Eye-Tracker Collected Dataset

TL;DR

The paper addresses biometric authentication using eye movement data collected from an eye-tracking-enabled VR headset (GBVR) and compares it to a high-end EyeLink-based dataset (GB) by downsampling to 250 Hz. It uses the Eye Know You Too (EKYT) DenseNet-based embedding network to produce 128-dimensional eye-movement embeddings from velocity channels, evaluated across monocular and binocular configurations with 4-fold cross-validation and held-out subjects. Key findings show GBVR binocular data achieve an excellent short-term $EER$ of $1.67 ext{%}$ and robust $d'$ while FRR remains a challenge, whereas long-term performance degrades; GB generally remains strongest, with GB-250Hz offering a different set of trade-offs (lower FRR but higher $EER$). The results demonstrate the biometric viability of ET-enabled VR headset data for EMB and highlight the advantage of binocular data in VR contexts, outlining practical considerations for deployment in authentication systems.

Abstract

Previous studies have shown that eye movement data recorded at 1000 Hz can be used to authenticate individuals. This study explores the effectiveness of eye movement-based biometrics (EMB) by utilizing data from an eye-tracking (ET)-enabled virtual reality (VR) headset (GazeBaseVR) and compares it to the performance using data from a high-end eye tracker (GazeBase) that has been downsampled to 250 Hz. The research also aims to assess the biometric potential of both binocular and monocular eye movement data. GazeBaseVR dataset achieves an equal error rate (EER) of 1.67% and a false rejection rate (FRR) at 10^-4 false acceptance rate (FAR) of 22.73% in a binocular configuration. This study underscores the biometric viability of data obtained from eye-tracking-enabled VR headset.

Figures (6)

• Figure 1: Simple block diagram of the methodology.
• Figure 2: Comparison of Kernel Density Estimations of Spatial Precision and Accuracy Measurements across datasets: GB, GB-250Hz (downsampled), and GBVR. (A) Illustrates the distribution of spatial precision measurements highlighting their respective medians. (B) Displays the distribution of spatial accuracy, also indicating median values. The plots demonstrate the variability within each dataset and visually compare the central tendencies and distribution shapes between the datasets.
• Figure 3: Similarity score distribution for GBVR monocular study (left) and binocular study (right).
• Figure 4: ROC Curve for GBVR monocular study (left) and binocular study (right).
• Figure 5: Similarity score distribution for GB-250Hz (left) and GBVR binocular study (right).