User Identification with LFI-Based Eye Movement Data Using Time and Frequency Domain Features

2025-05-12

AI Summary

To address privacy leakage risks inherent in conventional camera-based eye-tracking systems, this paper proposes a novel laser interferometry-based eye movement biometrics paradigm. Leveraging high-sampling-rate, non-imaging raw eye movement signals acquired via laser interferometric measurement (LFI), we empirically demonstrate for the first time the inter-subject uniqueness of these signals as biometric traits. We design a lightweight, gaze-free time-frequency joint feature engineering pipeline—extracting velocity, displacement, and time-frequency spectral features—without requiring gaze estimation. A cross-task, multi-class SVM/Random Forest classification framework is then constructed. Under a 5-second sliding window, our method achieves an average identification accuracy of 93.14% on both static and dynamic tasks, with an equal error rate (EER) of only 2.52%, significantly outperforming existing non-imaging alternatives. This work establishes a deployable, contactless, and privacy-preserving authentication pathway for sensitive applications.

Technology Category

• Computer Vision: Motion & Tracking
• Machine Learning: Privacy
• Intelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

• Security and Privacy: Large-scale security measurements
• User Modeling, Personalization and Recommendation: User privacy protection in personalized systems
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics

Abstract

Laser interferometry (LFI)-based eye-tracking systems provide an alternative to traditional camera-based solutions, offering improved privacy by eliminating the risk of direct visual identification. However, the high-frequency signals captured by LFI-based trackers may still contain biometric information that enables user identification. This study investigates user identification from raw high-frequency LFI-based eye movement data by analyzing features extracted from both the time and frequency domains. Using velocity and distance measurements without requiring direct gaze data, we develop a multi-class classification model to accurately distinguish between individuals across various activities. Our results demonstrate that even without direct visual cues, eye movement patterns exhibit sufficient uniqueness for user identification, achieving 93.14% accuracy and a 2.52% EER with 5-second windows across both static and dynamic tasks. Additionally, we analyze the impact of sampling rate and window size on model performance, providing insights into the feasibility of LFI-based biometric recognition. Our findings demonstrate the novel potential of LFI-based eye-tracking for user identification, highlighting both its promise for secure authentication and emerging privacy risks. This work paves the way for further research into high-frequency eye movement data.

Problem

• Identifying users from LFI eye movement data using time-frequency features
• Developing multi-class model for biometric recognition without gaze data
• Assessing privacy risks and authentication potential of LFI eye-tracking

Innovation

• LFI-based eye-tracking for user identification
• Time and frequency domain feature analysis
• Multi-class classification model without gaze data