In recent years, eye-tracking data have been frequently used for precise modeling of user states, including cognitive load, physical load, and cybersickness in virtual reality (VR). However, these data can also expose sensitive biometric and behavioral signatures of the users that allow for re-identification and the inference of sensitive demographics via linkage attacks. Existing privacy-preserving methods (e.g., differential privacy, federated learning, data anonymization, etc.) often compromise the fidelity of user state estimation, creating a research gap in balancing privacy with application utility. To address this, we propose a novel model that mitigates privacy leakage from eye-tracking data while preserving high accuracy in user state prediction. Our approach ranks eye-tracking features by their contribution to prediction and applies selective perturbation to high-risk features. Experimental results show significant reductions in demographic inference accuracy (gender: 95.4% to 60.0%, race: 86.3% to 49.7%, age: 76.3% to 55.3%), with only a 9.7% average accuracy drop for cognitive load, physical load, and cybersickness classification. Compared to a local differential privacy baseline, our method achieves higher utility preservation and stronger demographic suppression, yielding improvements of 19.1%, 20.6%, and 18.3% for age, gender, and race, respectively. These findings validate the potential for privacy-aware modeling in VR systems and offer a scalable path toward ethical, secure, and high-performance deployment of eye-tracking technology in real-world applications.

Extended Reality (XR) systems are gaining widespread usability due to their impressive adaptability capabilities; yet, the critical challenge they face is user privacy. The subtle behavioral patterns revealed through human motion in XR environments serve as significant indicators that can be used to infer user characteristics. Accordingly, gender classification utilizing behavioral motion data has emerged as a critical research focus in this field. Traditional machine learning models have achieved success using gait or inertial data in surveillance and mobile applications. However, there is still limited work that targets XR-specific motion signals captured through headsets, controllers, and body tracking devices. The proposed interpretable framework, "YouKnowWho," aims to address this challenge by developing an explainable model for classifying gender from behavioral motion data. The framework consists of two sequential deep learning architectures: Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU). The performance of this framework is assessed using accuracy, precision, recall, and F1-score. Additionally, Integrated Gradients are used to analyze the importance of each feature in predicting gender. The results demonstrate that YouKnowWho is capable of classifying gender from behavioral motion data with a remarkable accuracy of 68.2%, and it can identify specific features that can be preserved or safeguarded to ensure user privacy.