Motion-Based User Identification

March 2026

What is the paper about?

The paper investigates the identification of users across different eXtended Reality (XR) apps solely based on their motion data. Using information from head-mounted display and controller tracking, machine learning models are employed to recognize individuals from their unique motion patterns in different XR environments.

What are the results?

• Dataset:
  • A new XR motion dataset with over 60 hours of motion data from 49 users in five XR applications (four XR games with distinct tasks and action patterns and one social XR app without a predefined task)
• Models:
  • Similarity learning model: This models learns a representation (embedding) of motion sequences where movements from the same user are close together in a shared feature space, while movements from different users are further apart. Instead of directly predicting a user, it compares motion embeddings and determines whether two samples likely belong to the same person. A key advantage of this approach is that new users can be added without retraining the model, since their motion patterns can simply be embedded and compared to existing samples.
  • Classification model: This model directly predicts the identity of a user by assigning the motion input to one of the known users from the training set. It learns user-specific patterns during training and outputs the most likely identity as a classification result. This approach requires all users to be included in the training data and cannot naturally handle previously unseen users without retraining the model.
• Performance:
  • The similarity learning model achieved an average identification accuracy of 78.5% across all users and applications. When reference and query data came from the same application, accuracy was about 83%, but dropped to around 18%* when identifying users **across different applications.
  • With the classification model, 46.7% validation accuracy and 43.2% test accuracy for user identification was reached, with higher performance in applications that involve more consistent movements (e.g., rhythm games).
• Key findings:
  • Current state-of-the-art models are able to reliably identify users within a single VR application, but their ability to generalize across different XR applications remains limited.

What are possible fields of application?

• (Continuous) Biometric authentication in XR systems
• User identification for personalization and analytics
• Implicit security and access control

How does the research in the paper contribute to shaping the metaverse?

• Using motion patterns for identification: The research demonstrates that motion data from XR devices contains distinctive behavioral patterns that can be used to recognize individuals. This highlights motion tracking as a potential foundation for behavioral identity systems in immersive environments.
• Progress toward behavioral biometrics in XR: The results show that reliable user identification within single VR applications is already achievable. This indicates strong potential for motion-based biometrics as a future (continuous) authentication mechanism.
• Privacy awareness: The work highlights that motion tracking data can unintentionally reveal personal identities, raising important privacy considerations for future XR systems. Understanding these risks helps to guide the design of safer and privacy-aware metaverse platforms.
• Benchmark dataset for further research: By releasing a large dataset containing motion data from multiple XR applications, a valuable benchmark asset for developing and evaluating future motion-based identification and authentication methods in XR is provided.

Reference

Schach, L., Rack, C., McMahan, R. P. & Latoschik, M. E. (2026). Motion-based user identification across XR and metaverse applications by deep classification and similarity learning. Frontiers in Virtual Reality, 7. https://doi.org/10.3389/frvir.2026.1743491