Understanding User Identification in Virtual Reality through Behavioral Biometrics and the Effect of Body Normalization

Title of the Paper

Understanding User Identification in Virtual Reality Through Behavioral Biometrics and the Effect of Body Normalization

Paper Information

• Subject Area: User identification in virtual reality based on behavioral biometrics
• Keywords: Identification, virtual reality, behavioral biometrics, body normalization, usable security, data collection, deep learning, user experience, experimental study, continuous authentication

Research Background and Issues

• Problems and Challenges:

  1. Current user identification methods in virtual reality systems often rely on traditional approaches such as passwords and PINs, which suffer from poor user experience (interrupting immersion) and security vulnerabilities (e.g., shoulder surfing attacks).
  2. Behavioral biometric identification in virtual reality faces two core challenges: how to achieve identification through real tasks instead of artificial ones, and how users' physiological characteristics (e.g., height and arm length) affect the accuracy of the identification model.

• Significance:

  • Multi-user VR scenarios require fast, seamless, and secure user identification to enable adaptive user interfaces, load personalized settings, or provide authentication support.
  • Improving the accuracy of implicit authentication (i.e., requiring no active input from users) can reduce user burden and enhance the overall user experience.

• Research Motivation and Related Work:

  • Previous studies have shown that behavioral biometrics, such as head/controller movements, can be used for user authentication in VR, but their relatively low accuracy limits practical applications.
  • Some research has explored the impact of virtual body representation on user experience and behavior in VR, but its role in user identification has not been systematically evaluated.
  • Behavioral biometric verification heavily relies on differences in users' physiological characteristics. This study investigates the impact of normalizing users' height and arm length on identification accuracy.

Proposed Solution

• Methods and Solutions:

  • A task-driven implicit user identification system is proposed, collecting behavioral data as users perform everyday VR tasks (e.g., bowling, archery).
  • Introduced "body normalization" between physical controllers and virtual bodies by adjusting users' virtual height and arm length to create consistent virtual body proportions, thereby studying the influence of physiological characteristics.

• Innovations:

  1. Investigated the impact of virtual body normalization on user identification, a first in behavioral biometric research.
  2. Proposed a cross-day user identification method, enhancing the practical applicability and stability of the data.
  3. Utilized deep learning models (LSTM and Multi-Layer Perceptron (MLP)) for efficient data classification and compared the effects of different feature combinations.

• Implementation Steps:

  1. Designed realistic task scenarios for bowling and archery and implemented spatial motion data collection.
  2. Applied two body normalization methods:
     • Arm length normalization (standardizing virtual arm length).
     • Height normalization (setting virtual height to a fixed value).
  3. Conducted experiments:
     • Recruited 16 participants for lab experiments, with data collection spread over two days.
     • Tested four normalization conditions for each task (bowling and archery), collecting comprehensive behavioral data.
  4. Trained and validated the data using deep learning models, focusing on cross-day data variability.

Research Findings

• Specific Results:

  1. User identification accuracy was highest with height normalization, reaching 90% in the archery scenario, a 27-percentage-point improvement over no normalization.
  2. Normalization allowed the model to focus more on participants' behavioral differences rather than physiological noise.
  3. Among various feature combinations, using only the position vectors between the controller and the head-mounted device (F3 feature set) achieved the best identification results.

• Advantages:

  1. Significant performance improvement compared to non-normalized methods, especially in cross-day data validation.
  2. Users do not need to actively participate in the authentication process, greatly enhancing the system's seamless experience.

• Experimental and Evaluation Results:

  • The highest accuracy in the archery scenario was 90%, while in the bowling scenario, it was 68%.
  • Identification accuracy significantly improved in most cases after applying height normalization.
  • Although the majority of participants did not notice the normalization operation, a small number reported a sense of unnaturalness in the virtual body proportions during tasks.

• Limitations and Future Directions:

  1. Limitations:
     • Small sample size with only 16 participants, and a short experiment duration (two days).
     • The accuracy in the bowling scenario still requires further improvement.
     • The potential impact of virtual anthropomorphic effects has not been studied in conjunction with more complex tasks and environments.
  2. Future Directions:
     • Expand the experiment to include more users and scenarios.
     • Explore the effects of other types of virtual body adjustments on identification performance.
     • Test the concealment and user acceptance of body normalization in VR applications with higher immersion levels.