1 of 1

  • We can achieve high activity inference accuracies with AI techniques (SVM and LLM)
  • Our LLM test accuracy results could be improved with further fine-tuning the expert knowledge and the prompts for activity inference

Conclusion and Future Work

  • Comparing accuracies between SVM and LLM
    • SVM test accuracy achieves 99.33%
    • LLM achieves an accuracy of 90.6% which is close to the SVM result (no training data)
  • Used Android Studio to develop an application to extract head-mounted display (HMD) and controller IMU data from Meta Quest
  • Designed six motions (two head motions and four hand motions) for data collection
  • Preprocessed and denoised data in MATLAB to create accurate waveforms and 3D graphs

Experiment Setup

  • Machine learning methods, such as Support Vector Machine (SVM), can learn to classify human activities by constructing support vectors to differentiate features of different classes
  • Large Language Models (LLMs) have strong generalization capability in reasoning and inferring private activity information from motion sensor readings, without training efforts from the attacker

Decoding Privacy with AI

  • Augmented Reality/Virtual Reality (AR/VR) technologies have been rapidly gaining popularity in recent years
  • Motion sensor data encodes various types of the user's private information, such as activity information and preferences
  • This project studies sensor data management in commercial AR/VR headsets and analyze the potential of private information leakage

References

[1] Xu, H., Han, L., Yang, Q., Li, M. and Srivastava, M., 2024, February. Penetrative AI: Making LLMs Comprehend the Physical World. In Proceedings of the 25th International Workshop on Mobile Computing Systems and Applications (pp. 1-7)

[2] Brownlee, J., 2018. 1D Convolutional Neural Network Models for Human Activity Recognition. Mach. Learn. Mastery, 26, p.2021.

AcknowledgementWe would like to thank our project advisor and mentors for their support and guidance throughout this project.

Privacy Leakage Study and Protection for Virtual Reality Devices

2

3

Team: Dirk Catpo Risco (MS), Brody Vallier (UG), Emily Yao (HS)

Project Advisor: Prof. Yingying (Jennifer) Chen

PhD Students as Mentors: Changming Li, Honglu Li, Tianfang Zhang

\

  • User movements can be captured via built-in motion sensors (i.e., accelerometer and gyroscope) embedded in AR/VR headset and controllers
  • Motion sensor recordings capture human motions in terms of linear acceleration and angles in a three-dimensional space

Background

Motivation and Objectives

  • Utilizing SVM upon statistical features (e.g., mean, max, min, etc.) from motion sensor data to classify human activities
  • Designing LLM prompts based on identified effective statistical features
    • Explaining the goal of the task and data types to be received
    • Expert knowledge about how to utilize the effective statistical features

    • Providing a response structure for results

Methodology

Scan for

Project Website

LLM

ML

Large Training Efforts

No Training

Small Groups of Data

Large Groups of Data

Activity Recognition

Activity Matched

Collect Motion Data

Input into SVM

Compare Results

Preprocess Motion Data

Construct LLM Prompt

Input into LLM

Right Controller

Left Controller

HMD

SVM Confusion Matrix

LLM Confusion Matrix

User Movement

Experiment Results