VisGuardian: A Lightweight Group-based Visual Privacy Control Technique For Smart Glasses in Home Environments

Paper Title

VisGuardian: A Lightweight Group-based Visual Privacy Control Technique For Smart Glasses in Home Environments

Publication Info

• Topic area: Privacy control techniques for AR glasses in home environments.
• Keywords: Visual privacy, AR glasses, group-based control, content-based permissions, YOLO, privacy-by-design, user study, real-time processing, home environments, AI applications.

Background and Problem

• Problem / challenge: Existing permission systems for AR glasses are inadequate for managing the continuous, context-sensitive visual data captured in home environments. Current methods either lack granularity, impose high cognitive loads, or fail to balance privacy and utility effectively.
• Significance: Home environments are highly sensitive due to intimate routines and the presence of bystanders. Robust privacy controls are essential to prevent data misuse, profiling, and breaches of personal privacy.
• Motivation and related work: Prior approaches include binary permission systems, multi-level sliders, and object-based controls, but these are either too rigid or inefficient for dynamic, object-dense scenarios. There is a gap in providing fine-grained, scalable, and user-friendly privacy controls tailored to home environments.

Solution

• Proposed approach: VisGuardian, a group-based, fine-grained visual privacy control technique for AR glasses, enabling users to manage permissions for multiple related objects efficiently.
• Novelty:
  1. Introduction of a group-based control mechanism that categorizes objects by privacy sensitivity, object type, or spatial proximity.
  2. Real-time object detection and sanitization using YOLOv10, achieving low latency and minimal battery impact.
  3. User-friendly interaction design with in-situ overlays and memory of user preferences for seamless operation.
• Procedure and key techniques:
  1. Detect sensitive objects in real-time using a fine-tuned YOLOv10 model.
  2. Allow users to select a single object and apply privacy settings across related groups.
  3. Render semi-transparent overlays to obscure sensitive objects and remember user preferences for future interactions.
  4. Evaluate the system through technical benchmarks and user studies in simulated home environments.

Results

• Concrete findings:
  • Detection precision (mAP50): 0.6704 with 14.0 ms latency and 1.7% increase in hourly battery consumption.
  • User study (N=24): VisGuardian reduced permission control time (15.20s vs. 20.36s for slider-based control) and clicks compared to baselines.
• Advantage over baselines:
  • Faster and more efficient than slider-based and object-based controls.
  • Higher user satisfaction and perceived ease of use.
• Experiments / evaluation:
  • Technical evaluation on COCO and LVIS datasets compared YOLOv10 to other models (e.g., Faster R-CNN, YOLOv8).
  • User study in a simulated home environment with 12 tasks across four scenarios (e.g., health monitoring, household management, social interaction, work assistance).
  • Metrics: permission control time, number of clicks, perceived privacy protection, usability ratings.
• Limitations and future work:
  • Static taxonomy for privacy objects; no user-defined categories.
  • Evaluation limited to controlled environments; real-world variability not tested.
  • Focus on visual data; future work could extend to other sensors (e.g., audio).

Summary

VisGuardian introduces a group-based, fine-grained visual privacy control system for AR glasses, addressing the challenges of managing sensitive data in home environments. By leveraging real-time object detection and group-based settings, it reduces user effort and enhances usability compared to existing methods. Technical evaluations confirm its efficiency and accuracy, while user studies demonstrate its effectiveness in balancing privacy and utility. Future work will explore dynamic taxonomies, real-world deployment, and extensions to other sensor modalities.