Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

GaitGuard: Protecting Video-Based Gait Privacy in Mixed Reality

Diana Romero, Athina Markopoulou, Salma Elmalaki

TL;DR

GaitGuard addresses the privacy risks of video-based gait in mixed reality by offloading real-time gait protection to a companion device. It introduces GaitExtract to clinically-ground gait features from egocentric MR video and an adaptive mitigation strategy that targets gait-critical frames, leveraging temporal sparsity to preserve visual quality. Through a comprehensive 233-configuration evaluation, the approach demonstrates substantial privacy gains (up to 68% profiling accuracy reduction) with high visual fidelity (SSIM ~0.97, PSNR ~30 dB) and real-time performance (29 FPS) across up to 10 users. The work offers a practical, deployable defense against gait-based profiling in collaborative MR and highlights the importance of adaptive, temporally-aware privacy techniques.

Abstract

Mixed Reality (MR) systems capture continuous video streams that expose bystanders' and collaborators' gait patterns -- a biometric revealing sensitive attributes including age, gender, and health conditions. We show that video-based gait profiling achieves 78\% accuracy (15.6$\times$ random chance) on unprotected MR feeds, motivating \textbf{GaitGuard}, a real-time defense operating on a companion mobile device. GaitGuard introduces \textbf{GaitExtract}, an automated gait feature extraction pipeline adapted from clinical analysis for egocentric MR perspectives. Through systematic evaluation of 233 mitigation configurations, we characterize privacy-utility-performance trade-offs. A key insight is that gait features derive primarily from transient events (heel strikes, toe-offs). We exploit this temporal sparsity through adaptive mitigation that selectively processes only gait-critical frames, achieving a 68\% reduction in profiling accuracy while preserving visual quality (SSIM: 0.97) at 29~FPS. \textbf{GaitGuard} scales to 10 simultaneous users with under 10ms latency. A qualitative study of 20-participants confirms that the users preferred a solution such as \textbf{GaitGuard} which provides privacy guarantees.

GaitGuard: Protecting Video-Based Gait Privacy in Mixed Reality

TL;DR

GaitGuard addresses the privacy risks of video-based gait in mixed reality by offloading real-time gait protection to a companion device. It introduces GaitExtract to clinically-ground gait features from egocentric MR video and an adaptive mitigation strategy that targets gait-critical frames, leveraging temporal sparsity to preserve visual quality. Through a comprehensive 233-configuration evaluation, the approach demonstrates substantial privacy gains (up to 68% profiling accuracy reduction) with high visual fidelity (SSIM ~0.97, PSNR ~30 dB) and real-time performance (29 FPS) across up to 10 users. The work offers a practical, deployable defense against gait-based profiling in collaborative MR and highlights the importance of adaptive, temporally-aware privacy techniques.

Abstract

Mixed Reality (MR) systems capture continuous video streams that expose bystanders' and collaborators' gait patterns -- a biometric revealing sensitive attributes including age, gender, and health conditions. We show that video-based gait profiling achieves 78\% accuracy (15.6 random chance) on unprotected MR feeds, motivating \textbf{GaitGuard}, a real-time defense operating on a companion mobile device. GaitGuard introduces \textbf{GaitExtract}, an automated gait feature extraction pipeline adapted from clinical analysis for egocentric MR perspectives. Through systematic evaluation of 233 mitigation configurations, we characterize privacy-utility-performance trade-offs. A key insight is that gait features derive primarily from transient events (heel strikes, toe-offs). We exploit this temporal sparsity through adaptive mitigation that selectively processes only gait-critical frames, achieving a 68\% reduction in profiling accuracy while preserving visual quality (SSIM: 0.97) at 29~FPS. \textbf{GaitGuard} scales to 10 simultaneous users with under 10ms latency. A qualitative study of 20-participants confirms that the users preferred a solution such as \textbf{GaitGuard} which provides privacy guarantees.
Paper Structure (37 sections, 18 figures, 7 tables)

This paper contains 37 sections, 18 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. Gait Analysis Methods
  4. Gait Privacy Beyond Identification
  5. Datasets for Gait Analysis
  6. Privacy in MR
  7. Threat Model
  8. GaitGuard System Design
  9. Architecture and Deployment Strategy
  10. GaitExtract
  11. Mitigation Design Space
  12. Adaptive Mitigation & Gait Event Dataset
  13. Implementation
  14. Evaluation
  15. Methodology
  16. ...and 22 more sections

Figures (18)

  • Figure 1: Collaborative MR on a shared virtual object: a bystander walks into the headset's view while a remote desktop user assists, exposing moving users to potential video-based gait-profiling.
  • Figure 2: Gait cycle highlighting toe-off and heel strike events pirker2017gait.
  • Figure 3: Overview of GaitGuard.[Left] Multiple users interact in a collaborative MR app. GaitGuard: ❶ intercepts raw frames from HoloLens 2; ❷ GaitGuard applies gait mitigation; ❸ releases mitigated frames to the app—protecting against adversaries on the ❹ server or ❺ remote clients. [Right]GaitGuard implemented as three pipelined threads: A captures frames, B detects humans and mitigates gait, C transmits output—maintaining real-time MR performance. An honest-but-curious adversary may reside in the telemetry manager or remote desktop client.
  • Figure 4: GaitExtract: A gait-profiling pipeline using OpenPose to extract temporal features and estimate step length with minimal calibration.
  • Figure 5: Design space of gait mitigation strategies. Mitigation techniques are explored along four aspects: ❶ location of application, ❷ mask size, ❸ type of transformation, and ❹ adaptability. Each axis presents trade-offs in privacy protection, visual quality, and computational efficiency.
  • ...and 13 more figures