MIRAGE: Full-Body Bystander Privacy for Smart Glasses with Consent-Based Restoration

A three-tier architecture that enforces full-body privacy, supports synthetic replacement, and enables consent-based restoration via encrypted recovery material.

Muhammad Umair; Muhammad Danial Maqbool; Fatima Arshad Cheema; Kapal Dev; Muhammad Hamad Alizai; Muhammad Ali Siddiqi; Naveed Anwar Bhatti· 2026· DOI 10.48550/arXiv.2609.24537

The core problem

Smart glasses with continuous video recording capabilities pose a significant privacy threat to bystanders. While conventional sanitization methods focus on face de-identification, they overlook the rich biometric information contained in full-body appearance, such as gait, posture, and silhouette. These cues enable person re-identification (ReID) even after face sanitization, as demonstrated by prior work. MIRAGE addresses this gap by enforcing full-body privacy for all visible bystanders. The system not only masks complete bodies but also supports synthetic full-body replacement to maintain visual plausibility. Crucially, it retains encrypted recovery material that allows authorized restoration of the original video upon consent. This digest presents the architecture, methodology, and results of MIRAGE, highlighting its effectiveness against both standard and adaptive adversaries.

Innovation

MIRAGE was evaluated on a Raspberry Pi 5 (CPU-only proxy for smart-glasses compute), companion phones, and a cloud generative backend. The system achieved 0.948 Average Precision (AP) and 0.976 Average Recall (AR) for full-body detection, accurately identifying the complete visible body. For privacy protection, bounding box masking reduced learned silhouette-based ReID to essentially random guessing, with 10.86% Rank-1 accuracy compared to an 11.12% measured chance level. Against an adaptive adversary retrained on MIRAGE's sanitized pose signals, Rank-1 gait identification dropped from 90.25% to 26.20%, removing 72.5% of the adversary's identification advantage. These results demonstrate that MIRAGE effectively mitigates full-body biometric re-identification while maintaining high detection performance. The system's ability to generate synthetic replacements and support consent-based restoration was also validated, showing that encrypted recovery material can be securely stored and later used to restore original video when authorized.

Smart glasses with continuous video recording capabilities pose a significant privacy threat to bystanders. While conventional sanitization methods focus on face de-identification, they overlook the rich biometric information contained in full-body appearance, such as gait, posture, and silhouette. These cues enable person re-identification (ReID) even after face sanitization, as demonstrated by prior work. MIRAGE addresses this gap by enforcing full-body privacy for all visible bystanders. The system not only masks complete bodies but also supports synthetic full-body replacement to maintain visual plausibility. Crucially, it retains encrypted recovery material that allows authorized restoration of the original video upon consent. This digest presents the architecture, methodology, and results of MIRAGE, highlighting its effectiveness against both standard and adaptive adversaries.

MIRAGE employs a three-tier architecture: (1) on-device processing on smart glasses (proxied by a Raspberry Pi 5), (2) companion phone for intermediate processing and storage, and (3) cloud generative backend for synthetic replacement and encrypted recovery. The on-device tier performs real-time full-body detection and bounding box masking. The companion phone manages encrypted recovery material and coordinates with the cloud. The cloud backend generates synthetic full-body replacements that preserve scene context while removing biometric identifiers. The system uses a generative model to synthesize plausible body replacements, and encryption ensures that recovery material is only accessible with proper consent. The architecture is designed to be computationally feasible on resource-constrained smart glasses, leveraging the companion phone and cloud for heavier tasks. The pipeline is illustrated in the Mermaid diagram below.

Why it matters

The results indicate that MIRAGE significantly enhances bystander privacy in smart glasses recordings by addressing full-body biometrics, a often-overlooked vector. The near-chance performance of silhouette-based ReID after masking confirms that simple bounding box masking is sufficient to disrupt learned silhouette features. The substantial reduction in gait identification accuracy against an adaptive adversary highlights the robustness of the approach even when the adversary is aware of the sanitization method. The three-tier architecture balances computational load, enabling real-time operation on constrained devices. However, the reliance on a cloud backend for synthetic replacement and encryption may raise concerns about latency and trust. Future work could explore on-device generative models and decentralized consent mechanisms. The encrypted recovery material ensures that restoration is only possible with proper authorization, aligning with privacy regulations. Overall, MIRAGE provides a practical framework for full-body privacy in wearable cameras, with potential extensions to other domains such as AR/VR and surveillance.

Who should read this

CS practitioners and researchers

Opening member content…