Mitigating Hidden Privacy Risks in Next-Generation Wearables Using Adversarial Representation Learning
IEEE Access · 2026 · 14 · 119123-119136
DOI: 10.1109/access.2026.3713681 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Ragad Ahmed und 2 weiteren Autor:innen. Die Arbeit nennt 49 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Faculty Research through the American University of Sharjah.
Abstract (Originalsprache)
Wearable motion sensors embedded in devices such as augmented reality (AR) glasses and wrist-worn wearables capture rich kinematic signals that enable accurate activity recognition but may also expose sensitive demographic attributes through unintended inference. This paper investigates privacy risks in inertial motion data and proposes a privacy–utility-preserving representation learning framework based on adversarial training. We formulate activity (script) recognition as the utility objective and gender inference as the demographic privacy leakage. The proposed framework operates on IMU signals from accelerometer and gyroscope sensors in head-mounted glasses and wrist-mounted devices. We propose a structured experimental pipeline including data preprocessing, temporal segmentation, baseline modeling, and adversarial learning. We evaluate our approach on large-scale egocentric IMU data collected from Meta’s Project Aria devices using the Nymeria dataset. Baseline results show that 144-dimensional motion features achieve strong utility performance while leaking sensitive information, with a gender inference F1-score of 85%. To mitigate this risk, we introduce PRIMA, a Privacy-Resilient Inertial Motion Autoencoder that suppresses sensitive attributes while preserving task-relevant information. PRIMA significantly reduces gender inference performance (F1-score = 33.53%) while maintaining high activity recognition performance (F1-score = 83.71%). These results demonstrate that adversarial representation learning provides a practical foundation and a promising direction for privacy preservation in next-generation wearable and AR systems.
Förderung
- Faculty Research through the American University of Sharjah
Zitieren
APA
Ahmed, R., Aklson, A., & Aburas, R. M. (2026). Mitigating Hidden Privacy Risks in Next-Generation Wearables Using Adversarial Representation Learning. IEEE Access, 14, 119123-119136. https://doi.org/10.1109/access.2026.3713681
Vancouver
Ahmed R, Aklson A, Aburas RM. Mitigating Hidden Privacy Risks in Next-Generation Wearables Using Adversarial Representation Learning. IEEE Access. 2026;14:119123-119136. doi:10.1109/access.2026.3713681
BibTeX
@article{Ahmed2026,
author = {Ahmed, Ragad and Aklson, Alex and Aburas, Reham Mohamed},
title = {Mitigating Hidden Privacy Risks in Next-Generation Wearables Using Adversarial Representation Learning},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {119123--119136},
doi = {10.1109/access.2026.3713681}
}
EUVIMED