Loading...
Developing Privacy Frameworks for Motion Sensor Data in Next-Generation Wearable Devices
Ahmed, Ragad Moustafa
Ahmed, Ragad Moustafa
Date
2025-12
Author
Advisor
Type
Thesis
Degree
Citations
Altmetric:
Files
Description
A Master of Science thesis in Computer Engineering by Ragad Moustafa Ahmed entitled, “Developing Privacy Frameworks for Motion Sensor Data in Next-Generation Wearable Devices”, submitted in December 2025. Thesis advisor is Dr. Reham Aburas and thesis co-advisor is Dr. Alex Abraham Aklson. Soft copy is available (Thesis, Completion Certificate, Approval Signatures, and AUS Archives Consent Form).
Abstract
Wearable motion sensors embedded in devices such as AR glasses, smartwatches, and wristbands capture rich kinematic information that enables accurate activity recognition but simultaneously reveal sensitive demographic traits through unintended inference channels. This thesis investigates the privacy risks inherent in inertial motion data and proposes a privacy-utility preserving representation learning framework based on adversarial training. Using the Nymeria dataset (the world’s largest egocentric motion collection), the study focuses specifically on IMU-derived linear and angular velocity signals from head-mounted and wrist-mounted devices. Although Nymeria has been widely used as a benchmark for foundation models in vision and motion AI, this work presents the first systematic privacy analysis of its sensor modalities. A structured methodological pipeline was developed, including large-scale preprocessing, segmentation, handcrafted feature extraction, baseline modelling, and adversarial representation learning. Baseline results show that the 144-dimensional motion features strongly support script classification but also leak gender information, with a gender inference AUC of 0.889.The proposed Adversarial Representation Learning with Autoencoder (ARA) model suppresses this leakage while maintaining activity-recognition utility: adversarial training reduces gender AUC to 0.5614 (near random guessing) and preserves activity classification utility with an AUC of 0.9738. This thesis represents the first adaptation of an ARA-style adversarial anonymization framework to IMU sensor data from AR glasses and wristbands, demonstrating that adversarial representation learning extends effectively beyond image embeddings to wearable-sensor modalities. Overall, this work establishes an empirical foundation for understanding the privacy vulnerabilities of egocentric motion datasets and provides a practical mechanism for mitigating demographic inference risks in next-generation wearable ecosystems.