EECS Seminar: Can Sharing Anonymous Wrist-worn Accelerometry Data Re-identify You?

Abstract: This talk will present recent work in mHealth privacy showing that sharing wrist-worn accelerometry data collected from daily life carries a re-identification risk of 96%. Re-identification is achieved using a novel deep learning model called WristPrint that recognizes unique fingerprints in the micro-motion of each person’s wrist and the sequence of their movements. A new loss function helps the model achieve open-set characteristics, to easily generalize across a large population, making the attack more potent.

Although this research impacts users who usually collect and openly share raw accelerometry data in the context of research studies, most commercial devices do not share raw sensor data. They track step counts and activity states. Whether and how much re-identification risk is embedded in such data needs further investigation. Further research is also needed to discover effective mitigation approaches.

Bio: Santosh Kumar is currently the Lillian & Morrie Moss Professor of Computer Science at University of Memphis and director of NIH-funded mHealth research centers called MD2K and mDOT. His research develops wearable AI to enable the development, optimization and privacy-aware deployment of sensor-triggered health interventions. Open-source software developed by his team has been used to conduct scientific studies nationwide, producing hundreds of terabytes of wearable sensor data. His team has used these data to develop AI models for detecting stress, smoking, craving, cocaine use, brushing and flossing from wearables.