Files in this item



application/pdfJUEN-DISSERTATION-2015.pdf (6MB)
(no description provided)PDF


Title:Maintaining privacy during continuous motion sensing
Author(s):Juen, Joshua Paul Joseph
Director of Research:Borisov, Nikita
Doctoral Committee Chair(s):Borisov, Nikita
Doctoral Committee Member(s):Schatz, Bruce; Caesar, Matthew; Choudhury, Romit
Department / Program:Electrical & Computer Engineering
Discipline:Electrical & Computer Engineering
Degree Granting Institution:University of Illinois at Urbana-Champaign
Subject(s):Computer Security
Machine Learning
Big Data
Phone Sensors
Differential Privacy
Abstract:Mobile devices contain sensors which allow continuous recording of a user's motion allowing the development of activity, fitness and health applications. With varied applications, the motion sensors present new privacy problems which require protection. This dissertation builds on previous work with activity and fitness machine learning techniques demonstrating the ability to predict medical values from motion data using smartphones. We conduct two clinical trials collecting a data set of eighty-eight patients and forty-five hours of monitoring to analyze the privacy implications of releasing motion data. We extract a comprehensive set of statistical features from all available smartphone sensors and evaluate feature selection techniques and machine learning models. We find we can predict user identity, phone identity, speed, FEV1/FVC, and activity from the motion signal. Designing a privacy protection mechanism for motion data requires a precise understanding of how the signal predicts the sensitive information. We develop algorithms to conduct private feature selection which identifies features useful for prediction. We find that simply blocking all private features significantly reduces the usefulness of the signal for other predictions. We develop a sensitivity estimation framework to calibrate the noise for each private feature requiring an order of magnitude less noise than differential privacy sensitivity. We find adding noise to private features calibrated using the sensitivity estimate is effective at reducing the prediction of five tested target predictions. Our methods hide both user and phone identification while allowing other prediction but cannot hide activity, FEV1/FVC and speed without significantly lowering the accuracy of other predictions. Our methods are still effective when the attacker has prior knowledge of the noise distribution. The methods presented in this dissertation demonstrate the need for privacy in motion data and provide a framework for protecting sensitive user information in motion readings.
Issue Date:2015-08-18
Rights Information:2015 Joshua Paul-Joseph Juen
Date Available in IDEALS:2016-03-02
Date Deposited:2015-12

This item appears in the following Collection(s)

Item Statistics