Replacement AutoEncoder: A Privacy-Preserving Algorithm for Sensory Data Analysis

摘要

An increasing number of sensors on mobile, Internet of things (IoT), andwearable devices generate time-series measurements of physical activities.Though access to the sensory data is critical to the success of many beneficialapplications such as health monitoring or activity recognition, a wide range ofpotentially sensitive information about the individuals can also be discoveredthrough these datasets and this cannot easily be protected using traditionalprivacy approaches. In this paper, we propose an integrated sensing framework for managing accessto personal time-series data in order to provide utility while protectingindividuals' privacy. We introduce \textit{Replacement AutoEncoder}, a novelfeature-learning algorithm which learns how to transform discriminativefeatures of multidimensional time-series that correspond to sensitiveinferences, into some features that have been more observed in non-sensitiveinferences, to protect users' privacy. The main advantage of ReplacementAutoEncoder is its ability to keep important features of desired inferencesunchanged to preserve the utility of the data. We evaluate the efficacy of thealgorithm with an activity recognition task in a multi-sensing environmentusing extensive experiments on three benchmark datasets. We show that it canretain the recognition accuracy of state-of-the-art techniques whilesimultaneously preserving the privacy of sensitive information. We use aGenerative Adversarial Network to attempt to detect the replacement ofsensitive data with fake non-sensitive data. We show that this approach doesnot detect the replacement unless the network can train using the users'original unmodified data.