A System for Privacy Preserving Collaborative Machine Learning on Sensor Data

Abstract

Sensor data is everywhere and we have both the machine learning algorithms and big data infrastructures to make sense of it. Unfortunately, people are reluctant to share their sensor data due to privacy concerns. In this paper I describe a system that collects sensor data in a privacy-preserving manner, builds machine-learning classifiers based on that data, and then lets users use those classifiers. My work is heavily inspired by a similar project called “Pickle,”which I discuss in detail. My paper has three parts. First, I describe how I ensure user privacythrough data perturbation; in my case this consists of adding additive and multiplicative noise to the feature data. Second, I describe the architecture of my system, which uses scalable third party tools. Third, I discuss topics related to my system, such as adoption incentives, data obfuscation techniques, and design of similar systems. I evaluate my work by quantifying the tradeoffs between privacy and classification accuracy incurred by data perturbation and listing the space and time constraints of my implementation. Future work should evaluate these privacy-accuracy tradeoffs in specific domains such as speech recognition.