Privacy-preserving inference in crowdsourcing systems

Machine learning has widely been used in crowdsourcing systems to analyze the behavior of their mobile users. However, it naturally raises privacy concerns, as personal data needs to be collected and analyzed in the cloud, and results need to be sent back to the users to improve their local estimates. In this paper, we focus on the use of a specific type of learning algorithms, called maximum a posteriori (MAP) inference, in crowdsourcing systems, and use a crowdsourced localization system as an example. With MAP inference, the accuracy of each estimate of the user state may be improved by analyzing other users' estimates. Naturally, the privacy of the user state needs to be protected. Within the general framework of differential privacy, we show how private user states can be perturbed while preserving statistically accurate results. For the crowdsourcing system, we design a non-interactive mechanism for a group of users to perform inference without revealing their true states to any other party. The mechanism is implemented and verified in an indoor localization system. By comparing with the state-of-the-art, we have shown that our proposed privacy-preserving mechanism produces highly accurate results efficiently.