A framework for context-aware privacy of sensor data on mobile systems

We study the competing goals of utility and privacy as they arise when a user shares personal sensor data with apps on a smartphone. On the one hand, there can be value to the user for sharing data in the form of various personalized services and recommendations; on the other hand, there is the risk of revealing behaviors to the app producers that the user would like to keep private. The current approaches to privacy, usually defined in multi-user settings, rely on anonymization to prevent such sensitive behaviors from being traced back to the user---a strategy which does not apply if user identity is already known, as is the case here. Instead of protecting identity, we focus on the more general problem of choosing what data to share, in such a way that certain kinds of inferences---i.e., those indicating the user's sensitive behavior---cannot be drawn. The use of inference functions allows us to establish a terminology to unify prior notions of privacy as special cases of this more general problem. We identify several information disclosure regimes, each corresponding to a specific privacy-utility tradeoff, as well as privacy mechanisms designed to realize these tradeoff points. Finally, we propose ipShield as a privacy-aware framework which uses current user context together with a model of user behavior to quantify an adversary's knowledge regarding a sensitive inference, and obfuscate data accordingly before sharing. We conclude by describing initial work towards realizing this framework.