Attacks and defenses in the location privacy literature largely consider that users' data available for training wholly characterizes their mobility patterns. Thus, they hardwire this information in their models. We show that, in practice, training information cannot capture users' behavior with perfect certainty, and hence state-of-the-art defenses overestimate the level of privacy they provide. To tackle this problem, we propose a new model for user mobility in sporadic scenarios, that we call the blank-slate model. This model acknowledges the uncertainty of designers about real user behavior: it starts as a "blank page", and learns the actual user behavior as she queries the service provider. We use the blank-slate model to develop both an attack, the Profile Estimation Based Attack (PEBA), and a design technique, called output-feedback, that can be used to create LPPMs that progressively learn the user behavior. Using real data, we empirically show that our proposals outperform state-of-the-art mechanisms designed based on previous hardwired models. Our learning LPPMs do not require bootstrapping with training data and are easy to compute, so they could be easily deployed in current location privacy solutions.
View on arXiv