Private Client-Side Profiling with Random Forests and Hidden Markov Models

Abstract

Abstract. Nowadays, service providers gather fine-grained data about users to deliver personalized services, for example, through the use of third-party cookies or social network profiles. This poses a threat both to privacy, since the amount of information obtained is excessive for the purpose of customization, and authenticity, because those methods employed to gather data can be blocked and fooled. In this paper we propose privacy-preserving profiling techniques, in which users perform the profiling task locally, reveal to service providers the result and prove its correctness. We address how our approach applies to tasks of both classification and pattern recognition. For the former, we describe client-side profiling based on random forests, where users, based on certified input data representing their activity, resolve a random forest and reveal the classification result to service providers. For the latter, we show how to match a stream of user activity to a regular expression, or how to assign it a probability using a hidden Markov model. Our techniques, based on the use of zero-knowledge proofs, can be composed with other protocols as part of the certification of a larger computation. 1