Abstract

Today the method du jour for statistical analysis of user behavior is to gather lots of user data, anonymize it (more-or-less), and then analyze that data. The need for doing statistical analysis drives many companies to gather large amounts of user data, often without the users’ awareness. My research group at MPI -SWS has been exploring approaches for doing statistical analysis without gathering user data. Rather, user data is kept on user devices, and queries are pushed to these devices. The resulting answers are anonymized and fuzzed such that 1) no single party can associate data with individual users, and 2) the aggregate answers are differentially private. In this talk, I will present a general approach that we will present in NSDI this year. I will outline the shortcomings of this approach, and follow with some enhancements that scale better in specific applications domains, namely web analytics and behavioral advertising.

Bio: Paul Francis is a tenured faculty at the Max Planck Institute for Software Systems in Germany. Paul has held research positions at Cornell University, ACIRI , NTT Software Labs, Bellcore,and MITRE , and was Chief Scientist at two Silicon Valley startups. Paul’s research centers around routing and addressing problems in the Internet and P2P networks. Paul’s innovations include NAT , shared-tree multicast, the first P2P multicast system, the first DHT (as part of landmark routing), and Virtual Aggregation. Recently Paul has become interested in designing advertising systems that protect user privacy while allowing for effective targeting.