Abstract

Large-scale distributed collection of contextual information is often essential in order to gather statistics and train machine learning models. The ability to do so in a privacy-preserving way enables a number of computational scenarios that would be hard, or outright impossible, to realize without strong security guarantees. In this talk, we present the design and deployment of practical techniques for privately gathering statistics from large data streams. We build on efficient cryptographic protocols for private aggregation and on data structures for succinct data representation, namely, Count-Min Sketch and Count Sketch. We then show how to use these techniques to instantiate real-world privacy-friendly systems, supporting, among others, recommendations for media streaming services and crowd-sourced mobility analytics.

We then focus on how to identify and quantify possible privacy leakage from the aggregate statistics. We frame the problem in terms of the advantage an adversary has, from the aggregates, in profiling or inferring membership target users, and present two novel frameworks to quantify such leakage vis-a-vis inference attacks even when differential privacy protections are used.