Differential Guarantees for Cryptographic Systems

Jan 11, 2017

Download: MP4 Video Size: 139.0MB  
Watch on YouTube

Abstract

Differential privacy aims at learning information about the population as a whole, while protecting the privacy of each individual. With its quantifiable privacy and utility guarantees, differential privacy is becoming standard in the field of privacy-preserving data analysis. On the other hand, most cryptographic systems for their privacy properties rely on a stronger notion of indistinguishability, where an adversary should not be able to (non-negligibly) distinguish between two scenarios. Nevertheless, there exists some cryptographic system scenarios for which the notion of indistinguishability is known to be impossible to achieve. It is natural to ask if one can define differential privacy-motivated privacy notions to accurately quantify the privacy loss in those scenarios. In this talk, we will study two such scenarios.

Our first scenario will consider (non-)uniform randomness employed in cryptographic primitives. It is well-known that indistinguishability-based definitions of cryptographic primitives are impossible to realize in systems where parties only have access to non-extractable sources of randomness. I will demonstrate that it is, nevertheless, possible to quantify this secrecy (or privacy) loss due to some non-extractable sources (such as the Santha-Vazirani sources) using a generalization of indistinguishability inspired by differential privacy.

Our second scenario will capture privacy properties of anonymous communication networks (e.g., Tor). In particular, I will present our AnoA framework that relies on a novel relaxation of differential privacy to enables a unified quantitative analysis of properties such as sender anonymity, sender unlinkability, and relationship anonymity.