IMR:MM-1B: New directions in Privacy-Preserving Telemetry

点击翻译按钮获取中文摘要

Almost all modern-day devices are connected to the internet. Home appliances, smart watches, phones, cars, industrial tools, and even body weight scales are connected. For example, Samsung Smart TV reports back to Samsung every choice and every click that a consumer makes as well as the duration that a consumer watches any program. Internet browsers collect user browsing data. Internet Service Providers collect user IP access data. The torrent of data collected is typically used to improve user experience, service, and target advertisements. However, collection of ever more intrusive data regarding each individual consumer or organization comes at the price of tremulous invasion of privacy, and increases the risk that the data collected can be harvested for other (unintended) purposes, such as counter-intelligence, political campaigns, as well as identity theft and other criminal activity. How do we allow organizations to collect aggregate statistics regarding streaming data without violating individual consumer privacy?This research aims to explore novel ways to compute aggregate statistics on streaming data in a privacy-preserving way, extending systems such as PRIO, PRIO+, and Poplar. The framework is that users or devices send their data in a secret-shared way to two servers which then communicate with each other to compute telemetry data while not revealing (to each other or anyone else) users’ individual data. This approach was adopted, for example, by the Firefox browser in a Mozilla project titled “Origin Telemetry”. The goals of this research are to explore even more efficient methods to privately compute telemetry data in this setting by exploring how to generalize streaming algorithms (without privacy) that were pioneered by Alon, Matias, and Szegedy to streaming algorithms with privacy. More specifically, can we compute frequency moments in a privacy-preserving and efficient manner by two servers receiving a (secret-shared) stream? While this seems like a very specialized question, it is, in fact, generalizable, as shown by the PI in the paper titled “Zero-One Frequency Laws”. This brings us to the even more interesting question: how to classify all functions that can be privately computed over streaming secret-shared data in one pass and with poly-logarithmic memory. Our goal is to provide new tools for the questions of privacy-preserving analysis of streaming large volume data, specifically called out in section MM-1B of the NSF Internet Measurement Research call for proposals. If successful, our methods will allow large-scale and efficient computation of all aggregate statistics that can be computed in small memory in a privacy-preserving way. This project will also advance the state-of-the-art performance of MPC on streaming data to Internet-size streaming computations. Lastly, our proposal calls for robust training of graduate and undergraduate students, including actively seeking minorities and female students to enter cryptographic research.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.