CAREER: Foundations of Privacy-Preserving Collaborative Learning

点击翻译按钮获取中文摘要

Collaborative machine-learning techniques allow multiple data owners to collaborate to train better machine-learning models by increasing the volume and diversity of data. In many real-world scenarios, however, the data is privacy-sensitive, as is the case for healthcare records, financial transactions, or geolocation data. Privacy-preserving machine-learning techniques can facilitate machine-learning applications while protecting the privacy of sensitive data. This project aims to develop an efficient, secure, and trustworthy collaborative learning paradigm to address several critical challenges in the real-world application of privacy-preserving collaborative learning. The outcomes of the project will allow multiple data owners to collaborate to train machine-learning models without revealing any sensitive data, which will improve the performance of machine-learning applications by increasing the volume and diversity of data. It will also facilitate novel applications in fields where data is scarce and collaboration has traditionally been limited due to privacy challenges, such as better drug and vaccine discovery in healthcare. The research will be strongly integrated with education, through mentoring of undergraduate students, development of new undergraduate and graduate courses, and machine-learning workshops for K-12 students and teachers, with the goal of building a diverse and inclusive machine learning workforce. Privacy-preserving machine learning is expected to revolutionize the future of data-driven collaborative applications, by allowing large-scale machine-learning applications without revealing any sensitive data, but its real-world adoption has been limited by several major barriers, including the communication bottleneck, security, and trustworthiness. The research will address these fundamental challenges by introducing a new approach rooted in information and coding theory. The research is organized in three main thrusts: 1) develop the foundations of communication-efficient privacy-preserving collaborative learning; 2) realize a privacy-preserving machine-learning paradigm with provable security and fairness guarantees; and 3) enable privacy-preserving machine learning in arbitrary network topologies, including centralized, decentralized, and dynamic topologies, and networks with heterogeneous computing and communication resources. The research is rooted in coding and information theory, and incorporates stochastic optimization, distributed computing, and cryptography. The insights gained from the research will enable privacy-aware machine learning applications that are: 1) accessible by users with bandwidth and computational limitations, such as consumer devices in mobile edge networks; 2) secure, by preventing adversaries from injecting unwanted behavior into the decision process; and 3) fair in its decisions towards all communities in society, without revealing any sensitive data and personal information.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.