CIF: Small: Information-theoretic privacy and security for personalized distributed learning

点击翻译按钮获取中文摘要

Personalized recommendation systems have been widely deployed with centralized data collection over the web and mobile applications with great success. Concurrently, they have created significant and justified public concerns about privacy. Society is in the midst of crossing a new frontier where millions of devices ranging from simpler Internet-of-Things devices (sensing) to (semi) autonomous vehicles (cars, drones) are connected over networks. A natural question is whether and how one can leverage large-scale local data collection in this emerging ecosystem to collaboratively build distributed personalized learning systems. This motivates the central question of this project, namely, how to design personalized learning models with information-theoretic privacy and security guarantees. The research outcomes of this project will be broadly disseminated, through publications, involvement of the PI in teaching and interaction with industry. One would ideally like to design personalized systems that can leverage large-scale collaboration, maintain privacy of local data, and require trust only on one's own devices as opposed to other entities. This project therefore explores how to design privacy schemes which also give good personalized learning performance, and how to design robust collaborative schemes that give good personalized learning performance despite malicious participants. In particular, in the task on privacy for personalized learning, the project plans to explore designs of personalized privacy mechanisms that are robust to iterative interactions necessary for collaborative learning and analyze its privacy-performance trade-off using information-theoretic and statistical tools. In the task on security for personalized learning, the project leverages ideas from high-dimensional robust statistics and information theory to develop robust mechanisms with theoretical guarantees to enable personalized learning in the presence of malicious participating devices, including investigating when collaboration is beneficial. The successful completion of the project will advance the state of the art and build bridges between information theory and trustworthy federated/distributed machine learning and optimization, through formal privacy and security guarantees without adversary computational assumptions.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.