SaTC: CORE: Small: Robust and Private Federated Analytics on Networked Data

University Of California-San Diego, La Jolla CA

Investigators

Abstract

Federated learning with local differential privacy -- where raw data stays on device, and only sanitized updates are sent to a central server -- has enabled a number of AI applications on sensitive data while still preserving user privacy. The goal of this project is to advance the theory and practice of federated learning beyond "a single user contributing a record" to a broader networked setting where multiple users are connected into a social network. The main challenge in deploying federated analytics in this setting is that there are multiple criteria that need to balanced together with the privacy-accuracy tradeoff. These are communication efficiency -- as the client nodes are typically low-bandwidth, robustness -- as the distributed and networked setting leaves the door open to adversaries, and more complex privacy leaks, which could arise as a result of the networked setting. The goal of the project is to address these challenges by combining ideas from statistics, privacy-preserving algorithms as well as graph algorithms and ultimately developing a broad and general suite of algorithms for private and robust federated analytics in networked settings. The project team will also engage in community-building activities by organizing privacy workshops, as well as outreach activities in high schools. 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.