Clients send private inputs to a set of servers that jointly compute a function without seeing the data in the clear.
Private collection.
Servers holding secret-shared client data train models without seeing the training data.
Private processing.
Clients query a set of servers without revealing which records they retrieve.
Private retrieval.

Many of the most exciting developments in technology are premised on clever use of massive datasets. At the same time, today's practices for collecting, processing, and sharing data bring profound risks — the theft of individuals' medical records, businesses' financial data, and governments' secrets. The goal of this project is to develop new, lightweight privacy-protecting tools that let individuals, businesses, and governments reap the benefits of large datasets while providing strong, cryptographic privacy guarantees for user data.

Our systems operate in a setting with a small number of infrastructure servers and many millions of clients. The servers gather, process, and share user data while revealing as little as possible — to themselves and to system users — about any individual's data, and privacy holds as long as at least one server is honest.

Publications

  1. Heli: Heavy-Light Private Aggregation Ryan Lehmkuhl, Henry Corrigan-Gibbs, Emma Dauterman, and David J. Wu USENIX Security Symposium, 2026 · Code
  2. Nudge: A Private Recommendations Engine Alexandra Henzinger, Emma Dauterman, Henry Corrigan-Gibbs, and Dan Boneh USENIX Security Symposium, 2026 · Code