14:00 – 15:00 – Louis Aslett (Durham)
Title: Confidential Accept-Reject
Abstract: Federated learning enables statistical models to be fitted using distributed data sets, without the need to bring those data sets to a single location. Although this holds promise for collaborative model fitting while preserving the data privacy of each participant, a careful analysis may still be required to understand the privacy implications of both federated learning outputs and summaries exchanged during fitting. We present work-in-progress, proposing a federated, privacy-preserving accept-reject mechanism which exploits modern homomorphic secret sharing methods. This mechanism enables a range of Monte Carlo algorithms involving an accept-reject step to be converted into federated equivalents, also with the potential to complement privacy properties of existing federated Monte Carlo algorithms that already incorporate an accept-reject step.
In addition to this, we provide a practical software implementation enabling live federated learning across the internet between different parties, secured by both encrypted and signed shares, to authenticate participants and prevent man-in-the-middle attacks.
Refreshments available between 15:00 – 15:30, Huxley Common Room (HXLY 549)