Speaker: Matt Kusner (Alan Turing Institute)

Title: Secure Computation & Machine Learning: recent developments and next steps

Abstract: While there have been massive practical developments in both cryptography and machine learning, they have largely been independent of each other, especially in their most practical aspects. Even the most recent works on learning and prediction on encrypted data are largely based on tricks that apply to either the machine learning model or the employed cryptographic techniques. This is clearly suboptimal. For machine learning to thrive in cryptographic settings we need to start from first-principles, simultaneously characterize the constraints imposed by privacy techniques and learning models, and design new protocols and models that take both into account.
In this talk we will present recent works (ICML'18) where we developed cryptographic solutions to the problems of private prediction and training, with applications to fairness and outsourced prediction. We will discuss how these works themselves are instances of the situation discussed above, and some potential ways of closing the gap between cryptography and ML in practice going forward.