Abstract: Local Differential Privacy (LDP) is an approach that allows
a central server to compute on data submitted by multiple
users while maintaining the privacy of each user (2). LDP
is a very efficient approach to security; however, as privacy
increases, the accuracy of these computations decreases.
Multi-Party Computation (MPC) is a process by which multiple parties work together to compute the output of a function without revealing their own information (11). MPC is highly secure and accurate for such computations, but it is very computationally expensive and slow. The proposedhybrid privacy model harnesses the benefits of both LDP and MPC to create a secure, accurate, and fast algorithm for machine learning.
Authors: Yavor Litchev, Abigail Thomas (MIT)