Abstract

In the classical multi-party computation setting, multiple parties jointlycompute a function without revealing their own input data. We consider avariant of this problem, where the input data can be shared for machinelearning training purposes, but the data are also encrypted so that they cannotbe recovered by other parties. We present a rotation based method using flowmodel, and theoretically justified its security. We demonstrate theeffectiveness of our method in different scenarios, including supervised securemodel training, and unsupervised generative model training. Our code isavailable at https://github.com/ duchenzhuang/flowencrypt.

 

Quick Read (beta)