Abstract

We present a framework for experimenting with secure multi-party computationdirectly in TensorFlow. By doing so we benefit from several properties valuableto both researchers and practitioners, including tight integration withordinary machine learning processes, existing optimizations for distributedcomputation in TensorFlow, high-level abstractions for expressing complexalgorithms and protocols, and an expanded set of familiar tooling. We give anopen source implementation of a state-of-the-art protocol and report onconcrete benchmarks using typical models from private machine learning.

 

Quick Read (beta)