Optimizing Local Computation in Secure Matrix Multiplication for Outsourced Neural Networks

Data and computation are increasingly being out-sourced to the cloud for a variety of applications. However, traditional cloud computing comes with privacy and security risks, as it requires clients to trust the cloud with sensitive data and computations. Secure multi-party computation (MPC) is a type of cryptography that allows a set of parties to jointly compute functions on sensitive data without learning any information about the input data. Thus, MPC enables secure computing across untrusted cloud service providers.Secure machine learning (e.g., neural network training and inference) is of significant interest. However, practical performance of machine learning under MPC can often be a challenge. In this work, we investigate the performance of several approaches to secure matrix multiplication and convolution - two important building blocks of neural networks. We implement and evaluate optimizations using PyTorch on top of the MP-SPDZ MPC framework, and then demonstrate them in a multi-cloud setting for two machine learning applications: a small image classification task for benchmarking purposes, and a large object detection task to demonstrate feasibility of a complex application.