SecureML: A System for Scalable Privacy-Preserving Machine Learning

TLDR

Machine learning is widely used for predictive modeling, and its accuracy improves with large, diverse datasets, but collecting such data raises privacy concerns. This work proposes efficient protocols for privacy‑preserving machine learning in linear regression, logistic regression, and neural network training via stochastic gradient descent. The protocols operate in a two‑server, non‑colluding model using secure two‑party computation, introduce secure arithmetic on shared decimals, offer MPC‑friendly sigmoid and softmax alternatives, and are implemented in C++. Experiments show the protocols are several orders of magnitude faster than prior state‑of‑the‑art methods for privacy‑preserving linear and logistic regression, scale to millions of samples with thousands of features, and constitute the first privacy‑preserving system for neural‑network training.

Abstract

Machine learning is widely used in practice to produce predictive models for applications such as image processing, speech and text recognition. These models are more accurate when trained on large amount of data collected from different sources. However, the massive data collection raises privacy concerns. In this paper, we present new and efficient protocols for privacy preserving machine learning for linear regression, logistic regression and neural network training using the stochastic gradient descent method. Our protocols fall in the two-server model where data owners distribute their private data among two non-colluding servers who train various models on the joint data using secure two-party computation (2PC). We develop new techniques to support secure arithmetic operations on shared decimal numbers, and propose MPC-friendly alternatives to non-linear functions such as sigmoid and softmax that are superior to prior work. We implement our system in C++. Our experiments validate that our protocols are several orders of magnitude faster than the state of the art implementations for privacy preserving linear and logistic regressions, and scale to millions of data samples with thousands of features. We also implement the first privacy preserving system for training neural networks.