Efficient Deep Learning on Multi-Source Private Data

TL;DR

This paper introduces Myelin, a deep learning framework that combines privacy-preserving techniques like trusted hardware and differential privacy to enable secure, collaborative model training without data exposure.

Contribution

It presents a novel framework, Myelin, that integrates multiple privacy-preserving methods for efficient and fully private deep learning.

Findings

Establishes a baseline performance for private deep learning.

Demonstrates the feasibility of combining trusted hardware and differential privacy.

Provides insights into the trade-offs between privacy and model accuracy.

Abstract

Machine learning models benefit from large and diverse datasets. Using such datasets, however, often requires trusting a centralized data aggregator. For sensitive applications like healthcare and finance this is undesirable as it could compromise patient privacy or divulge trade secrets. Recent advances in secure and privacy-preserving computation, including trusted hardware enclaves and differential privacy, offer a way for mutually distrusting parties to efficiently train a machine learning model without revealing the training data. In this work, we introduce Myelin, a deep learning framework which combines these privacy-preservation primitives, and use it to establish a baseline level of performance for fully private machine learning.