基於同態加密和混淆電路之遠程醫療心電圖信號的隱私保護診斷

Abstract

憑藉模型準確度與雲端運算能力的發展，產業界逐漸有了機器學習服務的出現。然而隨著使用者對隱私保護日漸重視，使得使用醫療資料這類高度隱私相關的服務變得難以推廣。我們將利用同態加密與混淆電路並基於現存的架構與工具實作出具有隱私保護能力的心電圖診斷系統。除了能對加密後的心電圖進行診斷，該系統還能同時保護服務供應商的診斷模型免於外流，並且能提供匹敵甚至是優於現存遠端心電圖診斷服務的產能。本文將從介紹使用到的工具與架構開始並說明實作時所面對的各種困難，最後比較各式改進效能的方法以及提出實務上可行的方案作為參考。

Machine learning as a Service (MLaaS) is a new industry trend driven by powerful machine learning models and cloud computing resources. However, with the rise of awareness of personal privacy issues, these services become challenging to promote, especially when the data involved are highly sensitive, such as medical or financial data. In this work, we implement a privacy-preserving ECG inference system with the aid of the homomorphic encryption scheme and garbled circuits based on some existing analytic frameworks and libraries. Our system could perform inference on encrypted ECG signals and avoid the service provider’s model from leakage. Moreover, it provides a higher capacity than real-world ECG interpretation services. First, we will introduce the cryptographic schemes and related tools we referenced, then address a series of issues we confronted while implementing our system. At last, we demonstrate the effectiveness of the various mechanisms we applied to our system that helped us overcome those issues and obtain a feasible performance for real-world application.