DIFFERENTIAL PRIVACY AND HOMOMORPHIC ENCRYPTION–BASED PRIVACY-PRESERVING ENSEMBLE LEARNING FOR MEDICAL DIAGNOSIS

Machine learning in healthcare is increasingly being used for disease diagnosis, prognosis and patient care. But medical data are also sensitive, and the exposure of such personal health information by unauthorized access or model inference is a serious concern. To solve these problems, this paper introduces a privacy-preserving ensemble architecture called Differentially Privacy and Homomorphic Encryption Based Privacy-Preserving Ensemble Learning (DP-HE-PPEL) for medical data classification which consists of differentially private AdaBoost (DP-AB), differentially private random forest (DP-RF) and homomorphic encryption-based support vector machine (HE-SVM) classifiers. The proposed system employs several models to achieve better robustness and accuracy with the protection of formal privacy. The final decision is the result of a majority vote among all three (DP-AB, DP-RF, HE-SVM). As a mimicry of privacy-preserving inference, noisy Laplace-protected input features are used to emulate secure computation protocols. Experimental results show that the ensemble performances at optimal level and keeps good predictive accuracy with reasonable privacy, which follows at a minor cost for noisy data. In particular, our proposed system can trade off between privacy and robustness so that it is appropriate for integrating into real healthcare applications. Overall, this paper presents a holistic approach to privacy-preserving medical data classification that unifies differential privacy, simulated secure inference and ensemble learning.