A Study of Differentially Private Machine Learning in Healthcare

The field of Machine Learning (ML) has been engaged in intensive research for a while to build an efficient and effective intelligent system for the early identification of chronic diseases such as cancer and diabetes and has recently seen some promising findings. The bulk of the initiatives are aimed at classifying illness onset and minimizing cases of maltreatment. As a supervised learning problem, its accuracy is mostly determined by the training data, which is labeled data on actual patients that is highly privacy-sensitive. Privacy leakage can occur at any point in the machine learning lifecycle, from model training through model deployment, and can lead to a membership inference attack, model inversion attack, and reconstruction attack. As a result, safeguarding users' privacy is critical in healthcare issues, but little has been done to satisfy this demand. In this paper, we propose differential privacy-based Logistic Regression and Naive Bayes models on breast cancer classification and diabetes prediction. We evaluate the two models using the popular Wisconsin Diagnostic Breast Cancer (WDBC) dataset, and Pima Indians Diabetes dataset and depict the privacy requirement and model accuracy trade-off.