A Differentially Private Classification Algorithm with High Utility for Wireless Body Area Networks

Abstract

The advancement of the wireless body area networks (WBAN) and sensor technologies allows us to collect a variety of physiological and behavioral data from human body. And appropriate application of machine learning methods can greatly promote the development of e-health. Nevertheless, the collected data contains personal privacy information. When using the machine learning methods to analyze the collected data, some information of the training data will be stored in the learning models unconsciously. To handle such information disclosure problem, we propose a differentially private classification algorithm based on ensemble decision tree with high utility for wireless body area networks. In order to improve the accuracy and stableness of classification, the bagging framework of ensemble learning is used in our algorithm. We aggregate the results of multiple private decision trees as the final classification in a weight-based voting way. For each private decision tree trained on the bootstrap samples, we offer a novel privacy budget allocation strategy that allows the nodes in larger depth to get more privacy budget, which can mitigate the problem of excessive noise introduced to leaf nodes to some extent. The better classification accuracy and stableness of this new algorithm, especially on small dataset, are demonstrated by simulation experiments.