Multi-Party Privacy-Preserving Machine Learning and Its Applications

Conducting machine learning and statistical analytics in a distributed manner while maintaining data privacy can be beneficial to a wide variety of scientific investigations involving human subject data. As there is usually a tension between privacy protection and aspiration for high quality data analytics, developing and adapting machine learning algorithms under a rigorous and customizable framework for privacy is highly urgent and desirable. Moreover, the advance of personal mobile devices and modern communication technology has foster a rapid growth of distributed collection and storage of data. How to perform distributed data analysis tasks (classifier learning, hypothesis testing, etc.) without access to raw personal data becomes a challenging yet intriguing problem. The two research areas involved in addressing this problem, multi-party machine learning and privacy-preserving techniques, are separately well established. However, an interdisciplinary integration of the research efforts from both areas has been lacking until recent years. In this dissertation, our primary goal is to bridge the gap by designing different privacy-preserving machine learning models (logistic regression, feed-forward neural network and transfer learning) using different privacy protection techniques (differential privacy and traditional cryptographic techniques). The model performance, especially the utility-privacy trade-off is further evaluated on data sets from a variety of domains. Our work provides new perspectives and solutions to current privacy concerns, and hopefully directions for future research.