Differentially private auditing and monitoring

Machine learning methods and algorithms are an increasingly large component of modern data analytic practices and many businesses and institutions have a natural interest in using collected data for training these algorithms. However, while having appropriate data is important for many of these learning applications, issues can arise when this underlying data is sensitive. Data contributing individuals may not feel comfortable having their data used in for machine learning tasks. At the same time, privacy enhancing technologies (PETS) such as differential privacy may be difficult to understand for both data holders and data analysts. One example application is anomaly detection, in which an auditor could use privacy preserving learning algorithms to provide privacy protections/guarantees. This thesis describes three research contributions aimed at groups seeking to incorporate privacy protection in problems related to search and monitoring. First, we propose the concept of anomaly-restricted differential privacy and provide a privacy preserving anomaly detection algorithm. The goal is to provide potential auditors a way to perform anomaly detection while protecting the privacy of non-anomalous individuals. Second, we provide a differentially-private active learning algorithm and a web-based machine learning tool that implements the algorithm in an online stream-based environment. Analysts can use such a system to understand privacy/utility tradeoffs for anomaly detection or other learning tasks in active learning environments. Finally, we show that sensing systems proposed for “smart buildings” can reveal private information about an individual’s movements, even when the reported data are room occupancy counts. The attack strategy demonstrates some of the privacy challenges facing infrastructure-based sensing systems, where data is only indirectly collected from individuals.