Towards Practical Differentially Private Mechanism Design and Deployment

As the collection of personal data has increased, many institutions face an urgent need for reliable protection of sensitive data. Among the emerging privacy protection mechanisms, differential privacy offers a persuasive and provable assurance to individuals and has become the dominant model in the research community. However, despite growing adoption, the complexity of designing differentially private algorithms and effectively deploying them in real-world applications remains high. In this thesis, we address two main questions: 1) how can we aid programmers in developing private programs with high utility? and 2) how can we deploy differentially private algorithms to visual analytics systems? We first propose a programming framework and system EKTELO which can be used to author programs for a variety of statistical tasks that involve answering counting queries. In the framework, programs are described as compositions of reusable modules and automatically satisfy differential privacy. Moving on to the second question, we investigate the challenges of deploying differentially private algorithms in visualization tasks. Specifically, we conduct a study to better understand the relationship between noise introduced for privacy protection, visual analytics tasks, visualization, and accuracy. We also look at the influence of uncertainty in differentially private visualization and propose an approach to effectively represent uncertainty in two-dimensional location data. Third, we demonstrate how to direct deployment of differentially private algorithms causes both efficiency and accuracy issues in an interactive visualization dashboard. To address these challenges, we propose a DashGuard, a private dashboard where a smart middle layer processes front-end queries issued to the back-end EKTELO private engine. Through reuse and pre-computation of measurement, the middle layer provides benefits in accuracy, efficiency, and privacy budget consumption