Building Differentially Private Systems- Opportunities and Challenges

Abstract: Differential privacy has emerged as the gold standard for privacy preserving data analysis. Over the last decade, researchers from a variety of fields (including theory, databases, machine learning and security) have identified algorithms that ensure differential privacy and characterized asymptotic lower bounds on the error needed to solve tasks like answering counting queries and learning. In this breakout session, we will discuss the opportunities and barriers in developing systems for data analysis that ensure accuracy as well as provable guarantees of privacy. We will briefly discuss application domains where differential privacy implementations are being used or designed to release/analyze  data. Challenges we may discuss include: (a) customizing differential privacy to settings with complex data types with multiple sensitive entities (like relational data), streaming data (like location traces), and correlated data (like in social networks) -- settings where differential privacy may not directly apply; (b) authoring safe differentially private software tools that do not permit side channel attacks, (c) building algorithms with optimal error on finite datasets and the challenges with data dependent algorithm design, and (d) developing end-to-end differentially private algorithms and benchmarks for evaluating their error.