Policy Driven Data Sharing with Provable Privacy Guarantees

Companies such as Google or Facebook collect a substantial amount of data about their users to provide useful services. The release of these datasets for general use can enable numerous innovative applications and scientific research. However, such data contains sensitive information about users, and simple anonymization techniques have been shown to be ineffective to ensure users’ privacy. These privacy concerns have motivated the development of algorithms that share data with provable privacy guarantees including differential privacy. However, the focus of differentially private algorithm design has been on simplified problem settings. Real world applications must satisfy to complex privacy policies (beyond whether an individual is in or out of the dataset) and adhere to complex constraints, which hinders the deployment of differentially private algorithms.

This dissertation presents a novel policy-driven approach to design provable privacy guarantees for complex settings. This policy-driven approach results in a useful class of provable privacy definitions, named as Blowfish privacy, (a) generalize differential privacy to handle complex privacy preferences and constraints, (b) unify several variants of differential privacy that are used in practice, and (c) allow the creation of new well founded privacy definitions that allow flexible trade-offs between privacy, accuracy, and performance, based on the application’s requirements. The usefulness of this approach is shown in two use cases of data sharing: (1) analyzing location data which involves complex data types and privacy preferences, and (2) scaling private record linkage which involves secure computations between multiple parties. This work concludes with directions for future privacy research in private data analysis.