Enabling Data Sharing Through Controlled Data Release

Data sharing helps agents unlock the value of data in many scenarios. Government agencies share data to perform various functionalities; organizations share data to improve tasks such as fraud detection or drug discovery; and data markets create opportunities for agents to benefit from a large pool of external data. At the same time, data sharing scenarios are highly diverse, and each faces unique sharing constraints and requirements, including and not limited to the content and format of the data shared, the purpose for which data can be shared, and who is allowed to access the output of sharing under what conditions. These idiosyncrasies make enabling data sharing on a broader scale a challenging process. At the heart of these challenges is helping agents achieve controlled data release across all stages of sharing. Before sharing happens, agents should have enough insight into the outcome of data sharing to judge whether they will allow the sharing to happen. Once they decide to share, they need a sharing mechanism and supporting infrastructure to them achieve their sharing goals without violating their sharing constraints. Even after the sharing has happened, agents still want control over the purpose for which the data can be used. Many existing mechanisms already aim to help agents achieve controlled data release at different stages of data sharing, both technical and non-technical. However, these mechanisms remain limited in important ways: they may only support a narrow class of sharing scenarios, only target one stage of sharing without considering the others, or focus on ex-post punishment rather than enforcement. As a result, many more potentially beneficial sharing scenarios remain to be unblocked. This thesis develops systems and mechanisms to help agents achieve end-to-end controlled data release. In particular, our solution allows agents to discover the value of sharing in a data sharing scenario, and to express and enforce their preferences and constraints with an efficient and general sharing mechanism. To this end, we make the following contributions (i) providing a general programming model that permits developers to write applications that can programmatically incorporate diver sharing requirements, (ii) building an efficient sharing infrastructure that is general enough to support any computation, and (iii) designing a instrumental-value oriented data market that gives agents clear signal of the value of data before they commit to sharing. Together, these contributions provide a more complete foundation for achieving controlled data release on a broader scale, and help pave the way for enabling many more sharing scenarios that are currently too costly or too complex to realize.