Customizing differential privacy to meet legal interpretations of privacy and deliver accurate data
Read the research co-authored by Tumult Labs co-founder Ashwin Machanavajjhala on how to formalize complex privacy requirements mandated by law using novel notions and algorithms based on differential privacy.
Summary:
National statistical agencies (like those that work for the U.S. Census Bureau) need to be able to publish summaries based on combined employer-employee data, where the privacy of both employees (individuals) and employers (establishments) is mandated by law. Privacy requirements involving multiple entities are not captured by standard differential privacy definitions and algorithms. In this work we identified legal statutes and their current interpretations that regulate the publication of these data, and designed custom privacy notions and algorithms ensuring these custom privacy notions. Our novel algorithms ensured the appropriate privacy required by law and were able to release summaries of the data with error comparable or even better than that of releases made using legacy statistical disclosure limitation techniques.