Statistics Seminar

Date: November 9, 2022

Speaker: Xiao-Li Meng

Website:

Note: Joint with Political Science’s MEAD seminar series

Title: Privacy, Data Privacy, and Differential Privacy

Abstract: This talk invites curious minds to contemplate the notion of data privacy.  It first traces the evasive concept of privacy to a legal right, derived from the frustration of the husband of a socialite attracting tabloids when yellow journalism and film photography became popular in 1890s.  More than a century later, the rise of digital technologies and data science has made the issue of data privacy a central concern for essentially all enterprises, from medical research to business applications, and to census operations. Differential privacy (DP), a theoretically elegant and methodologically impactful framework developed in cryptography, is a major milestone in dealing with the thorny issue of properly balancing data privacy and data utility.  However, the popularity of DP has brought both hype and scrutiny, revealing several misunderstandings and subtleties that have created confusions even among specialists.  The technical part of this talk is therefore devoted to explicating such issues from a statistical perspective, particularly via the prior-to-posterior semantics of DP.  This semantics yields an intuitive statistical interpretation of DP, albeit it does not correspond in general to the commonly understood and desired data privacy protection.  The determination of whether the traditional data swapping method is DP demonstrates the subtleties and their consequences when overlooked.  Ultimately, the talk aims to highlight the challenges and research opportunities in quantifying data privacy, what DP does and does not protect, and the need to properly analyze DP data.  (This talk is based on three articles with James Bailie and Ruobin Gong.)