Saying it's Anonymous Doesn't Make It So: re-identifications of "anonymized" law school data

Results Summary: One protocol used a physical data enclave; two purported to produce a k-anonymity version of the data; and a fourth protocol developed a statistical model of the data. None of the protocols provided the privacy protection promised or commensurate with common expectations under public records laws. We demonstrated important lessons. k-anonymity guarantees that an adversary cannot do better than guessing that a name matches to at least k records, or vice versa. All fields that can be used for linking are included in the k-anonymization, which in today's networked society means all fields or requires justifying any non-included field. None of the "k-anonymity" protocols provided k-anonymity protection. Small group re-identifications can be as harmful as unique re-identifications. Physical data enclaves cannot thwart hiding or memorizing sensitive information. Adversarial testing on de-identified data can point out vulnerabilities and improve real-world practice. All four protocols left the records of Black and Hispanic test takers significantly more identifiable than the records of Whites. The Superior Court of California denied Sander's request for compelled disclosure of the data.