No silver bullet: De-identification still doesn’t work

Paul Ohm’s 2009 article Broken Promises of Privacy spurred a debate in legal and policy circles on the appropriate response to computer science research on re-identification techniques. In this debate, the empirical research has often been misunderstood or misrepresented. A new report by Ann Cavoukian and Daniel Castro is full of such inaccuracies, despite its claims of “setting the record straight.”

In a response to this piece, Ed Felten and I point out eight of our most serious points of disagreement with Cavoukian and Castro. The thrust of our arguments is that (i) there is no evidence that de-identification works either in theory or in practice and (ii) attempts to quantify its efficacy are unscientific and promote a false sense of security by assuming unrealistic, artificially constrained models of what an adversary might do.

Specifically, we argue that:

1. There is no known effective method to anonymize location data, and no evidence that it’s meaningfully achievable.
2. Computing re-identification probabilities based on proof-of-concept demonstrations is silly.
3. Cavoukian and Castro ignore many realistic threats by focusing narrowly on a particular model of re-identification.
4. Cavoukian and Castro concede that de-identification is inadequate for high-dimensional data. But nowadays most interesting datasets are high-dimensional.
5. Penetrate-and-patch is not an option.
6. Computer science knowledge is relevant and highly available.
7. Cavoukian and Castro apply different standards to big data and re-identification techniques.
8. Quantification of re-identification probabilities, which permeates Cavoukian and Castro’s arguments, is a fundamentally meaningless exercise.

Data privacy is a hard problem. Data custodians face a choice between roughly three alternatives: sticking with the old habit of de-identification and hoping for the best; turning to emerging technologies like differential privacy that involve some trade-offs in utility and convenience; and using legal agreements to limit the flow and use of sensitive data. These solutions aren’t fully satisfactory, either individually or in combination, nor is any one approach the best in all circumstances.

Change is difficult. When faced with the challenge of fostering data science while preventing privacy risks, the urge to preserve the status quo is understandable. However, this is incompatible with the reality of re-identification science. If a “best of both worlds” solution exists, de-identification is certainly not that solution. Instead of looking for a silver bullet, policy makers must confront hard choices.