Acknowledgement: Daniel Barth-Jones kindly engaged with me after this blog was initially published, and pointed out several significant factual errors, for which I am grateful.
In 2014, the New York Taxi & Limousine Company (TLC) released a large “anonymised” dataset containing 173 million taxi rides taken in 2013. Soon after, software developer Vijay Pandurangan managed to undo the hashed taxi registration numbers. Subsequently, privacy researcher Anthony Tockar went on to combine public photos of celebrities getting in or out of cabs, to recreate their trips. See Anna Johnston’s analysis here.
This re-identification demonstration has been used by some to bolster a general claim that anonymity online is increasingly impossible.
On the other hand, medical research advocates like Columbia University epidemiologist Daniel Barth-Jones argue that the practice of de-identification can be robust and should not be dismissed as impractical on the basis of demonstrations such as this. The identifiability of celebrities in these sorts of datasets is a statistical anomaly reasons Barth-Jones and should not be used to frighten regular people out of participating in medical research on anonymised data. He wrote in a blog that:
As a health researcher, Barth-Jones is understandably worried that re-identification of small proportions of special cases is being used to exaggerate the risks to ordinary people. He says that the HIPAA de-identification protocols if properly applied leave no significant risk of re-id. But even if that’s the case, HIPAA processes are not applied to data across the board. The TLC data was described as “de-identified” and the fact that any people at all (even stand-out celebrities) could be re-identified from data does create a broad basis for concern – “de-identified” is not what it seems. Barth-Jones stresses that in the TLC case, the de-identification was fatally flawed [technically: it’s no use hashing data like registration numbers with limited value ranges because the hashed values can be reversed by brute force] but my point is this: who among us who can tell the difference between poorly de-identified and “properly” de-identified?
And how long can “properly de-identified” last? What does it mean to say casually that only a “minuscule proportion” of data can be re-identified? In this case, the re-identification of celebrities was helped by the fact lots of photos of them are readily available on social media, yet there are so many photos in the public domain now, regular people are going to get easier to be identified.
But my purpose here is not to play what-if games, and I know Daniel advocates statistically rigorous measures of identifiability. We agree on that — in fact, over the years, we have agreed on most things. The point I am trying to make in this blog post is that, just as nobody should exaggerate the risk of re-identification, nor should anyone play it down. Claims of de-identification are made almost daily for really crucial datasets, like compulsorily retained metadata, public health data, biometric templates, social media activity used for advertising, and web searches. Some of these claims are made with statistical rigor, using formal standards like the HIPAA protocols; but other times the claim is casual, made with no qualification, with the aim of comforting end users.
“De-identified” is a helluva promise to make, with far-reaching ramifications. Daniel says de-identification researchers use the term with caution, knowing there are technical qualifications around the finite probability of individuals remaining identifiable. But my position is that the fine print doesn’t translate to the general public who only hear that a database is “anonymous”. So I am afraid the term “de-identified” is meaningless outside academia, and in casual use is misleading.
Barth-Jones objects to the conclusion that “it’s virtually impossible to anonymise large data sets” but in an absolute sense, that claim is surely true. If any proportion of people in a dataset may be identified, then that data set is plainly not “anonymous”. Moreover, as statistics and mathematical techniques (like facial recognition) improve, and as more ancillary datasets (like social media photos) become accessible, the proportion of individuals who may be re-identified will keep going up.
Both sides of this vexed debate need more nuance. Privacy advocates have no wish to quell medical research per se, nor do they call for absolute privacy guarantees, but we do seek full disclosure of the risks, so that the cost-benefit equation is understood by all. One of the obvious lessons in all this is that “anonymous” or “de-identified” on their own are poor descriptions. We need tools that meaningfully describe the probability of re-identification. If statisticians and medical researchers take “de-identified” to mean “there is an acceptably small probability, namely X percent, of identification” then let’s have that fine print. Absent the detail, lay people can be forgiven for thinking re-identification isn’t going to happen. Period.
And we need policy and regulatory mechanisms to curb inappropriate re-identification. Anonymity is a brittle, essentially temporary, and inadequate privacy tool.
I argue that the act of re-identification ought to be treated as an act of Algorithmic Collection of PII, and regulated as just another type of collection, albeit an indirect one. If a statistical process results in a person’s name being added to a hitherto anonymous record in a database, it is as if the data custodian went to a third party and asked them “do you know the name of the person this record is about?”. The fact that the data custodian was clever enough to avoid having to ask anyone about the identity of people in the re-identified dataset does not alter the privacy responsibilities arising. If the effect of an action is to convert anonymous data into personally identifiable information (PII), then that action collects PII. And in most places around the world, any collection of PII automatically falls under privacy regulations.
It looks like we will never guarantee anonymity, but the good news is that for privacy, we don’t actually need to. Privacy is the protection you need when you affairs are not anonymous, for privacy is a regulated state where organisations that have knowledge about you are restrained in what they do with it. Equally, the ability to de-anonymise should be restricted in accordance with orthodox privacy regulations. If a party chooses to re-identify people in an ostensibly de-identified dataset, without a good reason and without consent, then that party may be in breach of data privacy laws, just as they would be if they collected the same PII by conventional means like questionnaires or surveillance.
Surely we can all agree that re-identification demonstrations serve to shine a light on the comforting claims made by governments for instance that certain citizen datasets can be anonymised. In Australia, the government is now implementing telecommunications metadata retention laws, in the interests of national security; the metadata we are told is de-identified and “secure”. In the UK, the National Health Service plans to make de-identified patient data available to researchers. Whatever the merits of data mining in diverse fields like law enforcement and medical research, my point is that any government’s claims of anonymisation must be treated critically (if not skeptically), and subjected to strenuous and ongoing privacy impact assessment.
Privacy, like security, can never be perfect. Privacy advocates must avoid giving the impression that they seek unrealistic guarantees of anonymity. There must be more to privacy than identity obscuration (to use a more technically correct term than “de-identification”). Medical research should proceed on the basis of reasonable risks being taken in return for beneficial outcomes, with strong sanctions against abuses including unwarranted re-identification. And then there wouldn’t need to be a moral panic over re-identification if and when it does occur, because anonymity, while highly desirable, is not essential for privacy in any case.