Hacker News Reader

Comment by GuB-42 | Hacker News Reader

The underlying claim here is that de-identification doesn't work, the articles then explores the consequences of that.

But the real question should be: why doesn't de-identification work and how to make it work. It is a technical problem, and I thought it was more or less solved. But the author here things it is just placebo. If so, what is the problem exactly? Is there a fundamental problem with the very idea of de-identification? Is there a "bug" in the process? What level is required to conduct these attacks? Can in individual do it? A cybercrime gang? A nation state? Is it only theoretical?

Depending on that, the answer could be very different.

View on HN

It's a mathematical problem, not a technical problem. You can remove features from a dataset that, on their own, serve to uniquely identify someone, but that doesn't change that the remaining features still each partition the population into smaller and smaller subsets until eventually the subset is one or only a few people. Once you get down to "only a few people," metadata may be able to bring that to one, i.e. there are a few hundred people it could potentially be, but only one was actually at the hospital the data was collected from at the time it was collected. You don't necessarily even need the hospital to leak that metadata. Maybe this person was the only one who even lived near that hospital near that time.

The only way to anonymize the data in a foolproof way is to remove so many features that it becomes useless for statistical research.

Modeling-wise, regression, classification, and clustering alike all rely on being able to construct a model that minimizes entropy. But anonymity relies upon maximizing entropy. This fundamental conflict can't really be resolved. You pick some point in the middle of the spectrum that ends up serving neither purpose well.

Math problems aren't technical problems? Do you mean it's a mathematically-impossible technical problem?

Noise and obfuscation can be added to the data such that the unique person it identifies doesn't exist. For example changing the dates by some limited random amount. Replacing values with roughly-similar choices (replace engineer with scientist, replace Mexico with el Salvador, etc.) The technical problem is doing that in a way that doesn't harm its research value, never allows the final set of possible similar identities to be below some lower limit, and can't be reversed of course.

Would you be willing to provide a few references for this? My professional work has skirted around the discussion at times, and I would be fantastically interested in learning more.

De-identification is in the eye of the data steward and their legal folks.

So it’s the definition of de-id that doesn’t actually “work” in that it’s possible to reidentify small amounts from de-id data, and that adds up over time.

For example, HIPAA considers data de-id if you remove 19 fields or expert determine that it’s de-id. [0]

What’s expert determination, who’s an export? That’s up to me to decide and my lawyers to accept.

The bug is that if half a percent of people each de-id datasets can be reidentified, likely acceptable in HIPAA, then each data released adds up for reidentified people. And more datasets allow for triangulation and linking to reidentify.

The article calls this out as a risk, not a certainty as it’s unclear if anyone is doing this. But the process would be something like: 1) buy HIPAA de-identified data since it doesn’t require patient consent 2) reidentify patients using other data publicly for sale (marketing data, voter registration, etc) 3) new data is not longer HIPAA restricted and fully identified health records can be sold for whatever you like (eg, super targeted drug marketing)

[0] https://www.hhs.gov/hipaa/for-professionals/privacy/special-...

The US federal government requires users to agree to terms in return for receiving potentially identifiable data. Usually those terms include something like, "I agree to not attempt to use these data to identify any individuals." Granted, an agreement carries more weight when it's with the federal government (lying to them is criminal). But this kind of clause can prevent legal groups from reliably trading in identified data, even if it's from a private hospital. If they break the agreement, they'll probably never again receive new data from the provider, and stale data is low value.

> It is a technical problem, and I thought it was more or less solved.

As a technical problem, it is more or less understood to be unsolvable.

The core issue is the more information you strip away, the more sanitized the data is, the less useful it is. Once the statistics of your "representative" set are no longer that (even worse, weirdly biased) you are very limited in what you can do. This works ok if you already 100% know the data you need and don't, but otherwise it falls apart pretty quickly.

On the other hand, relatively small numbers of innocuous data points about individuals will identify them uniquely.

These two points are in fundamental contention.

It doesn’t work because the data can be correlated against other sets of data.

When my wife almost died from an ectopic pregnancy, Enfamil was confident enough of the would-be due date based on data pieces together by a broker to FedEx us a box of formula on the due date.

Third parties get a real time feed of insurance claims, hospital admissions, prescriptions and other data. They put the puzzle together and make an assertion of valuable diagnoses. (Pregnancy, diabetes, etc)

They don’t need to be 100% accurate. The collateral damage of reminding someone who lost a child and nearly lost their life doesn’t have a cost to them.

> They don’t need to be 100% accurate.

In many cases, abusing such data will also be profitable even if the result estimate is highly noisy.

Just as an example, let's assume that an insurance company knows your given name when you apply and they have statistical data which shows that if your name belongs to one set of names, you are smoking with a probability of 51%. If you have another name, you are smoking with a probability of 49%, because the correlation is noisy. It is still more lucrative for the company to exclude applicants from the "more likely to smoke" names group, because smoking causes cost. You might never have tried a single cigarette in your life but with the information the company has, it is more profitable to not insure you.

Of course the above is an exaggerated example which would not work out quantitatively (the profit they make by insuring you might be higher than the expected value of risk from smoking) but the principle still holds; it is used for credit scoring and similar things.

It’s more insidious than that. Based on proprietary databases, insurers can target advertising or marketing at you based on profiles that you don’t know exist.

If you want to avoid insuring a protected class, you might be able to identify them based on those databases, and use advertising to avoid the folks you don’t want. (This is common for student apartments)

Some medical data, such as genome sequence data, is unique to an individual (where the uniqueness is partly dependent on the length of sequence data generated - short sequences may not be unique). Even worse, whole exome or genome data also describes your lineage and may reveal sensitive information about relatives, living and not. Deidentification of such genomic data is especially difficult.

But the real question should be: why doesn't de-identification work and how to make it work.

How much health data is required to distinguish a specific individual even without any "identifying" information? Surely there must be many people with unique combinations of conditions and treatments. If that is so then a data set that includes those people could only ever be pseudonymous, not truly anonymous.

What other data sets might then be combined with the health data to match a pseudonymous record with an identifiable person? Payment records? Travel records? Time off work? Insurance claims?

I don't know much about how the US healthcare system works, but maybe before asking how to make de-identification of health data work we should be asking whether it could ever work effectively at all.

It appears there is a spectrum of de-identification tools, some being better than others.

https://en.m.wikipedia.org/wiki/Data_re-identification

I suspect that the more data they have for a particular patient (many visits over multiple years) makes the re-identification process easier. The article mention financial data, if dates and amounts of charges aren't masked or altered that could be cross-referenced with another data source to deduce the person's identity.

People have mentioned some good cases, but I think the most straightforward one is radiology.

Imagine you have high quality 3D scans of people. You can strip the names off of the files, but anyone can volume render the data to see what they look like (inside and out).

There are some papers on adjusting the data, while hopefully preserving anything medically important, but that's obviously a challenging problem.

Because it's an arms race situation.

The industry of de-identification has a long history in fact.

It is typical for many types of criminals to employ hard-to-identify outfit and appearance. The work of investigators is to overcome this hardship and identify the criminal. The co-evolution of methods on both sides will never end.

Reply on news.ycombinator.com