Data Anonymization and Re-identification: Some Basics Of Data Privacy

From his Strata Conference on Data Science, Tim O'Reilly tweeted with dismay the recent California court decision that the zipcode is now to be classified as "personally identifiable information". "No more demographics" he lamented. A little later he retweeted a response that "apparently 87% of US residents can be uniquely identified by zip+DOB+gender: bit.ly/qysMqs" and later followed up with "Here's a reference for the claim that zip code, gender and DOB uniquely identify 87% of individuals: http://www.citeulike.org/user/burd/article/5822736 via @crdant".

These tweets are odd and disturbing. The zip/DOB/gender finding is a basic one in studies of privacy, published years ago by Latanya Sweeney of Carnegie Mellon University. I gave a talk at work on privacy a year ago, and this was one of the first references I came across. Tim O'Reilly has been pushing an agenda of Open Data, particularly Open Government Data, for the last couple of years, and yet it looks as if he isn't aware of the basic privacy issues around such data. Can that really be the case?

If it is, then here, to help Tim along, are some notes from my talk as a kind of introduction to data privacy, or at least to data-anonymization and re-identification. A great resource on some of these issues from a legal perspective is Paul Ohm's 2009 paper "Broken Promises of Privacy: Responding to the Surprising Failures of Anonymization" (PDF), University of Colorado Law Legal Studies Research Paper No. 09-12. It's long, but it's so well written it's an easy read. Much of these notes originated with this paper, in one form or another.

A few years ago Netflix ran its highly successful and widely publicised crowdsourced prize competition, in which it released a data set of users and their movie ratings and let competitors download them and search for patterns. The data consisted of a customer ID (faked), a movie, the customer's rating of the movie, and the date of the rating.

In the FAQ for the competition, Netflix said this:

Q. Is there any customer information in the dataset that should be kept private?

A. No, all customer identifying information has been removed; all that remains are ratings and dates. This follows our privacy policy… Even if, for example, you knew all your own ratings and their dates you probably couldn't identify them reliably in the data because only a small sample was included (less than one tenth of our complete dataset) and that data was subject to perturbation.

This certainly looked reasonable enough, but Arvind Narayanan and Vitaly Shmatikov of the University of Texas had other ideas.¹ First, they looked at the claim that the data was perturbed by asking acquaintances for their rankings. They found that only a small number of the ratings were perturbed at all, which makes sense because perturbing the data gets in the way of its usefulness.

In the Netflix data set, different users have distinct sets of movies that they have watched. The data set is sparse (most people have not seen most movies), and there are many different movies available, so individual tastes and viewing histories leave a clear fingerprint. That is, if you knew what movies someone watched, you could pick them out of the data set because no one else would have seen the same combination.

A closer look showed that with 8 ratings (of which 2 may be completely wrong) and dates that may have a 14-day error, 99% of the records in the Netflix data set uniquely identify an individual. For 68% of records, two ratings and dates are sufficient. Various combinations of information are sufficient to identify users, eg 84% by 6 of 8 movies outside the top 500.

But of course there is no personally identifiable information in the data set. So is this a privacy issue? It is when you have another data set to look at. The researchers took a sample of 50 IMDB users. The IMDB data is noisy – there is no ranking, for example. Still, they identified two users whose Netflix records were 28 and 15 standard deviations away from the next best. One from ratings, another from dates.

So despite Netflix's best efforts, the data set included enough information to identify some individuals. Partly because of this, a planned follow-up competition was scrapped, and the whole enterprise of crowdsourcing recommender algorithms was given a possibly terminal blow.

Just to be clear, this set of notes is not about the following things:

• Encryption
• Restricting access to data
• Lost USB keys and CDs

It is about these:

• Deliberately released data that turns out to infringe on privacy
• HIPAA, EU Data Directive, corporate rules for handling customer data
• Advertising and ISPs
• Gov 2.0, data.gov, and "openness"

It's about claims such as: "Attorneys on Monday accused Google of intentionally divulging millions of users' search queries to third parties in violation of federal law and its own terms of service" (October 26 2010)

"MySpace and some popular applications on the social-networking site have been transmitting data to outside advertising companies that could be used to identify users, a Wall Street Journal investigation has found" (October 23, 2010)

"Facebook users may inadvertently reveal their sexual preference to advertisers in an apparent wrinkle in the social-networking site's advertising system, researchers have found" (October 22, 2010)

(These claims are a year old, found in the week before I gave the talk. I'm sure there are many more.) The Facebook case was one in which advertisers (for a nursing program I believe) asked to target their ads specifically at females and at men interested in other men. But unlike, for example, an ad about a gay bar where the target demographic is blatantly obvious, a male user reading the ad text would have no idea that it had been targeted solely at a very specific demographic, and that by clicking it he would reveal to the advertiser both his sexual preference and a unique identifier (cookie, IP address, or e-mail address if he signs up on the advertiser's site). "Furthermore (the researchers wrote) such deceptive ads are not uncommon; indeed exactly half of the 66 ads shown exclusively to gay men (more than 50 times) during our experiment did not mention 'gay' anywhere in the ad text."

Indeed we do. Europe and the USA adopt different approaches to balancing privacy and utility, with the US adopting industry-specific rules (HIPAA for health, FERPA for education, Driver's Privacy Protection Act, FDA regulations, Video Privacy Protection Act etc), while the EU has taken a global approach with the Data Protection Directive. But both approaches are based on a common set of concepts and assumptions.

The big thing is that there is an assumption that data can be anonymized, and once it is then you can share it, because where's the harm? Both sets of rules are built on the idea that there is such a thing as personally identifiable information (PII) and that you can hide it, while still releasing data that is useful. The release process is "release and forget" because if data is properly anonymized why do you have to track what's done with it? There is a faith in the anonymization process, and that faith was broken by the Netflix study and a couple of other related studies.

Let's go back to a time before HIPAA, when the debate was focused in terms of how much anonymization you needed to do. Here are some quotations from Latanya Sweeney's paper (PDF), that Tim O'Reilly appeared unaware of.

"Figure 1" below is a simple Venn diagram with two intersecting circles. The left circle holds medical data: Ethnicity, Visit Date, Diagnosis, Procedure, Medication, Total Charge. The right circle holds a voter list: Name, Address, Date Registered, Party Affiliation, Date Last Voted. And in the intersection is ZIP, Date of Birth, Sex.

The National Association of Health Data Organizations (NAHDO) reported that 37 states in the USA have legislative mandates to collect hospital level data and that 17 states have started collecting ambulatory care data from hospitals, physicians offices, clinics, and so forth. The leftmost circle in Figure 1 contains a subset of the fields of information, or attributes, that NAHDO recommends these states collect; these attributes include the patient's ZIP code, birth date, gender, and ethnicity.

In Massachusetts, the Group Insurance Commission (GIC) is responsible for purchasing health insurance for state employees. GIC collected patient-specific data with nearly one hundred attributes per encounter along the lines of the those shown in the leftmost circle of Figure 1 for approximately 135,000 state employees and their families. Because the data were believed to be anonymous, GIC gave a copy of the data to researchers and sold a copy to industry.

For twenty dollars I purchased the voter registration list for Cambridge Massachusetts and received the information on two diskettes. The rightmost circle in Figure 1 shows that these data included the name, address, ZIP code, birth date, and gender of each voter. This information can be linked using ZIP code, birth date and gender to the medical information, thereby linking diagnosis, procedures and medications to particularly named individuals.

For example, William Weld was governor of Massachusetts at that time and his medical records were in the GIC data. Governor Weld lived in Cambridge Massachusetts. According to the Cambridge Voter list, six people had his particular birth date; only three of them were men; and, he was the only one in his 5-digit ZIP code. [Editor's note: a 5-digit zip code may have several thousand people in it.]

The example above provides a demonstration of re-identification by directly linking (or "matching") on shared attributes. The work presented in this paper shows that altering the released information to map to many possible people, thereby making the linking ambiguous, can thwart this kind of attack. The greater the number of candidates provided, the more ambiguous the linking, and therefore, the more anonymous the data.

In a theatrical flourish, Dr. Sweeney sent the Governor's health records (which included diagnoses and prescriptions) to his office.

Now, of course, health information in the US is governed by HIPAA, but according to HIPAA, "de-identified" health information is unregulated. De-identified means either: a statistician says it is de-identified, or the 18 Personally Identifying Information (PII) identifiers are suppressed or generalized. These PIIs are things like Name, e-mail address, social security numbers, computer IP addresses, and so on.

The EU doesn't list specifics. Instead it says that PII is "anything that can be used to identify you". But what does that cover? IP addresses perhaps? Here is Google in their argument to the EU:

• we "are strong supporters of the idea that data protection laws should apply to any data that could identify you. The reality is though that in most cases, an IP address without additional information cannot."
• "We believe anonymizing IP addresses after 9 months and cookies in our search engine logs after 18 months strikes the right balance."
• "we delete the last octet after nine months (170.123.456.XXX)"

The Latanya Sweeney result was the first to show that once you can mix and match data sets, PII is just not enough to provide privacy. And nowadays, of course, data mining multiple data sets is big business.

Let's step back a little and look at the technical side of anonymization. There are four basic methods for anonymizing data:

Replacement - substitute identifying numbers
Suppression - omit from the released data
Generalization - for example, replace birth date with something less specific, like year of birth
Perturbation - make random changes to the data

Then you have to measure how private a data set. Latanya Sweeney came up with the notion of k-anonymity to define this. Here's how it works.

Think about a table, with rows and attributes. Each attributes is either part of a quasi-identifier (like a name or address), or is sensitive information (like the fact you had an operation on a particular afternoon). A quasi-identifier is a set of attributes that, perhaps in combination, can uniquely identify individuals. Sensitive information includes the attributes that we want to keep private. Your driving license number is an identifier; our driving record is sensitive information. The table satisfies k-anonymity iff each sequence of values in any quasi-identifier appears with at least k occurrences. Bigger k is better.

So here is a table with 11 rows.

If we remove all the attributes except for the problem we have a very anonymized data set (k = 11):

On the other hand, if we just remove the name and generalize the zip code and date of birth we have a less anonymized set. Exercise: convince yourself that k=2 for this set.

Of course, the issue is utility. There is a tradeoff between keeping the data useful for research and maintaining privacy. Researchers and attackers are doing the same thing after all: looking for useful patterns in the data. With the k=2 data set you can ask questions about correlation of problems with gender, or with geography to some extent (although not very specific geographical factors, like toxic leaks).

It would be nice if you could make the data set anonymous for the purposes of attackers, but still useful for researchers. But it turns out you can't. In a paper called The Cost of Privacy: Destruction of Data-Mining Utility in Anonymized Data Publishing, Justin Brickell and Vitaly Shmatikov investigated the problem. They took a set of different sanitization methods and compared it to a data set with trivial sanitization (removal of identifiers). Here are the results.

The left bar of each pair is the privacy (smaller = more private), The right represent the utility to the researcher (bigger = more useful). Anonymization seeks to shorten left without shortening the right, but the results show, depressingly, that small increases in privacy cause large decreases in utility.

OK, if you insist.

If we are going to take a new look, we need to recognize that privacy is not a binary issue, and it is not a property of a single data set. We need to worry about reidentification attacks that do not reveal sensitive information. As Paul Ohm writes: "For every person on earth, there is at least one fact about them stored in a computer database that an adversary could use to blackmail, discriminate against, harass, or steal the identity of him or her… the 'database of ruin' containing this fact but now splintered across dozens of databases on computers around the world."

Privacy is erased incrementally as successive queries reduce uncertainty and narrow in on an individual. The way to quantify this reduction in uncertainty is to use the idea of information entropy, adopted from the thermodynamic concept, and usually identified by H. The information gained in a query is

H(before) – H(after)

as you increase your knowledge of a system, the entropy (loosely, disorder) decreases.

So what is the formula for H?

For a set of outcomes {i,…}, each with probability p_i, the information entropy is:

H = – SUM p_i log₂(p_i)

(excuse the lack of greek sigma), and is measured in bits. The logarithm appears because the probability of two independent events occurring is the product of the probabilities of each event, but the information we gain from observing two independent events is the sum of the information we gain from each event

Take a simple example: a coin toss. Before the toss, there are two outcomes with equal probabilities, so

H = -(1/2) log₂(1/2) – (1/2)log₂(1/2)

= – log₂(1/2)

= log₂(2)

= 1 bits

which makes sense if you think about it, because the coin could be heads (1) or tails (0).

After the toss, H = log₂(1) = 0 : there is no uncertainty left and we have complete information about the system.

In the same way, a dice roll has (before rolling it) an entropy of

H = log₂(6) ~ 2.6 bits

So if the challenge is to identify one person in the population of the world, how much information entropy is there? The identity of a random, unknown person is just under 33 bits (233 ~ 8 billion). Hence the web site 33bits.org.

Learning a fact about the individual reduces the uncertainty (reduces information entropy). So if you learn that the star sign is Capricorn then that's -log₂(1/12) = log₂(12) = 3.58 bits of information.

If you find out other independent pieces of information you add up the contributions to find out how much the the entropy has been reduced. So a ZIP code may provide 23.81 bits of information, a birthday 8.51, and gender 1 bit for a total of 33.32 bits: it probably identifies one individual.

The Netflix de-anonymization paper used these ideas a bit. The a priori entropy of the data set (additional information required for complete de-anonymization) is 19 bits (2¹⁹ = 524288, which is about the number of individuals in the data set). Individual movies give from 1 to 18 bits of information, depending on what you know about them (dates within 14 days, rankings +/- 1). Very popular movies gave little information, but niche movies viewed by few individuals yielded many bits of information. So little auxiliary information is needed to re-identify records in the database. In a theoretical excursion, the researchers showed that de-anonymization is going to be robust against noise, and does not need much additional information, so long as the data set is large and sparse enough.

I can do nothing better than quote from Paul Ohm to summarize the privacy dilemma we find ourselves in.

"Abandoning PII is a disruptive and necessary first step, but it is not enough alone to restore the balance between privacy and utility that we once enjoyed. How do we fix the dozens, perhaps hundreds, of laws and regulations that we once believed reflected a finely calibrated balance but in reality rested on a fundamental misunderstanding of science?

Techniques that eschew release-and-forget may improve over time, but because of inherent limitations like those described above, they will never supply a silver bullet alternative. Technology cannot save the day, and regulation must play a role.

Ohm notes that the US sectoral approach to regulation sets the privacy bar too low, by focusing on explicitly listed PII. Meanwhile the EU, by saying "anything that can be used to identify you" would, if interpreted in the light of modern de-anonymization techniques, be too high.

The direction to take, says Ohm, is to focus on the people, not the anonymization, and to distinguish private from public release. We need to codify notions of trust and practices and apply strong sanctions against re-identification. This will put more administrative and procedural burdens in our future, but is needed to preserve privacy.

And, to return to Tim O'Reilly's tweet, open data advocates and big data enthusiasts need to pay more attention to these issues rather than relying, as some do, on personally identifying information as a sufficient solution.