Managing data privacy is becoming an increasingly difficult challenge for massive corporations littered with data silos. New data regulations–from the EU to the US to China— illustrate that this challenge is really just beginning.
This trend underscores the importance of anonymization – one of the most important tools in a data scientist’s “privacy toolbox.” Data anonymization is a technique that can be used to protect private information in your data while preserving, to varying degrees, the utility of that data; however, as we’ll see, this tool is only best put to use in combination with others, and not as a standalone strategy to protect your data.
What’s this thesis based upon? In a few words, the very real limits to anonymization. And, of course, Judd Apatow.
Anonymization, Utility, and Judd Apatow
To start with, think of anonymization as a technique to remove an individual’s identifying information from a dataset so that the remaining data cannot be linked to that individual. But that’s not the end of the game: the remaining data needs to actually be useful as well. This is what I call the “privacy vs. utility tradeoff.” If a dataset is perfectly anonymized, there is no risk in identifying an individual from that data, but that data also might (and probably will) be useless. In the cybersecurity world, there’s a saying that the safest computer is one that won’t function. And here the same point applies: ensuring anonymity usually requires sacrificing utility.
In many cases, anonymization techniques can be “fragile,” which is to say that even once you believe the utility vs. privacy tradeoff is balanced, the security of anonymized datasets can be dependent upon a variety of external factors that are hard to control. To illustrate this, I’m going to walk you through an example of what’s called a “linkage attack,” whereby an adversary uses knowledge external to the anonymized data to crack it.
This particular attack began when an analyst acquired all New York City taxi data for a given period of time using New York’s freedom of information act. This data included pickup and dropoff locations, pickup and dropoff times, fares, and tip amounts. At first glance, the data didn’t raise too many privacy flags on its own, which is why New York released the data in the first place. But then a Northwestern University graduate student used timestamped photos of celebrities getting into cabs in New York City to find specific cab rides within the released data. This meant he could determine how much a certain celebrity tipped the driver, making what was thought to be anonymous data suddenly newsworthy material.
Gawker picked up the story and published an article that exposed the tipping habits of Judd Apatow and a number of other celebrities. I’ve reproduced this linkage attack below.
The below picture of Judd Apatow and his wife, actress Leslie Mann, and the accompanying map of their trip both, appeared in the original Gawker exposé:
Photo credit: Gawker
SELECT pickup_datetime, dropoff_datetime, tip_amount FROM nyc_taxi WHERE pickup_datetime >= '2013-06-21 11:28:00' and pickup_datetime < '2013-06-21 11:29:00' AND pickup_latitude >= 40.719 and pickup_latitude <= 40.720 AND pickup_longitude <= -74.009 and pickup_longitude >= -74.011
Which results in a single entry, Judd’s trip:
|6/21/13 11:28||6/21/13 11:35||$2.10|
There are other ways to implement a similar link attack. If you had a taxi receipt, for example, you could use the cost of the ride and dropoff date/time based on when the receipt was created to discover the trip. For what it’s worth, New York did try to anonymize the data by hashing (a technique to mask data) the taxi medallion number (though even that attempt was crackable through a rainbow table). Note, however, that we didn’t even need to crack the masked medallion number to successfully perform the link attack above.
From Cardinality to Anonymity
The achilles heel of anonymization is in the columns that contain the most unique values, or what’s called high cardinality columns. Notice in the successful link attacks perpetrated with the NYC taxi data, the timestamp and location data was leveraged; because these columns have very high cardinality, only a small set of pickups and dropoffs occur simultaneously with the photograph, so one can easily isolate to a single result at the same date/time and location.
What should NYC have done differently to protect Judd’s tipping habits?
In addition to masking the taxi medallions, the city could have applied further anonymization to this dataset. If they generalized the pickup/dropoff to the nearest hour, then the link attack would fail. The same query I used above would return zero results because no entry would exist between those two minutes, since every entry would be rounded to the hour – my query is too specific for the data. This is a technique called k-anonymization. In k-anonymization, high cardinality values are generalized, or “blurred”, to try to prevent link attacks. In fact, you can think of k-anonymization as giving you more results because in this case I can no longer query for specifics within that hour – the data is coarser. Here, we’re not just protecting sensitive values; we’re forced to protect columns that simply have many unique values, what are sometimes termed quasi-identifiers.
Unfortunately, k-anonymization is by no means foolproof. Even if we round the pickup/dropoff date/time in the data to prevent from this attack, the adversary could still break NYC’s efforts to preserve the anonymity of the dataset. If the adversary knew the location of the pickup and the dropoff, for example, and both data points together were unique, likely even within the hour, this would again break the attempts at anonymization.
This is because the pickup/dropoff locations also have high cardinality. The location coordinates could also be k-anonymized (generalize the latitude and longitude) to preserve anonymity, but that’s where the privacy vs utility tradeoff rears its head. Are we taking the privacy preserving measures so far that the data is now becoming useless? To prevent potential attacks, are we also preventing useful analysis? Do we also need to k-anonymize the cost of the ride? That’s fairly high cardinality as well… . . . you get the idea.
There is a different technique termed differential privacy, which does have mathematical guarantees on privacy. You are limited to understanding aggregate trends in your data, and have limits on how many queries you can actually make. This is extremely powerful because of the privacy guarantees; unlike k-anonymization, however, there remains a significant utility tradeoff due to the types of queries and limits on those queries.
Who Really Cares? The Europeans.
So who cares about all of this? At the very top of the list is the Europeans, who implemented the General Data Protection Regulation (GDPR) to protect data subjects’ private information. A violation can result in a fine equal to four percent of a company’s global revenue. Which makes the GDPR the most serious data regulation on the planet, and it applies to any business using European data, so if you’re a data scientist working with global data, chances are that the GDPR will apply to you.
When it comes to anonymization, the GDPR has some very blurred guidelines (no pun intended). Let’s delve into some legalese. The recitals section of the GDPR is where the lawmakers comment on the text of the laws themselves. In recital 26, they state that the GDPR does “not apply to anonymous information,” which is defined as information that “does not relate to an identified or identifiable natural person or to personal data rendered anonymous in such a manner that the data subject is not or no longer identifiable.” Anonymization, then, seems to be a potential “get out of jail free card,” meaning that if a dataset is truly anonymized, the GDPR wouldn’t apply.
Elsewhere in the GDPR, this concept of anonymization is distinguished from pseudonymisation, which is defined as “the processing of personal data in such a manner that the personal data can no longer be attributed to a specific data subject without the use of additional information, provided that such additional information is kept separately and is subject to technical and organisational measures to ensure that the personal data are not attributed to an identified or identifiable natural person.” This is essentially what NYC attempted to do when they masked the taxi medallions.
Under the GDPR, anonymized data is data that is “no longer identifiable,” and that data isn’t regulated by the GDPR. Pseudonymized data, on the other hand, requires additional data to identify the data subject… which begs the question: is there such a thing as data that can’t be identified given additional information? New York certainly thought so, but the answers are far from certain. At worst, there might be no acceptable form of anonymization under the GDPR.
Making Anonymization Work for You
Given the inherent risks, anonymization has to be just one piece of the privacy puzzle. Anonymization techniques need to take place within a larger information governance framework that can provide both anonymization capabilities and the ability to dynamically control data “states” based on the user’s attributes and purpose at that time. That includes providing an environment for executing analyses, rather than dumping data to third parties and losing visibility on how that data is being acted upon, which is what some are tempted to do using “anonymized” data sets.
So what does this really mean in practice? It means that anonymization without context isn’t going to cut it. More precisely, a dynamic data abstraction layer is key.
A dynamic data abstraction layer is a thin layer that sits on top of all your data, making decisions based on your policies and how data is presented to consumers. This allows organizations to restrict access based on certain purposes (analytical context), or certain attributes of a user. It also allows organizations to audit all actions against the data through a unified view, enabling report generation and a complete, real-time understanding of what policies are being enforced when and by whom. Most importantly, it actually enables better data science because this abstraction layer provides a consistent and unified view of the data, which makes sharing analysis easier, as the data comes from the same place for everyone.
Let’s return to our taxi example to see what an abstraction layer would look like in practice.
If the New York could have exposed their data through a data abstraction layer, they would have had the ability to link purposes, user attributes, and context to how the data is protected. Through that abstraction layer, for example, they could mandate that users only access their data for certain purposes. And if a user acknowledged his or her purpose as finding aggregate trends in the data, the taxi data could be dynamically protected through differential privacy for that purpose at that time. If the analyst then wished to analyze individual routes from a specific location, locations could be kept accurate, but k-anonymization could be applied to the pickup and dropoff times (again, dynamically, for that purpose at that time). Not only would finding Judd’s tip break the acknowledgement of purpose set forth at time of access; the dynamic anonymization, plus audit, can guarantee this won’t happen.
Anonymisation techniques can provide privacy guarantees, but only when they are engineered appropriately. Balancing the privacy vs utility tradeoff is all about context, and an abstraction layer is what allows for that context. If New York had taken advantage of these additional tools, celebrity journalists might not have been happy. But for those who care about the privacy and security of our data, it would represent a win, and Judd’s tipping habits would have been protected.
Originally posted at www.kdnuggets.com