Information about us is being constantly collected, through our phones and the services we use online. This data is hugely valuable but also highly personal, and often sensitive. This raises a crucial question: can we use this data without disclosing people’s private information? We studied Diffix, a system developed and commercialized by Aircloak to anonymise data by adding noise to SQL queries sent by analysts. In a manuscript we just published on arXiv, we show that Diffix is vulnerable to a noise-exploitation attack. In short, our attack uses the noise added by Diffix to infer people’s private information with high accuracy. We share Diffix’s creators opinion that it is time to take a fresh look at building practical anonymization systems. However, as we increasingly rely on security mechanisms to protect privacy, we need to learn from the security community: secure systems have to be fully open and part of a larger layered security approach. Privacy is hard, it is time to admit that we won’t find a silver bullet and start engineering systems.

In the last decade, the amount of personal data being collected and used has exploded. With half of the world population soon having access to the Internet and with the Internet of Things becoming a reality, this is unlikely to stop anytime soon. While data has immense potential for economic development and scientific advancements, its collection and use raises legitimate privacy concerns. More than 80% of U.S. citizens are concerned about sharing personal information online. Data, directly or indirectly, contains sensitive information that could be used against individuals.

To prevent this and allow organizations to collected and use data while preserving people’s privacy, data is often anonymized. The idea behind this is that if the data is anonymous, if one can’t know that user MjJ17torTC is you, the data can’t be used against you. Data anonymization, also called de-identification, is a two steps process: first direct identifiers such as name, social security numbers or email addresses are removed and then noise is added to the dataset.

However, a large body of research has shown that pseudonymized and even anonymized data can often be easily linked back to you, re-identified. This increasing amount of evidence has led President (Obama)’s Council of Advisors on Science and Technology (PCAST) to conclude in 2013 that data anonymization “is not robust against near-term future re-identification methods” and that they “do not see it as a useful basis for policy”.

Privacy researchers have therefore been increasingly interested in the potential of question-and-answer (or query-based) systems as a way to use data without disclosing sensitive information.

The idea of question-and-answer systems is simple: rather than sharing the raw anonymized data with analysts (a model called release-and-forget), data holders could allow analysts to remotely ask questions from the data and only get aggregated answers back. For instance, a question could be: “What is the average income of men older than 50?”

However, avoiding direct access to the data is, alone, not sufficient to ensure that privacy is preserved. Without additional security measures, query-based systems are susceptible to a wide range of attacks. For instance, it does not prevent the analyst from accessing private information by asking the right question, such as how many users named Edward Snowden were diagnosed with cancer in 2017, or a combination of right questions. Attacks relying on multiple, seemingly innocuous, queries such as averaging attacks or intersection attacks have been developed over the years.

Diffix

German startup Aircloak, along with researchers from the Max Planck Institute for Software Systems, developed and commercialized a system called Diffix to protect SQL databases from rogue analysts. Diffix relies on a novel, patented, and proprietary approach, called sticky noise, which adds noise to each answer to a query with the noise being based on the query.

Aircloak says their approach allows analysts to ask an infinite number of queries, with a rich query syntax and minimal noise, all the while strictly preserving people’s privacy. According to them, Diffix (1) falls outside the scope of GDPR regulations, (2) has been guaranteed to deliver GDPR-level anonymity by the French Data Protection commission (CNIL) and (3) certified by TÜViT as fulfilling “all requirements for data collection and anonymized reporting”.

In a manuscript we just published on arXiv, we show that Diffix is vulnerable to a new attack we developed, which we call noise-exploitation attack. We show in the paper how an attacker can successfully learn the sensitive attribute (e.g. HIV status) of someone in the dataset protected by Diffix knowing a set of attributes that uniquely identify them in the dataset (say, age, ZIP code, and education level). Our main contribution is a novel (to the best of our knowledge) technique to exploit the noise added by Diffix as a “signal” to learn the target’s private information.

In order to understand how our attack works, let us briefly explain the “sticky noise” used by Diffix. When Diffix processes a new query, it computes the accurate result of the query on the dataset and then adds several layers of noise to the output. There are two types of noise layers: static noise layers, that depend on the query expression (i.e. the question asked), and dynamic noise layers, that depends on the set of users selected by the query (the query set). To prevent simple attacks, Diffix also does not report results whose number of users is below a certain threshold (randomly selected according to a normal distribution $N(4, \frac{1}{2})$), they call this bucket suppression. For a query $Q$, Diffix’s output is $\textsf{output} = \textsf{true_value} + \textsf{static}_Q + \textsf{dynamic}_Q$, where $\textsf{static}_Q$ and $\textsf{dynamic}_Q$ denote the sum of all static and dynamic noise layers respectively.

Our noise-exploitation attack relies on three steps to circumvent Diffix’s protection (all the details and numerical simulations are available in the manuscript):

  1. We can design queries that are similar enough that they will share part of their static noise. This allows us to cancel out some of Diffix’s noise.

  2. As Diffix’s noise depends on the query set, the noise itself leaks information about the query set. This is the core of our attack: analyzing the remaining Diffix’s noise we develop a statistical test to learn information about the data Diffix is protecting.

  3. We exploit logical equivalence between queries to circumvent some of the “stickiness” of the noise by repeating (almost) the same query. This allows us to obtain fairly independent noise samples which, when added to our statistical method, increase the power of our attack.

Combining these steps, we developed a powerful attack allowing an attacker to potentially infer a user’s attribute with high accuracy with very limited auxiliary information. For instance, we showed that knowing only four attributes could allow an attacker to learn a user’s private information with 99% accuracy making this an important vulnerability. While further research is needed, we do not see a direct way for Aircloak to prevent this attack.

Accuracy of our noise-exploitation attack on Diffix.
Accuracy of our noise-exploitation attack on Diffix as function of the number of attributes known to the attacker.

Moving Forward

The privacy of question-and-answer systems has been heavily researched under the theoretical framework of differential privacy introduced in 2006. While differential privacy proposes a solid and theoretically appealing mathematical foundation for the protection of privacy, it has so far mostly failed to offer a practical solution to the protection of modern datasets. While a few recent exceptions exist, many (including Diffix’s creator) believe that it will not be a viable solution for general use cases such as the one tackled by Diffix.

It is therefore time to start investigating new approaches to the privacy of query-based systems including privacy-through-security approaches. While theoretical frameworks with provable guarantees such as differential privacy are likely to continue playing an important role in privacy research, they cannot be the only—or maybe even the main—focus of the community. Moving forward, penetrate-and-patch adversarial approaches are likely to become a crucial part of building practical privacy-preserving systems even when differentially private. Even provable guarantees are not enough to make a system safe. Implementation issues, hypotheses, and design choices may always introduce vulnerabilities.

This adversarial engineering approach is powerful but hard as it requires us to change both our expectations of privacy-preserving mechanisms and our way of thinking about privacy. We need to accept that no system is perfect. There will be attacks, and some of them will succeed. We need to prepare for this and learn from best practices in security: ensuring that several layers of security exist, to not have all the data in one place (what Jean-Pierre Hubaux calls the Fort Knox approach), etc. We also need standards and systems to be completely transparent and open. Building secure systems requires anyone to be able to review the code without technical or legal barriers, propose solutions and build upon existing work.

Fresh approaches to protecting privacy are essential moving forward but we need the right environment around them.

UPDATE — May 17, 2017

Paul Francis —director at the MPI-SWS and co-founder of Aircloak— published on April 27 a blogpost and Felix Bauer —CEO of Aircloak— a statement regarding our attack. While both acknowledge the vulnerability we disclosed, they claim that “the conditions under which it could work are so rare as to be practically non-existent”. Their claim is based on an empirical analysis of open-data datasets on which the attack would only sometimes succeed, depending on properties of the dataset. We are currently evaluating their analysis and running our own experiments.

Paul Francis also stated on Twitter that this is a “immediate vulnerability disclosure” implying that we did not contact them before publishing the manuscript and blogpost. This is not accurate. We submitted the manuscript to Arxiv (which receives and publishes articles in bulk the day after), the goal being to protect us from potential cease and desist letters or other threats, and e-mailed Paul Francis and Aircloak right after. Paul Francis answered our e-mail a couple of hours later and did not ask or suggest we should delay the public disclosure. We have, as of today, not received a response from Aircloak.