New tools allow researchers to share and study sensitive data safely by applying ‘differential privacy’
The promise of big data lies in researchers’ ability to mine massive datasets for insights that can save lives, improve services and inform our understanding of the world.
These data may be generated by surfing the web, interacting with medical devices or passing sensors. Some data may be trivial, but in many cases, data are deeply personal. They can even influence our insurance premiums or the price we pay for a product online.
When planning a study, data scientists need to balance their desire to uncover new knowledge with the privacy of the people whom the data represent.
“The science of understanding human behavior, health, and interactions is being transformed by the ability of researchers to collect, analyze, and share data about individuals on a wide scale,” a team of Harvard University researchers wrote in a July 2014 paper, “Integrating Approaches to Privacy across the Research Lifecycle.”
However, the paper continued, “a major challenge for realizing the full potential of such data science is ensuring the privacy of human subjects.”
Initially, researchers believed that anonymizing data–erasing the names and replacing them with arbitrary identifiers–was enough to protect the identities and personal information of those who had agreed (knowingly or unknowingly) to contribute information. However, in a well-known study published in 2000, Latanya Sweeney led a team that uncovered the identities of patients, including then Massachusetts governor William Weld, by correlating anonymized data with other publicly available data.
In a more recent case, researchers Arvind Narayanan and Vitaly Shmatikov from The University of Texas at Austin partially de-anonymized a Netflix dataset containing half a million movie reviews. By cross-referencing that dataset with information in the Internet Movie Database, the researchers showed that attackers could potentially identify known users, compromising their data.
As cases of re-identification and de-anonymization emerge, researchers are exploring new, more robust approaches to privacy protection.
Salil Vadhan, a professor of computer science at Harvard University and former director of the Center of Research on Computation and Society, is among the researchers exploring an approach known as “differential privacy” that allows one to investigate data without revealing confidential information about participants. Initially introduced by Cynthia Dwork, Frank McSherry , Kobbi Nissim and Adam Smith, among others, in the mid-2000s, researchers continue to develop the concept today to apply it for real-world problems.
As the lead researcher for the National Science Foundation (NSF) supported “Privacy Tools for Sharing Research Data,” Vadhan and his team at Harvard are developing a new computer system that acts as a trusted curator–and identity protector–of sensitive, valuable, data. (The Sloan Foundation and Google, Inc. are providing the project with additional support.)
The system works like this: Researchers ask the virtual curator questions based on the data–for instance, “What percentage of individuals who have Type B blood are also HIV positive?” The computer returns an answer that is approximately accurate, but that includes just enough “noise” that no matter how hard someone tries, they cannot find out anything specific to any individual in the database.
“Even if an adversary tries to target an individual in the dataset, the adversary should not be able to tell the difference between the world as it is and one where that individual’s data is entirely removed from the dataset,” Vadhan said. “Randomization turns out to be very powerful.”
If the system is implemented simply, the level of privacy degrades with multiple queries, so one could keep asking questions until the point where identifying people in the database becomes possible. However, by judiciously increasing the amount of noise and carefully correlating it across queries, the system can maintain privacy protection, even in the face of very large number of questions.
Differential privacy has become a hot topic in recent years. A 2015 Science magazine article referred to differential privacy as one of the most promising technical solutions for protecting the data of students enrolled in Massive Open Online Courses (MOOCs). Projects including OnTheMap, used for U.S. census data, RAPPOR, a new product from Google, apply forms of differential privacy for data sharing.
Speaking at the NSF in early 2015, Vadhan explained how ideas from theoretical computer science inspired the development of differential privacy algorithms, which have are now entering the research ecosystem. Harvard’s Institute for Quantitative Social Science is planning to use differential privacy techniques to enable more researchers to share, retain control of, and credit for their data contributions as part of the Dataverse Network, a project that guarantees the long-term preservation of critical datasets.
Unlocked Scientific Potential
Dataverse is the largest public general-purpose research data repository in the world. However, the scientific community could access far more datasets that are currently not publicly available, if differential privacy’s promise is fulfilled, according to Gary King, Albert J. Weatherhead III University Professor at Harvard University and Director of the Institute for Quantitative Social Science.
“That’s why we’re so thrilled to be working on this project,” King said. “The social sciences are finally getting to the point in human history where we have enough information to move from studying problems to actually solving them. As we make progress on the privacy problem, we will be able to unlock more and more of the potential of this new information.”
The differential privacy tool Vadhan and his team are developing will allow the inclusion of datasets that were previously withheld because the information was too sensitive and privacy was uncertain.
“Currently, Dataverse is not equipped to handle datasets with privacy concerns associated with them,” Vadhan said. “If a researcher says that a dataset has identifiable personal information, it is not made available for download.”
Differential privacy doesn’t work for every type of research question. Vadhan pointed to regression, machine learning, and social network analysis as areas where there are very promising theoretical results, but challenges remain to making differential privacy work well in practice.
Differential privacy also doesn’t help when you’re looking for identity of a specific individual: as in the case of identifying terrorists or a match for a kidney donor. But that’s the point: each individual should be “hidden” even as they contribute to the greater good of any given study.
“This project could significantly enhance the state-of-the-art in privacy,” said Nina Amla, a program director at NSF who oversees the award. “They take a highly interdisciplinary approach which brings together deep expertise in computer science, social science, statistics, and law.”
The Latest on: Differential privacy
via Google News
The Latest on: Differential privacy
- Homomorphic Encryption Market: Securing Data like Never Beforeon July 15, 2019 at 4:28 am
However, considerable progress in the fields of differential privacy and privacy preservation techniques has been made over the years. New tools have emerged that can deliver encryption similar to ... […]
- Privacy enhancing technology for data analysison June 17, 2019 at 11:39 am
Differential privacy, which addresses privacy in disclosure rather than in computation, adds noise to the data so that when a statistic is released, information about an individual is not revealed. ... […]
- How federated learning could shape the future of AI in a privacy-obsessed worldon June 3, 2019 at 6:16 am
Updates sent from devices can still contain some personal data or tell you about a person, and so differential privacy is used to add gaussian noise to data shared by devices, Google AI researcher ... […]
- An Introduction to Differential Privacyon May 21, 2019 at 1:08 am
Differential privacy can be achieved by adding randomised “noise” to an aggregate query result to protect individual entries without significantly changing the result. Differentially private ... […]
- The next-gen Google Assistant is taking the privacy fight right to Siri—and it just might winon May 13, 2019 at 8:30 pm
To that end, Apple has been anonymizing and encrypting Siri data for years and using differential privacy to analyze data without collecting it. It’s been a line in the sand between Apple and Google ... […]
- As The Feds Blaze A Trail On Differential Privacy, Will Ad Tech Follow?on April 21, 2019 at 9:07 pm
"Data-Driven Thinking" is written by members of the media community and contains fresh ideas on the digital revolution in media. Today’s column is written by Victor Wong, CEO at Thunder Experience ... […]
- WTF is differential privacy?on April 9, 2019 at 9:05 pm
It’s a process used to aggregate data that was pioneered by Microsoft and is now used by Apple, Google and other big tech companies. In a nutshell, a differential privacy algorithm injects random data ... […]
- Introducing TensorFlow Privacy: Learning with Differential Privacy for Training Dataon March 6, 2019 at 8:02 am
To ensure this, and to give strong privacy guarantees when the training data is sensitive, it is possible to use techniques based on the theory of differential privacy. In particular, when training on ... […]
- Privitar Authors Review of Differential Privacy for the Office for National Statistics’ Quality Reviewon December 13, 2018 at 10:49 am
This has prompted researchers over the last fifteen years to explore a stronger, formal method for protecting privacy, known as differential privacy. Differential privacy brings a new approach to ... […]
- Differential Privacyon October 31, 2018 at 5:00 pm
Differential privacy is a formal mathematical framework for quantifying and managing privacy risks. It provides provable privacy protection against a wide range of potential attacks, including those ... […]
via Bing News