3/7/2017 Emily Scott
Professor Sewoong Oh was recently awarded a Google Faculty Research Award for his work in data privacy.
Written by Emily Scott
Professor Sewoong Oh was recently awarded a Google Faculty Research Award for his work in data privacy.
Google Faculty Research Awards are granted to faculty pursuing research in areas of mutual interest.
Oh’s work revolves around the need for data privacy. In today’s world, social data can be helpful in the decision-making processes of individuals and businesses wanting to extract information about people’s activities and opinions.
These data analyses can end up benefitting people, but it could mean they are revealing sensitive information — a critical breach of privacy.
As a result, there is an increasing tension between the need to share data for improved analyses in data analysis technologies, and the need to protect an individual’s privacy.
The need for data privacy appears in two separate contexts. The first, referred to as “local privacy,” involves data collection: people disclosing their personal information voluntarily, such as through social networking sites.
The second, referred to as “global privacy,” involves data release: institutions answering questions on information databases, or releasing these databases — for example, the U.S. government releasing census data, or Netflix releasing data so others can test state-of-the-art machine learning.
Differential privacy has been proposed as a formal mathematical notion of privacy that provides strong guarantees against arbitrary adversaries.
“Using local differential privacy as the measure of privacy, this proposal addresses the privacy aspects of social data processing, which is a fundamental topic of interest,” Oh said.
Oh explained that the fundamental question in designing privatization mechanisms is how to maximize the utility of privatized data.
“Typically, it is computationally challenging due to the non-convex nature of the problem,” he said.
However, computationally efficient methods to find the exact optimal mechanisms in certain contexts exist — for example, statistical analysis.
“The key idea is to exploit the geometric structures, which we call staircase mechanisms, of optimal mechanisms,” Oh said. “We propose to investigate this framework in two specific contexts: private multi-party computation and incentive designing mechanisms for sharing private data of strategic users.”