You are here

Ioannidis and Leeser Awarded $500K NSF MaSSIF Grant

July 24, 2017

Assistant Professor Stratis Ioannidis and Interim Chair and Professor Miriam Leeser of the Department of Electrical and Computer Engineering (ECE) have been awarded a $500,000 grant from the National Science Foundation to develop a “Massively Scalable Secure Computation Infrastructure Using FPGAs” to advance large-scale and secure data mining.

Every day, the collection of behavioral data takes place on a massive scale, from clinical trials performed by medical and scientific experts, to shopping conduct and economic trends by retail sales corporations. While mining these data can offer tremendous societal benefits resulting from sharing data among researchers and practitioners, it also gives rise to significant personal privacy concerns—an Engineering Grand Challenge facing society today. Ioannidis and Leeser’s research seeks to not only guarantee the safety of the personal data collected, but also to speed up the process by which it is encrypted.

“There are solutions to safeguard personal information, including behavioral data, through a technique called Secure Function Evaluation (SFE); the end user can access the output of a computation performed on the data, but not the original records,” says Ioannidis. “The problem with SFE is that it is computationally intensive; computations performed securely can be up to 700,000 times slower, which is prohibitive for large datasets.”

Leeser and Ioannidis propose a solution to this slow-down in the form of field-programmable gate arrays, or FPGAs, which are specialized hardware that can accelerate the SFE application. In 2016, the team were awarded a $67,000 Google Faculty Research Award to deploy and evaluate SFE algorithms over FPGAs, enabling companies that operate large data centers, like Google, to perform data mining operations efficiently while offering strong privacy guarantees.

PhD student Xin Fang has been working on the seed project stemming from this Google award, and his work has shown that the speed-up of secure data mining using FPGAs is possible on a small scale. With this new NSF grant, Ioannidis and Leeser will further their research in this area with potentially transformative results.

“By using FPGAs, we can work to apply SFE over large and changing data sets using parallel computing practically in real-time, while still preserving personal privacy,” says Leeser. “This has world-wide applications throughout industries, as well as toward the whole notion of the importance of protecting one’s data, from medical records to geolocation. If there is no chance that an individual’s personal data can be compromised, then we as a society are much more likely to become comfortable with prudent data collection. This can work for the betterment of everything from homeland security to the curing of diseases.”

Abstract Source: NSF

The statistical analysis of behavioral data collected through clinical trials, surveys, and experimentation, has a long history in academic disciplines like medicine, sociology, and behavioral economics. The privacy risks inherent in such studies are often at odds with the tremendous societal benefits resulting from sharing data among researchers and practitioners. Mining behavioral data at scale is also a ubiquitous practice among Internet companies, giving rise to significant privacy concerns. As the potential benefits to society are enormous, harnessing this data for the better good while protecting privacy is one of the grand challenges faced by our society today. This project addresses this challenge by bringing Secure Function Evaluation (SFE) of practical, real-life data mining and machine learning algorithms into the realm of practicality, through the development of a highly parallel, efficient, scalable computation platform for secure computation operating at a massive scale.

The project develops a Massively Scalable Secure computation Infrastructure using FPGAs (MaSSIF), accelerating secure computations over a cluster of FPGAs and leveraging benefits of both hardware acceleration and multi-device parallelism. MaSSIF significantly differs from previous implementations of SFE in that it is the first to accelerate secure computation primitives: specifically, Garbled Circuits (GC) with FPGAs on such a massively parallel scale. The algorithms considered are (a) computationally intensive, (b) non-trivial to parallelize under SFE, and (c) of considerable practical importance. MaSSIF advances the state of the art both through novel SFE algorithms, as well as in the design and optimization of accelerated, scalable systems for SFE.