UC San Diego

In today's data-driven world, we are conflicted with two opposing phenomena. On the one hand, collection and analysis of an enormous amount of data have resulted in rapid advances in technologies and services, especially the ones based on Artificial Intelligence (AI). On the other hand, existing and potential dangers of data misuse have created serious concern about data privacy. Privacy-preserving computation presents powerful cryptographic tools to tackle this conflict by enabling analysis on data with assurance of provable privacy guarantee. However, this capability comes with significant computation and communication overhead deterring its adoption in practical data-intensive applications. Moreover, understanding the details of the cryptographic methods often appears to be a daunting task for application developers. This dissertation contributes towards enabling data-intensive systems with provable privacy guarantee in realistic settings. Our work addresses the challenge of practical privacy-preserving computation from three directions. First, we develop open-source frameworks with efficient and scalable execution of privacy-preserving protocols as well as a rich programming interface to abstract the details of protocol execution from the users. Second, we speed up the computations required for the protocols through custom-designed hardware platforms. Our designs include both generic and application-specific accelerators achieving a minimum of 110× improvement in throughput-per-core over the best prior art. Third, we devise several practical privacy-preserving applications including secure localization, authentication with noisy keys, and k-nearest neighbor search on private data. Our most exciting application is a mixed protocol system for privacy-preserving AI with 4.7×–14.4× speed up over state-of-the-art.