CAREER: Scalable Oblivious-Private Computation with Applications to Databases and Systems

Sensitive information is increasingly stored in the cloud, from medical records and financial data to private messages. Encryption alone is not enough to protect it: even when data is encrypted, access patterns can quietly reveal the underlying secrets to a watching attacker. Specialized secure hardware now built into commercial computer chips offers an affordable way to protect data while it is in use, but it too can be defeated by attackers who observe these access patterns. This project develops the building blocks needed to close that gap, making it practical to compute on sensitive data at very large scale without revealing anything about its contents. Success will bring strong privacy guarantees to everyday services that society relies upon, including secure messaging, privacy-preserving artificial intelligence, confidential cloud databases, and collaborations in which hospitals or financial institutions pool information to detect rare diseases or fraud without exposing any individual's data. By making strong privacy practical at scale for the systems society increasingly depends on, these advances serve common interests in security and privacy. The main goal of this project is to make oblivious computation, which hides data access patterns from powerful adversaries, practical on realistic workloads of tens to hundreds of terabytes and beyond, a scale that prior approaches could not reach. The work combines oblivious algorithms with trusted execution environments to defend simultaneously against both access-pattern and software side-channel attacks, and is organized into three research thrusts. The first designs faster, more scalable oblivious primitives for core operations such as sorting, shuffling, and compaction, using new models of parallelism that balance computation against communication across many machines and that exploit emerging hardware acceleration. The second builds on these primitives to construct scalable oblivious memory and data structures, including high-throughput, low-latency key-value stores with richer functionality and improved input-output efficiency. The third integrates these results into scalable oblivious database systems, contributing new distributed operators for filtering, aggregation, and complex multi-way joins, together with a query optimizer that reveals no information about intermediate results. All methods will be formally proven secure, released as open-source prototypes on commercial secure hardware, and evaluated against the best existing approaches using standard database benchmarks. The project also integrates these advances into new and revised courses, research-driven capstone projects, a credit-based security colloquium, and annual regional workshops, creating a virtuous cycle of discovery and instruction. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.