Oblivious Query Processing
Motivated by cloud security concerns, there is an increasing interest in database systems that can store and support queries over encrypted data. A common architecture for such systems is to use a trusted component such as a cryptographic co-processor for query processing that is used to securely decrypt data and perform computations in plaintext. The trusted component has limited memory, so most of the (input and intermediate) data is kept encrypted in an untrusted storage and moved to the trusted component on ``demand.'' In this setting, even with strong encryption, the data access pattern from untrusted storage has the potential to reveal sensitive information; indeed, all existing systems that use a trusted component for query processing over encrypted data have this vulnerability. In this paper, we undertake the first formal study of secure query processing, where an adversary having full knowledge of the query (text) and observing the query execution learns nothing about the underlying database other than the result size of the query on the database. We introduce a simpler notion, oblivious query processing, and show formally that a query admits secure query processing iff it admits oblivious query processing. We present oblivious query processing algorithms for a rich class of database queries involving selections, joins, grouping and aggregation. For queries not handled by our algorithms, we provide some initial evidence that designing oblivious (and therefore secure) algorithms would be hard via reductions from two simple, well-studied problems that are generally believed to be hard. Our study of oblivious query processing also reveals interesting connections to database join theory.