Towards Practical Oblivious Join Processing.

In cloud computing, remote accesses over the cloud data inevitably bring the issue of trust. Despite strong encryption schemes, adversaries can still learn sensitive information from encrypted data by observing data access patterns. Oblivious RAMs (ORAMs) are proposed to protect against access pattern attacks. However, directly deploying ORAM constructions in an encrypted database brings large computational overhead. In this work, we focus on oblivious joins over a cloud database. Existing studies in the literature are restricted to either primary-foreign key joins or binary equi-joins. Our major contribution is to support general band joins and multiway equi-joins. For oblivious join without ORAMs, we extend the existing binary equi-join algorithm to support general band joins obliviously. For oblivious join with ORAMs, we integrate $B$ -tree indices into ORAMs for each input table and retrieve blocks through the indices in join processing. The key point is to avoid retrieving tuples that make no contribution to the final join result and bound the number of accesses to each $B$ -tree index. The effectiveness and efficiency of our algorithms are demonstrated through extensive evaluations over real-world datasets. Our method shows orders of magnitude speedup for oblivious multiway equi-joins in comparison with baseline algorithms.