Data Oblivious Genome Variants Search On Intel Sgx

We show how to build a practical, private data oblivious genome variants search using Intel SGX. More precisely, we consider the problem posed in Track 2 of the iDash Privacy and Security Workshop 2017 competition, which was to search for variants with high. 2 statistic among certain genetic data over two populations. The winning solution of this iDash competition (developed by Carpov and Tortech) is extremely efficient, but not memory oblivious, which potentially made it vulnerable to a whole host of memory- and cache-based side channel attacks on SGX. In this paper, we adapt a framework in which we can exactly quantify this leakage. We provide a memory oblivious implementation with reasonable information leakage at the cost of some efficiency. Our solution is roughly an order of magnitude slower than the non-memory oblivious implementation, but still practical and much more efficient than naive memory- oblivious solutions-it solves the iDash problem in approximately 5min. In order to do this, we develop novel definitions and models for oblivious dictionary merging, which may be of independent theoretical interest.