Optimizing TLB for access pattern privacy protection in data outsourcing

Abstract

Oblivious RAM (ORAM) is a protocol to hide access pattern to an untrusted storage. ORAM prevents a curious adversary identifying what data address the user is accessing through observing the bits flows between the user and the untrusted storage system. Basically, ORAM protocols store user’s data in shuffled form on the untrusted storage and substitute the original access with multiple access to random addresses to cover the real target. Such redundancy introduce significant performance overhead. Traditional Translation Lookaside Buffer (TLB) exploits temporal locality hide memory latency in DRAM systems. However, the ORAM locality is totally different and thus traditional TLB eviction strategy have a poor performance. In this paper, we propose O-TLB which exploits ORAM temporal locality and optimized TLB eviction strategy to reduce server-side memory I/O operations. Intuitively, exploiting locality for performance may expose this locality which breaks obliviousness. We challenge this intuition by exploiting locality based on server-side ORAM data structures. Unlike previous works, our approach do not sacrifice any provable security. Specifically, previous optimization works leaks access pattern through timing channel and do no fit with adaptive asynchronous obliviousness (AAOB) in a multiple users scenario. While in our method, the timing do not vary with locality of program and O-TLB optimization can be adopted directly keeping AAOB. Our simulation result show that with O-TLB scheme, the underlying ORAM server-side I/O performance is improved by 11%.