Techniques for performing secure computation on encrypted data: Chris Fletcher, MIT
- 10:00 am on Monday, December 10, 2012
- 12:00 pm on Monday, December 10, 2012
- MCS 137
Abstract: Privacy of data is a huge problem in cloud computing, and more generally in outsourcing computation. From financial information to medical records, sensitive data is stored and computed upon in the cloud. Computation requires the data to be exposed to the cloud servers, which may be attacked by malicious applications, hypervisors, operating systems or insiders. In the ideal scenario, no one other than the user sees the private data in decrypted form, as is achieved through the use of fully homomorphic encryption (FHE) techniques. The first part of the talk will focus on (a) techniques to run general purpose programs under FHE and (b) how some programs are naturally better suited for FHE than others. I will talk about the how ambiguity in program control flow and data structures leads to large overheads for certain programs, in addition to the crypto overheads already imposed by FHE (which impose about a billion times slowdown). Motivated by large FHE overheads, the second part of the talk describes how to approximate FHE with a tamper-resistant processor called Ascend. Ascend performs program obfuscation in hardware: given an untrusted program and private user data running within the Ascend chip, the chip's external input/output and power pins give off a signal that is independent of the private user data. I will discuss how strict periodic accesses to an Oblivious RAM obfuscate input/output behavior and how strict periodic accesses to on-chip circuits (e.g., on-chip caches) coupled with DPA-resistant techniques obfuscate Ascend's power signature. Surprisingly, Ascend incurs only a ~5X performance overhead running SPEC benchmarks. The trusted computing base is only the Ascend chip: no software (the user application, server operating system, etc) or anything outside the Ascend processor (external RAM or communication channels) is trusted. This is joint work with Marten van Dijk, Srini Devadas, Ling Ren and Xiangyao Yu.