GPU and CPU Parallelization of Honest-but-Curious Secure Two-Party Computation

Recent work demonstrates the feasibility and practical use of secure two-party computation. A major enabler of this feasibility has been the exploitation of the embarrassingly parallel nature of many tasks that are part of a secure computation protocol. Nonetheless, there are still various forms and levels of parallelization that may yet improve the performance of these protocols. For example, implementations on the SIMD architecture of modern GPGPUs requires significantly different approaches to implementation than the general purpose MIMD architecture of multi-core CPUs, which again differ from the needs of parallelizing on compute clusters.

In this work, we present the design and analysis of a general-purpose, Graphical Processing Unit (GPU)-optimized implementation of Yao's garbled-circuit protocol for two-party secure computation. We provide security where a malicious adversary can learn one arbitrary bit of the other user's private input, as proposed by Franklin and Mohassel, based on the protocol of Huang et. al. We implement nearly all of the modern protocol advancements, such as Free XOR, Pipelining, and OT extension. Our implementation is the first allowing entire circuits to be generated in parallel concurrently, and makes use of a modification of the XOR technique so that circuit generation is optimized for implementation on SIMD architectures of GPUs. We implement and compare the performance of the circuit evaluation step in Yao's protocol on both GPU and on multi-core CPUs. Finally, our system is the first to support multiple GPUs within a single machine and give initial evidence that the approach is feasible.

Author(s):

Nathaniel Husted    
Indiana University
United States

Steve Myers    
Indiana University
United States

abhi shelat    
University of Virginia
United States

Paul Grubbs    
Indiana University
United States