Cuot: Accelerating Oblivious Transfer on Gpus for Privacy-Preserving Computation

Privacy and security requirements are gaining importance rapidly as new applications emerge in several domains that work with sensitive user information. Various Secure Multiparty Computation (MPC) techniques in cryptography allow working with data in a privacy-preserving manner. A core building block in many popular MPC protocols is the Oblivious Transfer (OT). The main challenge in deploying MPC protocols is the significant overhead they incur compared to computing over plaintext data. Given the fundamental role of OT in constructing MPC protocols, efficiency improvements in generation of OTs would directly translate into improvements for a wide range of MPC protocols. This paper proposes cuOT, a novel framework for accelerating OT on the GPU based on a CUDA implementation. cuOT implements OT variants based on recent silent OT extension constructions which significantly reduce communication at the cost of higher computation. cuOT implements core building blocks of silent OT extension, incorporating several GPU specific optimizations to improve parallelization and reduce memory transfer overheads. Our evaluations show that cuOT is able to obtain up to (75×) and (15×) speedup in generating millions of OTs compared to CPU counterparts.