Piranha: A GPU Platform for Secure Computation

Secure multi-party computation (MPC) is an essential tool for privacy-preserving machine learning (ML). However, secure training of large-scale ML models currently requires a prohibitively long time to complete. Given that large ML inference and training tasks in the plaintext setting are significantly accelerated by Graphical Processing Units (GPUs), this raises the natural question: can secure MPC leverage GPU acceleration? A few recent works have studied this question in the context of accelerating specific components or protocols, but do not provide a general-purpose solution. Consequently, MPC developers must be both experts in cryptographic protocol design and proficient at low-level GPU kernel development to achieve good performance on any new protocol implementation. We present Piranha, a general-purpose, modular platform for accelerating secret sharing-based MPC protocols using GPUs. Piranha allows the MPC community to easily leverage the benefits of a GPU without requiring GPU expertise. Piranha contributes a three-layer architecture: (1) a device layer that can independently accelerate secret-sharing protocols by providing integer-based kernels absent in current general-purpose GPU libraries, (2) a modular protocol layer that allows developers to maximize utility of limited GPU memory with in-place computation and iterator-based support for non-standard memory access patterns, and (3) an application layer that allows applications to remain completely agnostic to the underlying protocols they use. To demonstrate the benefits of Piranha, we implement 3 state-of-the-art linear secret sharing MPC protocols for secure NN training: 2-party SecureML (IEEE S&P '17), 3-party Falcon (PETS '21), and 4-party FantasticFour (USENIX Security '21). Compared to their CPU-based implementations, the same protocols implemented on top of Piranha's protocol-agnostic acceleration exhibit a 16-48× decrease in training time. For the first time, Piranha demonstrates the feasibility of training a realistic neural network (e.g. VGG), end-to-end, using MPC in a little over one day. Piranha is open source and available at https://github.com/ucbrise/piranha . Effectively and easily interfacing with the GPU is a major barrier to MPC developers who wish to accelerate their protocols, but lack experience in programming optimized GPU kernels. Thus, a flexible abstraction is needed to support a wide array of MPC protocols while minimizing any domain-specific knowledge required. In this section, we discuss how Piranha addresses two primary challenges in providing extensible GPU support for MPC protocols: managing vectorized GPU data and supporting acceleration for integer-based computation.