Paper 2025/424

Matchmaker: Fast Secure Inference across Deployment Scenarios

Abstract

Secure Two-Party Computation (2PC) enables secure inference with cryptographic guarantees that protect the privacy of the model owner and client. However, it adds significant performance overhead. In this work, we make 2PC-based secure inference efficient while considering important deployment scenarios. We observe that the hitherto unconsidered latency of fetching keys from storage significantly impacts performance, as does network speed. We design a Linear Secret Sharing (LSS)-based system $LSS^M$ and a Function Secret Sharing (FSS)-based system $FSS^M$ for secure inference, optimized for small key size and communication, respectively. Notably, our highly-optimized and hardware-aware CPU-based $LSS^M$ outperforms prior GPU-based LSS systems by up to $50\times$. We then show that the best choice between $LSS^M$ and $FSS^M$ depends on the deployment scenario. In fact, under certain deployments, a combination of $LSS^M$ and $FSS^M$ can leverage heterogeneous processing across CPU and GPU. Such protocol-system co-design lets us outperform state-of-the-art secure inference systems by up to $21\times$ (geomean $3.25\times$).