Paper 2025/1679

SoK: Connecting the Dots in Privacy-Preserving ML - Systematization of MPC Protocols and Conversions Between Secret Sharing Schemes

Abstract

Privacy-preserving machine learning (PPML) has become increasingly important due to the need to protect sensitive data during training and inference. Secure multiparty computation (MPC) and homomorphic encryption (HE) have emerged as foundational technologies, enabling secure computation over private data. In this work, we provide a systematic comparative overview of MPC frameworks for PPML, focusing on protocols that introduce novel approaches rather than incremental improvements. Frameworks are analyzed based on computational and communication complexity, throughput, security guarantees, and applicability in small-party settings. Each underlying primitive in PPML is examined from an MPC perspective, highlighting its role and trade-offs. We also emphasize the diversity of secret-sharing schemes and associated interoperability challenges, proposing scheme conversions to facilitate efficient hybrid solutions. This Systematization of Knowledge guides researchers in identifying open problems and practitioners in selecting effective MPC-based frameworks for real-world PPML deployment.

Note: While we have made every effort to include all relevant MPC-based PPML works, given the rapid developments in this area, some may have been inadvertently missed. We welcome corrections or additions.