Secure Multi-Party Computation for Machine Learning: A Survey
Ian Zhou, Farzad Tofigh, Massimo Piccardi, Mehran Abolhasan, Daniel Franklin, Justin Lipman
Abstract
Machine learning is a powerful technology for extracting information from data of diverse nature and origin. As its deployment increasingly depends on data from multiple entities, ensuring privacy for these contributors becomes paramount for the integrity and fairness of machine learning endeavors. This review looks into the recent advancements in secure multi-party computation (SMPC) for machine learning, a pivotal technology championing data privacy. We evaluate these applications from various aspects, including security models, requirements, system types, and service models, aligning with the IEEE’s recommended practices for SMPC. Broadly, SMPC systems are divided into two categories: homomorphic-based systems, which facilitate computations on encrypted data, ensuring data remains confidential, and secret sharing-based systems, which disseminate data across parties in fragmented shares. Our literature analysis highlights certain gaps, such as security requisites, streamlined information exchange, incentive structures, data authenticity, and operational efficiency. Recognizing these challenges lead to envisioning a holistic SMPC protocol tailored for machine learning applications.