Ruyi: A Configurable and Efficient Secure Multi-Party Learning Framework With Privileged Parties

Abstract

Secure multi-party learning (MPL) enables multiple parties to train machine learning models with privacy preservation. MPL frameworks typically follow the peer-to-peer architecture, where each party has the same chance to handle the results. However, the cooperative parties in business scenarios usually have unequal statuses. Thus, Song et al. (CCS’22) presented pMPL , a hierarchical MPL framework with a privileged party. Nonetheless, pMPL has two limitations: (i) it has limited configurability requiring manually finding a public matrix that satisfies four constraints, which is difficult when the number of parties increases, and (ii) it is inefficient due to the huge online communication overhead. In this paper, we are motivated to propose Ruyi , a configurable and efficient MPL framework with privileged parties. Firstly, we reduce the public matrix constraints from four to two while ensuring the same privileged guarantees by extending the standard resharing paradigm to vector space secret sharing in order to implement the share conversion protocol and performing all the computations over a prime field rather than a ring. This enhances the configurability so that the Vandermonde matrix can always satisfy the public matrix constraints when given the number of parties, including privileged parties, assistant parties, and assistant parties allowed to drop out. Secondly, we reduce the online communication overhead by adapting the masked evaluation paradigm to vector space secret sharing. Experimental results demonstrate that Ruyi is configurable with multiple parties and outperforms pMPL by up to $ 53.87 \times $ , $13.91 \times $ , and $2.76 \times $ for linear regression, logistic regression, and neural networks, respectively.