International Association
for Cryptologic Research

Machine learning is widely used for a range of applications and is increasingly offered as a service by major technology companies. However, the required massive data collection raises privacy concerns during both training and inference. Privacy-preserving machine learning aims to solve this problem. In this setting, a collection of servers secret share their data and use secure multi-party computation to train and evaluate models on the joint data. All prior work focused on the scenario where the number of servers is two or three. In this work, we study the problem where there are $N \geq 3$ servers amongst whom the data is secret shared.

A key component of machine learning algorithms is to perform fixed-point multiplication with truncation of secret shared decimal values. In this work, we design new protocols for multi-party secure fixed-point multiplication where each of the $N$ parties have one share each of the two values to be multiplied and receive one share of the product at the end of the protocol. We consider three forms of secret sharing - replicated, Shamir, and additive, and design an efficient protocol secure in the presence of a semi-honest adversary for each of the forms. Our protocols are more communication efficient than all prior work on performing multi-party fixed-point multiplication. Additionally, for replicated secret sharing, we design another efficient protocol that is secure in the presence of a malicious adversary. Finally, we leverage our fixed-point multiplication protocols to design secure multi-party computation (MPC) protocols for arbitrary arithmetic circuits that have addition and fixed-point multiplication with truncation gates. All our protocols are proven secure using a standard simulation based security definition. Our protocols for replicated and Shamir sharing work in the presence of an honest majority of parties while the one for additive sharing can tolerate a dishonest majority as well.