Scalable Multi-Party Computation Protocols for Machine Learning in the Honest-Majority Setting

In this paper, we present a novel and scalable multi-party computation (MPC) protocol tailored for privacy-preserving machine learning (PPML) with semi-honest security in the honest-majority setting. Our protocol utilizes the Damgård-Nielsen (Crypto'07) protocol with Mersenne prime fields. By leveraging the special properties of Mersenne primes, we are able to design highly efficient protocols for securely computing operations such as truncation and comparison. Additionally, we extend the two-layer multiplication protocol in ATLAS (Crypto'21) to further reduce the round complexity of operations commonly used in neural networks. Our protocol is very scalable in terms of the number of parties involved. For instance, our protocol completes the online oblivious inference of a 4-layer convolutional neural network with 63 parties in 0.1 seconds and 4.6 seconds in the LAN and WAN settings, respectively. To the best of our knowledge, this is the first fully implemented protocol in the field of PPML that can successfully run with such a large number of parties. Notably, even in the three-party case, the online phase of our protocol is more than 1.4× faster than the Falcon (PETS'21) protocol.