Force: Highly Efficient Four-Party Privacy-Preserving Machine Learning on GPU

Tremendous efforts have been made to improve the efficiency of secure Multi-Party Computation (MPC), which allows $n$ $\geq$ 2 parties to jointly evaluate a target function without leaking their own private inputs. It has been confirmed by previous research that Three-Party Computation (3PC) and outsourcing computations to GPUs can lead to huge performance improvement of MPC in computationally intensive tasks such as Privacy-Preserving Machine Learning (PPML). A natural question to ask is whether super-linear performance gain is possible for a linear increase in resources. In this paper, we give an affirmative answer to this question. We propose Force, an extremely efficient Four-Party Computation (4PC) system for PPML. To the best of our knowledge, each party in Force enjoys the least number of local computations, smallest graphic memory consumption and lowest data exchanges between parties. This is achieved by introducing a new sharing type $X$-share along with MPC protocols in privacy-preserving training and inference that are semi-honest secure in the honest-majority setting. By comparing the results with state-of-the-art research, we showcase that Force is sound and extremely efficient, as it can improve the PPML performance by a factor of 2 to 38 compared with other latest GPU-based semi-honest secure systems, such as Piranha (including SecureML, Falcon, FantasticFour), CryptGPU and CrypTen.