Practical MPC+FHE with Applications in Secure Multi-PartyNeural Network Evaluation

The theoretical idea of using FHE to realize MPC has been there for over a decade. Existing threshold (and multi-key) FHE schemes were constructed by modifying and analyzing a traditional single-key FHE in a case-by-case manner, thus technically highly-demanding. This work explores a new approach to build threshold FHE (thereby MPC schemes) through tailoring generic MPC protocols to the base FHE scheme while requiring no effort in FHE redesign. We applied our approach to two representative Ring-LWE-based FHE schemes: CKKS [37, 38] and GHS [54, 55], producing GMPFHE-CKKS and GMPFHE-GHS. We developed MPC protocols based on GMPFHECKKS and GMPFHE-GHS which are secure against any number of passive but colluding adversaries. The online cost of our MPC protocol is O(|C |), as opposed to O(|C | · n2) for existing MPC protocols, and our offline cost is independent of |C |. We experimentally show that the GMPFHE-CKKS-based MPC protocol offers unparalleled amortized performance on multi-party neural network evaluation.