Paper 2025/972

Generalized BGV, BFV, and CKKS for Homomorphic Encryption over Matrix Rings

Abstract

Some of the most valuable applications of homomorphic encryption, such as encrypted machine learning inference, require efficient large-scale plaintext-ciphertext and ciphertext-ciphertext matrix multiplications. Current state-of-the-art techniques for matrix multiplications all build on the ability to pack many ciphertexts into a ciphertext and compute on them in a Single Instruction, Multiple Data (SIMD) manner. However, to fit the operation of matrix multiplication into this computational model, a large number of additional costly operations need to be performed, such as the rotation of elements between the plaintext slots. In this work, we propose an orthogonal approach to performing encrypted matrix operations with BGV-like encryption schemes, where the plaintext and ciphertext spaces are generalized to a matrix ring of arbitrary dimension. To deal with the inherent problem of noncommutativity in the case of matrix rings, we present a new superoperator technique to better represent linear and quadratic expressions in the secret key, which allows for the relinearization of ciphertexts after multiplication. The security of the modified encryption schemes is based on Module-LWE with module rank equal to the dimension of the matrices. With this construction, we demonstrate that Ring-LWE, Module-LWE, and LWE are potentially equally efficient for homomorphic encryption, both in terms of useful information density and noise growth, only for different sizes of matrices.