Paper 2024/1730

Secure and Efficient Outsourced Matrix Multiplication with Homomorphic Encryption

Abstract

Fully Homomorphic Encryption (FHE) is a promising privacy-enhancing technique that enables secure and private data processing on untrusted servers, such as privacy-preserving neural network (NN) evaluations. However, its practical application presents significant challenges. Limitations in how data is stored within homomorphic ciphertexts and restrictions on the types of operations that can be performed create computational bottlenecks. As a result, a growing body of research focuses on optimizing existing evaluation techniques for efficient execution in the homomorphic domain. One key operation in this space is matrix multiplication, which forms the foundation of most neural networks. Several studies have even proposed new FHE schemes specifically to accelerate this operation. The optimization of matrix multiplication is also the primary goal of our work. We leverage the Single Instruction Multiple Data (SIMD) capabilities of FHE to increase data packing and significantly reduce the KeySwitch operation count— an expensive low-level routine in homomorphic encryption. By minimizing KeySwitching, we surpass current state-of-the-art solutions, requiring only a minimal multiplicative depth of two. The best-known complexity for matrix multiplication at this depth is $\mathcal{O}(d)$ for matrices of size $d\times d$. Remarkably, even the leading techniques that require a multiplicative depth of three still incur a KeySwitch complexity of $\mathcal{O}(d)$. In contrast, our method reduces this complexity to $\mathcal{O}(\log{d})$ while maintaining the same level of data packing. Our solution broadly applies to all FHE schemes supporting Single Instruction Multiple Data (SIMD) operations. We further generalize the technique in two directions: allowing arbitrary packing availability and extending it to rectangular matrices. This versatile approach offers significant improvements in matrix multiplication performance and enables faster evaluation of privacy-preserving neural network applications.