2026 Volume 34 Pages 1124-1136
Homomorphic encryption (HE) is a promising solution for preserving data privacy in computation outsourcing. However, the large size of ciphertexts leads to high memory usage and communication costs. Previous techniques to reduce the data size of ciphertexts include (1) storing a seed for a pseudorandom generator instead of randomly generated elements in a ciphertext, and (2) hybrid homomorphic encryption, such as converting lightweight ciphertexts to HE ciphertexts. However, these approaches either yield high ciphertext expansion ratios or incur high computational costs. In this paper, we propose PACHE (Pattern-based Compression for Homomorphic Encryption). PACHE decomposes a set of original messages into a small number of shared pattern messages and their corresponding scalar values. To reduce the total data size, PACHE encrypts only the pattern messages rather than the entire original dataset by relaxing security guarantees, indistinguishability and privacy-preserving capability, for decompressed ciphertexts. During computation on a third-party server, PACHE performs low-latency decompression by computing the sum of products between the encrypted pattern messages and the unencrypted scalar values. We evaluate PACHE on a convolutional neural network (CNN) inference task. Compared to the seed-based baseline, PACHE reduces memory usage by 90.4% while increasing latency by 57.2% when using five patterns for a set of encrypted convolution filters.