@article{M4CF66800,
title = "Accelerating ResNet Inference on RNS-CKKS with Merged Bootstrapping",
journal = "The Journal of Korean Institute of Communications and Information Sciences",
year = "2025",
issn = "1226-4717",
doi = "10.7840/kics.2025.50.11.1669",
author = "Soomin Park, Eunsang Lee",
keywords = "Bootstrapping, Convolutional neural networks, homomorphic encryption, ResNet",
abstract = "Recent studies on convolutional neural networks based on fully homomorphic encryption(FHE) have gained momentum, and there are reports of implementing ResNet with high accuracy on the representative FHE residue number system variant of Cheon–Kim–Kim–Song(RNS-CKKS). However, existing approaches under‑utilize ciphertext slots, requiring frequent bootstrapping and therefore incurring substantial computational overhead. In response to this, our paper proposes a 'merged bootstrapping' method that, when performing residual network(ResNet) on multiple images simultaneously, merges multiple ciphertexts into a single ciphertext before bootstrapping. This effectively utilizes all slots of the ciphertext, significantly reducing computation time.
Experimental results using the RNS-CKKS scheme library, Lattigo, confirmed that when classifying 2, 4, and 8 Canadian Institute for Advanced Research-10(CIFAR-10) images with ResNet-20 using merged bootstrapping,"
}