Paper 2026/1279

BootNet: Homomorphic CNN Inference with Convolution and ReLU Fused in Bootstrapping

Abstract

Fully homomorphic encryption (FHE) enables privacy-preserving neural network inference but suffers from high overhead from homomorphic convolutions, polynomial activation approximations, and CKKS bootstrapping. This paper presents BootNet, a unified framework that fuses all three operations into a single bootstrapping invocation per CNN layer, achieving convolution, ReLU, and noise refresh simultaneously. Prior works are able to fuse convolution into bootstrapping using CinS encoding (NeuJeans, CCS 2024) or ReLU into functional bootstrapping (RBOOT, USENIX Security 2026), but combining both for end-to-end ImageNet inference introduces numerous undocumented challenges. BootNet resolves these through a suite of co‑design techniques and optimizations, including four representative solutions: (1) a split‑and‑merge bootstrapping schedule that halves bootstrapping calls for ResNet shortcuts; (2) an improved RBOOT configuration paired with a model quantization method that complete the activation function with fewer multiplication depth; (3) a fusion of NeuJeans' masking layer with RBOOT's arcsin step that yields additional depth reduction; (4) tailoring EvalRound+ (IEEE Access 2025) to fit slim bootstrapping and RBOOT for further depth reduction. We evaluate BootNet on multiple ResNet models on the ImageNet dataset. Compared to the state-of-the-art Orion (ASPLOS 2025), BootNet reduces end-to-end latency by 67-73% and storage by 76-90% while preserving plaintext accuracy. We also introduce BootNet-mini, enabling full ImageNet inference at ring degree $N=2^{15}$ (with over $120$-bit security via sparse-secret encapsulation) for the first time, while all previous works have to use $N=2^{16}$ or larger. BootNet-mini performs similarly in latency but reduces 30-55% storage requirement than BootNet, making it potentially more friendly for hardware acceleration.