Paper 2026/1255

Bootstrapping is All You Need: Secure Transformer Inference via Improved CKKS Functional Bootstrapping

Abstract

Fully homomorphic encryption (FHE) enables non-interactive secure transformer inference (NISTI). Due to the high cost of bootstrapping, conventional approaches typically choose parameters that support a large multiplicative depth to reduce bootstrapping frequency. However, larger depth directly increases ciphertext size, resulting in higher communication and computation overheads. In this paper, we introduce a novel functional bootstrapping (FBS) scheme that fundamentally reshapes the computation paradigm for NISTI: by fusing as many operations as possible into each bootstrapping operation, our approach significantly reduces the prescribed multiplicative depth. Our FBS achieves a trigonometric minimax approximation for the target function, making it well suited for precision-sensitive components such as transformer nonlinear layers. Furthermore, we incorporate linear layers into the slot-to-coefficient (S2C) transformation within FBS, thereby eliminating the need to evaluate them separately. Building on these innovations, we present a complete NISTI framework that achieves a 1.9$\times$ speedup in runtime (from 662.3s to 349.5s) and a 3$\times$ reduction in communication (from 48.3MB to 16.1MB) compared with the state-of-the-art.