Paper 2026/1502 Original: eprint.iacr.org/2026/1502 ↗

Efficient Privacy-Preserving LSTM Inference on Encrypted Sequential Data

Abstract

Recent advances in fully homomorphic encryption (FHE) have enabled privacy-preserving machine learning directly over encrypted data. As a representative recurrent architecture, the long short-term memory (LSTM) network is widely used for modeling sequential dependencies, yet existing FHE-based LSTM inference schemes still suffer from high latency and limited scalability. In this paper, we present an efficient privacy-preserving LSTM inference protocol on encrypted sequential data based on FHE. We insert a lightweight normalization module before nonlinear activations to bound the hidden states, thereby enabling accurate low-degree polynomial approximations of the Sigmoid, Tanh, and inverse square-root functions via a hybrid Remez and least-squares strategy. We implement the proposed protocol using the Lattigo library, incorporating ciphertext packing optimization, rotation minimization, and SIMD-based parallelization. Experiments show that on standard text classification benchmarks, our encrypted LSTM achieves competitive accuracy compared with plaintext models and consistently outperforms the state-of-the-art FHE-based method, achieving up to a 4.7x speedup.