Paper 2026/1091

Practical Homomorphic LSTM via Programmable Bootstrapping

Abstract

While deep learning is ubiquitous, centralized pro- cessing exposes sensitive sequential data—such as natural lan- guage—to untrusted servers, forcing an unacceptable privacy- utility trade-off. Fully Homomorphic Encryption (FHE) re- solves this by computing directly on encrypted data. However, standard neural networks ported to FHE suffer from severe latency bottlenecks, particularly because continuous non-linear activations dominate the computational budget. To overcome this, we introduce the Blind Spiking LSTM (BSLSTM), a TFHE-optimized recurrent architecture for privacy-preserving sequential inference. By co-designing the network with the cryptographic framework, we replace expen- sive continuous non-linearities with an efficient multi-threshold programmable bootstrapping paradigm. Evaluated on stan- dard NLP tasks, BSLSTM achieves an inference latency of 5.2 seconds for a 128-token sequence, significantly outperform- ing traditional homomorphic approaches while maintaining competitive accuracy. Operating at an amortized cost of 211 microseconds per bootstrapping operation, our work demon- strates the practical viability of low-latency, fully homomorphic inference for real-world applications.