SecPE

This repository is the implementation of paper: "SecPE: Secure Prompt Ensembling for Private and Robust Large Language Models (2024 ECAI)".
SecPE is the first secure prompt ensembling framework for private and robust LLM inference.

Run SecPE

Step1: Complie SEAL

We use Microsoft SEAL library version 3.6.6 for RNS-CKKS homomorphic encryption scheme. Since SEAL does not support boostrapping, we rely on the boostrapping implementation of FHE-MP-CNN.

You should build and install the modified SEAL library in SEAL-3.6-bs/

cd SEAL-3.6-bs
cmake -S . -B build
cmake --build build
cmake --install build

Step2: Complie SecPE

cmake .
make

This should produce a binary file bin/main.