NOIR: Privacy-Preserving Code Generation Based on Open-Source Large Language Models

NOIR is what the authors call the first privacy-preserving code generation framework that prevents the cloud from obtaining client prompts and generated code: the client sends prompt embedding vectors to the cloud LLM, decodes and generates code locally, and uses token-embedding-level local differential privacy and a data-independent random tokenizer to resist reconstruction and frequency analysis attacks from the cloud. On open-source large language models, under strong privacy, NOIR achieves Evalplus (MBPP, HumanEval) Pass@1 of 76.7 and 77.4, and BigCodeBench Pass@1 of 38.7 (only 1.77% lower than the original LLM).