Reproduce “InferDPT: Privacy-Preserving Inference for Closed-box Large Language Model” on OpenResearch

About this paper

Meng Tong, Kejiang Chen, Jie Zhang, Yuang Qi, Weiming Zhang, Nenghai Yu, Tianwei Zhang, Zhikun Zhang

InferDPT is a framework enabling privacy-preserving inference for closed-box Large Language Models during text generation, which leverages a novel differentially private mechanism called RANTEXT. The framework provides robust privacy protection against various inference and inversion attacks while maintaining text generation quality comparable to non-private LLM usage.

Key insights

• Despite perturbation, generated text from remote LLMs retains sufficient information from the original prompt, allowing a local model to effectively extract and reconstruct high-quality output.
• Dynamically generating token adjacency lists for Local Differential Privacy (LDP) mechanisms offers a better balance between protecting sensitive information and preserving the semantic utility of the text compared to static lists.
• A two-module framework, combining LDP-based perturbation and local model-based extraction, provides a practical pathway for achieving privacy-preserving inference in black-box LLMs for text generation.

Summary by alphaXiv.