PrivPrompt: A Protection Framework of Personal Identifiable Information for Large Language Model Prompts Based on Privacy Computing

When interacting with Large Language Models (LLMs), prompts play a crucial role, significantly influencing the accuracy and interpretability of model outputs. However, crafting precise prompts to elicit accurate and high-quality responses inevitably poses significant risks of Personally Identifiable Information (PII) leakage. To address this challenge, we introduce PrivPrompt, a privacy computing framework designed to protect PII in LLM prompts while evaluating its own effectiveness. By leveraging advanced privacy computing techniques, our approach integrates contextual attributes to define privacy types, thereby achieving high-precision PII entity identification. Specifically, through an in-depth analysis of key features in prompt desensitization scenarios, we develop adversarial generative methods that preserve essential semantic content while disrupting the linkage between identifiers and privacy attributes. Furthermore, we propose utility evaluation metrics for prompts, aiming to better balance privacy protection and usability. Our framework is not only applicable to prompt protection but also extendable to text-based applications where usability is critical. Experimental evaluations demonstrate that, compared to benchmarks and other models, our desensitized prompts offer superior privacy protection and maintain or improve model performance on downstream tasks.