Reproduce “DePrompt: Desensitization and Evaluation of Personal Identifiable Information in Large Language Model Prompts” on OpenResearch

About this paper

Xiongtao Sun, Gan Liu, Zhipeng He, Hui Li, Xiaoguang Li

DePrompt introduces a framework for desensitizing Personally Identifiable Information (PII) in Large Language Model prompts during the inference stage, balancing privacy protection with prompt usability. It achieves high-accuracy PII recognition through fine-tuned LLMs and employs an adversarial generative desensitization method that preserves semantic content while disrupting identifier-attribute linkages, outperforming traditional methods in usability.

Key insights

• Balancing privacy and usability requires a dynamic approach that leverages LLM's comprehension and generation capabilities to retain semantic content while disrupting PII linkages, rather than simply removing or anonymizing data.
• Defining and identifying PII with contextual attributes (semanticity, linkability, uncertainty) allows for a more nuanced desensitization strategy that adapts to the specific impact of different entities on prompt utility.
• An adversarial generative approach can confuse attackers by obfuscating PII and making traditional extraction difficult, yet still allow LLMs to comprehend the desensitized content, offering a superior trade-off compared to methods that introduce noise or remove information.

Summary by alphaXiv.