NAP^2: A Benchmark for Naturalness and Privacy-Preserving Text Rewriting by Learning from Human

TL;DR

This paper tackles privacy risks in interactions with cloud-based LLMs by proposing a human-inspired text rewriting task that sanitizes sensitive content through deleting and obscuring. It introduces NaP^2, a corpus built from manual rewrites and GPT-4 synthetic data grounded in PERSONA-CHAT privacy profiles, and a robust evaluation framework including a novel Privacy_NLI metric. Results show that a T5-Base model trained on NaP^2 with synthetic augmentation achieves strong privacy preservation and naturalness, outperforming differential privacy baselines and zero-shot LLMs, while maintaining semantic utility. The work provides a practical path for inference-time privacy protection and offers a reusable dataset for developing privacy-preserving rewriting systems on open-domain dialogues.

Abstract

The widespread use of cloud-based Large Language Models (LLMs) has heightened concerns over user privacy, as sensitive information may be inadvertently exposed during interactions with these services. To protect privacy before sending sensitive data to those models, we suggest sanitizing sensitive text using two common strategies used by humans: i) deleting sensitive expressions, and ii) obscuring sensitive details by abstracting them. To explore the issues and develop a tool for text rewriting, we curate the first corpus, coined NAP^2, through both crowdsourcing and the use of large language models (LLMs). Compared to the prior works on anonymization, the human-inspired approaches result in more natural rewrites and offer an improved balance between privacy protection and data utility, as demonstrated by our extensive experiments. Researchers interested in accessing the dataset are encouraged to contact the first or corresponding author via email.

Figures (3)

• Figure 1: Human evaluation of privacy leakage.
• Figure 2: Prompt template for T5-NaP$^2$
• Figure 3: Prompt used for LLM-based naturalness judgment.