Demystifying the Privacy-Utility Trade-off in LLM Interactions

Summary

Large language models often need context-rich instructions for daily tasks, but those instructions can expose sensitive user information. The paper argues that common privacy-preserving methods rely on static, context-agnostic rules that can substantially reduce utility, while the mechanisms behind this loss remain insufficiently understood. Its analysis identifies three mechanisms: the value of information depends on user intent, so data can shift from a critical constraint to dispensable noise; the choice between removing and replacing information should reflect whether a task depends more on factual integrity or structural coherence; and attributes can interact through synergistic dependencies or antagonistic redundancies, changing the scope of protection. Based on these findings, the authors introduce an intent-driven local protection framework built around a dynamic extraction-sanitization-restoration pipeline. A lightweight model, Veilmind-4B, drives the pipeline. The abstract reports that the approach reaches a low-leakage privacy point while preserving substantially higher response utility than existing privacy-oriented baselines, moving the trade-off closer to the Pareto frontier.