Dynamic privacy protection boosts usefulness of large language models
Demystifying the Privacy-Utility Trade-off in LLM Interactions
Artificial IntelligenceCryptography and Security
Summary
Large language models help with many tasks but need a lot of personal information to work well, which risks privacy. The authors found that privacy methods that treat all data the same harm usefulness a lot. They discovered that what to hide, how to hide it, and how information fits together depends on the user's goal and the task. Using this, they built a system that smartly protects privacy while keeping the model's helpfulness much higher than before.
What this means in practice
- •For ai product developers: Implement privacy features that maintain high task accuracy by tailoring data sanitization to user goals and task needs.$Commercial implications: Enables building privacy-aware AI products that protect user data without sacrificing performance, appealing to privacy-conscious markets.
- •For enterprise it security teams: Deploy adaptive sanitization methods improving privacy protection with less impact on AI system functionality for internal tools.
Authors
Zhenhua Liu, Zhanxu Xie, Junjie Yu, Tong Zhu, Lijun Li, Wenliang Chen
Abstract
The integration of Large Language Models into daily tasks relies on context-rich instructions, inevitably exposing sensitive user information. Current privacy-preserving methods typically employ context-agnostic static rules, causing severe utility degradation. However, the specific mechanisms governing how sanitization impacts downstream performance remain largely underexplored. To address this, we conduct a systematic analysis to deconstruct the privacy-utility trade-off, uncovering three underlying mechanisms: (1) Context-Dependent Utility, which first establishes when to sanitize by revealing that data value shifts from critical constraints to dispensable noise based on user intent; (2) Strategic Adaptation, which subsequently determines how to sanitize by dictating that the choice between removal and replacement depends on the task's reliance on factual integrity versus structural coherence; and (3) Combinatorial Interplay, which finally extends the protection scope by demonstrating that attributes form a semantic web of synergistic dependencies or antagonistic redundancies. Guided by these insights, we introduce an intent-driven local protection framework. By distilling a lightweight model Veilmind-4B to drive a dynamic extraction-sanitization-restoration pipeline, our approach reaches a low-leakage privacy point while preserving substantially higher response utility than existing privacy-oriented baselines, advancing the privacy-utility trade-off toward the Pareto frontier.