The integration of Large Language Models into daily tasks necessitates the use of context-rich instructions, potentially exposing sensitive user information. Current privacy-preserving methods often rely on context-agnostic static rules, resulting in significant utility degradation. This research investigates the mechanisms driving the privacy-utility trade-off within LLM interactions. The study identifies three key mechanisms: Context-Dependent Utility, Strategic Adaptation, and Combinatorial Interplay. The framework introduces an intent-driven local protection framework utilizing a lightweight model, Veilmind-4B, to implement a dynamic extraction-sanitization-restoration pipeline. This approach achieves a low-leakage privacy point while preserving substantially higher response utility than existing privacy-oriented baselines. The research suggests a shift toward the Pareto frontier for the privacy-utility trade-off.
Source: https://arxiv.org/abs/2609.10992