PrivacyPAD: A Reinforcement Learning Framework for Dynamic Privacy-Aware Delegation

2025-10-16

AI Summary

Users face an inherent trade-off between privacy and performance in LLM applications: uploading sensitive prompts to powerful cloud-based models risks data leakage, whereas privacy-preserving local small models suffer from inadequate capability. Existing static prompt rewriting methods impair semantic coherence and indiscriminately remove critical PII. Method: This work pioneers modeling privacy-aware prompt delegation as a sequential decision-making problem. We propose a reinforcement learning–based adaptive text chunking and routing framework that dynamically discriminates between maskable PII and task-critical information, jointly optimizing local masking and remote API invocation for fine-grained privacy–utility trade-offs. Contribution/Results: Evaluated on a high-density PII medical dataset, our approach significantly outperforms baselines, achieving state-of-the-art balance between privacy protection strength (e.g., PII anonymization fidelity) and downstream task performance (e.g., clinical QA accuracy).

Technology Category

• Machine Learning: Privacy
• Natural Language Processing: Prompt Engineering / Prompting
• Planning, Routing, and Scheduling: Planning with Language Models

Application Category

• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• User Modeling, Personalization and Recommendation: User privacy protection in personalized systems
• Security and Privacy: Privacy-enhancing technologies

Abstract

When users submit queries to Large Language Models (LLMs), their prompts can often contain sensitive data, forcing a difficult choice: Send the query to a powerful proprietary LLM providers to achieving state-of-the-art performance and risk data exposure, or relying on smaller, local models guarantees data privacy but often results in a degradation of task performance. Prior approaches have relied on static pipelines that use LLM rewriting, which shatters linguistic coherence and indiscriminately removes privacy-sensitive information, including task-critical content. We reformulate this challenge (Privacy-Conscious Delegation) as a sequential decision-making problem and introduce a novel reinforcement learning (RL) framework called PrivacyPAD to solve it. Our framework trains an agent to dynamically route text chunks, learning a policy that optimally balances the trade-off between privacy leakage and task performance. It implicitly distinguishes between replaceable Personally Identifiable Information (PII) (which it shields locally) and task-critical PII (which it strategically sends to the remote model for maximal utility). To validate our approach in complex scenarios, we also introduce a new medical dataset with high PII density. Our framework achieves a new state-of-the-art on the privacy-utility frontier, demonstrating the necessity of learned, adaptive policies for deploying LLMs in sensitive environments.

Problem

• Balancing privacy protection with task performance in LLM query delegation
• Dynamically routing sensitive data between local and remote models
• Distinguishing replaceable from task-critical private information in prompts

Innovation

• Uses reinforcement learning for dynamic query routing
• Learns policy balancing privacy leakage and performance
• Distinguishes between replaceable and task-critical PII