Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Zheng Hui, Yijiang River Dong, Sanhanat Sivapiromrat, Ehsan Shareghi, Nigel Collier

TL;DR

PrivacyPAD reframes privacy-utility trade-offs in LLM delegation as a sequential decision-making problem and learns a stateful, chunk-level routing policy that dynamically splits prompts between a local privacy-preserving model and a remote high-capacity model. By optimizing a composite reward that balances task quality with a non-linear privacy penalty, PrivacyPAD achieves a superior privacy-utility frontier across general and medical datasets, and shows particularly strong gains when the local model is weaker. The approach relies on semantic chunking, a Transformer-based policy, and PPO-based reinforcement learning, demonstrated on the PUPA and Med-PCD benchmarks with robust ablations illustrating the importance of stateful context and non-linear leakage penalties. This work underscores the potential of adaptive, learned delegation policies to enable trustworthy use of powerful LLMs in privacy-sensitive settings.

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.

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

TL;DR

PrivacyPAD reframes privacy-utility trade-offs in LLM delegation as a sequential decision-making problem and learns a stateful, chunk-level routing policy that dynamically splits prompts between a local privacy-preserving model and a remote high-capacity model. By optimizing a composite reward that balances task quality with a non-linear privacy penalty, PrivacyPAD achieves a superior privacy-utility frontier across general and medical datasets, and shows particularly strong gains when the local model is weaker. The approach relies on semantic chunking, a Transformer-based policy, and PPO-based reinforcement learning, demonstrated on the PUPA and Med-PCD benchmarks with robust ablations illustrating the importance of stateful context and non-linear leakage penalties. This work underscores the potential of adaptive, learned delegation policies to enable trustworthy use of powerful LLMs in privacy-sensitive settings.

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.
Paper Structure (50 sections, 8 equations, 2 figures, 6 tables)

This paper contains 50 sections, 8 equations, 2 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Training-Time vs. Inference-Time Privacy
  4. Reinforcement Learning for Privacy-Aware Control
  5. Problem formulation
  6. Quality Preservation
  7. Privacy Leakage
  8. PrivacyPAD
  9. Framework Overview
  10. Semantic Chunking
  11. Dynamic Routing and Response Generation
  12. The Policy Agent
  13. Input Representation
  14. Contextualization with Self-Attention
  15. Action Selection
  16. ...and 35 more sections

Figures (2)

  • Figure 1: A static rewriter (left) fails by redacting task-critical information. Our PrivacyPAD framework (right) succeeds by learning a policy to route sensitive and non-sensitive query chunks to the appropriate model, preserving both privacy and task performance.
  • Figure 2: Learning curves showing the reward during PPO training on Med-PCD (left) and PUPA (right).