Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Interaction-Aware Parameter Privacy-Preserving Data Sharing in Coupled Systems via Particle Filter Reinforcement Learning

Haokun Yu, Jingyuan Zhou, Kaidi Yang

TL;DR

This work tackles parameter privacy in data sharing across coupled cyber-physical systems by formulating a joint privacy-control objective that minimizes mutual information about a sensitive parameter $\Theta$ while preserving the data provider's control performance. It develops a particle-filter reinforcement learning framework to solve a Bellman-equation-based optimization, reducing history dependence and enabling continuous-state handling. A key advance is the belief-state representation via particle filters and an MGF-encoded actor-critic (MGF-A2C) that generates distortion policies $Y_t$ from the belief over $(\Theta, X^A_t)$. Experiments in a mixed-autonomy platoon show effective protection of HDV driving parameters against BI and RLS inference with only a small fuel-efficiency penalty, highlighting the method’s practical relevance for privacy-preserving collaborative systems.

Abstract

This paper addresses the problem of parameter privacy-preserving data sharing in coupled systems, where a data provider shares data with a data user but wants to protect its sensitive parameters. The shared data affects not only the data user's decision-making but also the data provider's operations through system interactions. To trade off control performance and privacy, we propose an interaction-aware privacy-preserving data sharing approach. Our approach generates distorted data by minimizing a combination of (i) mutual information, quantifying privacy leakage of sensitive parameters, and (ii) the impact of distorted data on the data provider's control performance, considering the interactions between stakeholders. The optimization problem is formulated into a Bellman equation and solved by a particle filter reinforcement learning (RL)-based approach. Compared to existing RL-based methods, our formulation significantly reduces history dependency and efficiently handles scenarios with continuous state space. Validated in a mixed-autonomy platoon scenario, our method effectively protects sensitive driving behavior parameters of human-driven vehicles (HDVs) against inference attacks while maintaining negligible impact on fuel efficiency.

Interaction-Aware Parameter Privacy-Preserving Data Sharing in Coupled Systems via Particle Filter Reinforcement Learning

TL;DR

This work tackles parameter privacy in data sharing across coupled cyber-physical systems by formulating a joint privacy-control objective that minimizes mutual information about a sensitive parameter while preserving the data provider's control performance. It develops a particle-filter reinforcement learning framework to solve a Bellman-equation-based optimization, reducing history dependence and enabling continuous-state handling. A key advance is the belief-state representation via particle filters and an MGF-encoded actor-critic (MGF-A2C) that generates distortion policies from the belief over . Experiments in a mixed-autonomy platoon show effective protection of HDV driving parameters against BI and RLS inference with only a small fuel-efficiency penalty, highlighting the method’s practical relevance for privacy-preserving collaborative systems.

Abstract

This paper addresses the problem of parameter privacy-preserving data sharing in coupled systems, where a data provider shares data with a data user but wants to protect its sensitive parameters. The shared data affects not only the data user's decision-making but also the data provider's operations through system interactions. To trade off control performance and privacy, we propose an interaction-aware privacy-preserving data sharing approach. Our approach generates distorted data by minimizing a combination of (i) mutual information, quantifying privacy leakage of sensitive parameters, and (ii) the impact of distorted data on the data provider's control performance, considering the interactions between stakeholders. The optimization problem is formulated into a Bellman equation and solved by a particle filter reinforcement learning (RL)-based approach. Compared to existing RL-based methods, our formulation significantly reduces history dependency and efficiently handles scenarios with continuous state space. Validated in a mixed-autonomy platoon scenario, our method effectively protects sensitive driving behavior parameters of human-driven vehicles (HDVs) against inference attacks while maintaining negligible impact on fuel efficiency.
Paper Structure (21 sections, 5 theorems, 43 equations, 4 figures, 2 tables)

This paper contains 21 sections, 5 theorems, 43 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Problem Statement
  3. Characterization of Interaction-Aware Privacy-Preserving Data-Sharing Policy
  4. Particle Filter Reinforcement Learning-Based Solution Method
  5. Particle formulation of belief state
  6. A2C solution algorithm with MGF encoder
  7. Experiments
  8. Experimental setup
  9. Privacy-preserving performance
  10. Tradeoff between privacy and control performance
  11. Conclusion
  12. Proof for Simplification of MI in Section \ref{['sec:Problem Statement']}
  13. Proofs of Sec. \ref{['sec: A Computationally Efficient Privacy-Preserving Policy']}
  14. Proof for Theorem \ref{['th:simplified policy theorem']}
  15. Derivation of Belief State Update
  16. ...and 6 more sections

Key Result

theorem 1

There exists a sequence of policy $\boldsymbol{\pi^{s,*}} \in \boldsymbol{\Pi^s}$ such that $\boldsymbol{\pi^{s,*}}$ is an optimal solution to the optimization problem eq:original optimization. Specifically, let $L(\cdot)$ denote the optimization problem eq:original optimization and $\boldsymbol{\pi

Figures (4)

  • Figure 1: Interaction Between System $A$ and System $B$.
  • Figure 2: Overview of the proposed privacy-preserving data-sharing framework.
  • Figure 3: Experiment settings and training results
  • Figure 4: Comparison of Bayesian attacker's belief over time for different values of $\theta$. Each figure shows the evolution of the attacker's belief when observing true data versus distorted (filtered) data.

Theorems & Definitions (5)

  • theorem 1
  • lemma 1
  • theorem 2
  • theorem 3
  • lemma 2