Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Privacy in Multi-agent Systems

Yongqiang Wang

TL;DR

Privacy in Multi-agent Systems surveys how distributed coordination models expose private data and reviews five main privacy approaches across static and dynamic consensus and distributed optimization. It highlights the challenges of applying static privacy techniques to dynamics, and presents representative methods: state/decomposition, dynamics-based privacy, partial homomorphic encryption, and differential privacy, including persistent noise strategies that preserve convergence. The paper also covers extensions to other MAS algorithms, and illustrates applications in robot networks and decentralized machine learning, underscoring the need for co-design between privacy mechanisms and coordination rules. Overall, it argues that secure, private MAS require integrated design choices to maintain both privacy guarantees and the accuracy of collaborative objectives, with practical impact for critical domains like power and transportation systems.

Abstract

With the increasing awareness of privacy and the deployment of legislations in various multi-agent system application domains such as power systems and intelligent transportation, the privacy protection problem for multi-agent systems is gaining increased traction in recent years. This article discusses some of the representative advancements in the filed.

Privacy in Multi-agent Systems

TL;DR

Privacy in Multi-agent Systems surveys how distributed coordination models expose private data and reviews five main privacy approaches across static and dynamic consensus and distributed optimization. It highlights the challenges of applying static privacy techniques to dynamics, and presents representative methods: state/decomposition, dynamics-based privacy, partial homomorphic encryption, and differential privacy, including persistent noise strategies that preserve convergence. The paper also covers extensions to other MAS algorithms, and illustrates applications in robot networks and decentralized machine learning, underscoring the need for co-design between privacy mechanisms and coordination rules. Overall, it argues that secure, private MAS require integrated design choices to maintain both privacy guarantees and the accuracy of collaborative objectives, with practical impact for critical domains like power and transportation systems.

Abstract

With the increasing awareness of privacy and the deployment of legislations in various multi-agent system application domains such as power systems and intelligent transportation, the privacy protection problem for multi-agent systems is gaining increased traction in recent years. This article discusses some of the representative advancements in the filed.
Paper Structure (28 sections, 4 equations, 5 figures, 2 tables)

This paper contains 28 sections, 4 equations, 5 figures, 2 tables.

Table of Contents

  1. Privacy in Multi-agent Systems
  2. Introduction
  3. Privacy protection for static average consensus
  4. Problem formulation
  5. Static average consensus
  6. Privacy in static average consensus
  7. Literature review
  8. Partially homomorphic encryption based approaches
  9. Decomposition based approaches
  10. Dynamics based approaches
  11. Differential privacy based approaches
  12. Observation based privacy
  13. Privacy protection for dynamic average consensus
  14. Problem formulation
  15. Dynamic average consensus
  16. ...and 13 more sections

Figures (5)

  • Figure 1: A step-by-step illustration of the confidential interaction protocol. Single arrows indicate the flow of computations; double arrows indicate data exchange via a communication channel. Shaded nodes indicate the computation done in ciphertext. Note that $a_{1 \shortrightarrow 2}$ and $a_{2 \shortrightarrow 1}$ are different from step to step ruan2019secure.
  • Figure 2: State-decomposition based privacy-preserving average consensus wang2019privacy. (a) Before state decomposition (b) After state decomposition
  • Figure 3: The two different gradient functions of node $1$ that lead to identical observations gao2022dynamics.
  • Figure 4: The interaction graph.
  • Figure 5: Comparison of Algorithm 1 in wang2022tailoring with the distributed gradient descent algorithm (DGD) in nedic2009distributed (under the same noise) and the differential-privacy approach for decentralized optimization PDOP in huang2015differentially (under the same privacy budget) using the MNIST image classification problem

Theorems & Definitions (1)

  • Definition 1