A multi-agent distributed consensus method and system with attack resilience and privacy protection

By combining distributed estimation of the zero eigenvector of the Laplacian matrix and dynamic masking function to update interactive information, the average consensus bias and privacy leakage problems of multi-agent systems in directed communication topologies are solved, achieving distributed consensus with attack resilience and privacy protection.

CN123074783APending Publication Date: 2026-11-10YANGTZE RIVER DELTA RES INST OF BEIJING JIAOTONG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611250263.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-18
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

In directed communication topologies, distributed consensus algorithms for multi-agent systems struggle to simultaneously achieve average consensus accuracy, resistance to spoofed data injection attacks, and state information privacy protection; existing methods cannot achieve all three simultaneously.

Method used

By using the left eigenvector of the zero eigenvalue of the Laplacian matrix to estimate it in a distributed manner, a virtual auxiliary state and a dynamic mask function are constructed to generate combined interaction information. Consensus state is updated during the iteration process, which enhances the system's attack resilience and privacy protection.

Benefits of technology

It achieves robustness against spoofed data injection attacks in directed communication topologies, while protecting the privacy of state information and ensuring the accuracy and security of distributed average consensus.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123074783A_ABST
    Figure CN123074783A_ABST
Patent Text Reader

Abstract

This invention discloses a multi-agent distributed consensus method and system with attack resilience and privacy protection, targeting multi-agent systems in strongly connected directed communication topologies. Each agent performs distributed average consensus updates based on neighbor interaction information; a distributed estimation mechanism of left eigenvectors is used to achieve directed graph average consensus. The system's robustness is improved under bounded spoofing attacks on communication links. Each agent constructs a virtual auxiliary state, which is coupled with the local physical consensus state for updates, reducing the impact of attack disturbances on the consensus process. Each agent broadcasts combined interaction information consisting of the physical consensus state, the virtual auxiliary state, and a dynamic masking function, preventing eavesdroppers from inferring the agent's true state information. This invention achieves multi-agent distributed average consensus in directed communication network topologies, resists spoofing attacks, and ensures the privacy of agent state information. It is applicable to scenarios such as unmanned swarms, sensor networks, and multi-micronet collaborative optimization.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of multi-agent distributed control technology, specifically to a multi-agent distributed consensus method and system with attack resilience and privacy protection. Background Technology

[0002] Multi-agent systems consist of multiple agents with sensing, communication, and computing capabilities, and are widely used in unmanned swarms, sensor networks, and networked control systems. Distributed consensus aims to achieve state consistency convergence among agents through local neighbor information exchange in the absence of a central node. However, in general strongly connected directed communication topologies, traditional distributed consensus algorithms typically converge to a weighted average determined by the left eigenvector of the network topology, making it difficult to guarantee convergence to the arithmetic mean of the initial states of all agents, thus affecting the accuracy of average consensus in directed networks.

[0003] Meanwhile, multi-agent systems rely on open communication networks for information exchange, making them vulnerable to malicious attacks such as spoofed data injection. This can cause consensus values ​​to deviate from the true target values, and even disrupt the system's collaborative stability. Furthermore, the state information frequently exchanged by agents during the consensus process may contain sensitive data such as local measurements, control parameters, and task information. Directly broadcasting the true state can easily lead to privacy leaks. Existing methods typically struggle to simultaneously achieve average consensus accuracy in directed communication topologies, resilience against spoofed data injection attacks, and the privacy of state information. Therefore, it is necessary to propose a distributed average consensus method for strongly connected directed multi-agent networks that can achieve attack-resistant consensus while protecting the privacy of the agents' true state information.

[0004] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0005] The technical problem to be solved by this invention is to overcome the above-mentioned technical defects and provide a multi-agent distributed consensus method and system with attack resilience and privacy protection. It addresses the problems of average consensus deviation in directed communication topologies, consensus offset caused by false data injection attacks, and privacy leakage during state information interaction. Through average consensus correction injection, virtual auxiliary state compensation, and dynamic masking mechanisms, it achieves distributed average consensus with attack resilience and privacy protection capabilities.

[0006] To address the aforementioned problems, the technical solution of this invention is a multi-agent distributed consensus method with attack resilience and privacy protection, the method comprising the following steps: Step 1: Each agent performs distributed estimation of the left eigenvector corresponding to the zero eigenvalue of the directed communication topology Laplace matrix based on neighbor interaction information, and calculates the injection term used to correct the average consensus bias of the directed graph. Step 2: Each agent constructs a virtual auxiliary state and generates a dynamic mask function based on local private parameters; each agent generates combined interaction information based on its own physical consensus state, virtual auxiliary state and dynamic mask function, and broadcasts the combined interaction information to neighboring agents; Step 3: Each agent performs elastic consensus iteration based on the combined interaction information of its neighbors, the average consensus correction injection item, and the virtual auxiliary state, in order to update the physical consensus state and virtual auxiliary state of each agent and obtain the distributed average consensus output.

[0007] Preferably, in step 1, each agent obtains the estimated value of the left eigenvector corresponding to the zero eigenvalue of the Laplacian matrix by broadcasting its local left eigenvector estimate and performing neighborhood consensus iteration. The specific process is as follows: Step 1.1: Construct a strongly connected directed communication network topology and a consistent iterative model; Consider a A strongly connected directed communication network consisting of several intelligent agents has the following network topology: ,in, A collection of intelligent agents, Let be a set of directed communication edges, and let its adjacency matrix be... When intelligent agents Able to receive intelligent agents When broadcasting information, ;otherwise The Laplace matrix of the network is: ,in, intelligent agent The set of incoming neighbors is: Intelligent agents The physical consensus state is initialized as follows: ,in, For intelligent agents The initial state value; Step 1.2: Update the estimated left eigenvector according to the consensus iterative algorithm. And calculate the average consensus correction injection item. ; intelligent agent The left eigenvector estimation vector is represented as Its initial value is ,in, For the first One standard basis vector; agent Update the estimated vector according to the following rules. : , in, Let be the consensus iteration step size; ,in, Represents the estimated vector The Each component defines an intermediate correction vector. for: , Furthermore, the average consensus correction injection term is obtained: , in, This is the preset initial correction vector.

[0008] Preferably, in step 2, each agent constructs a virtual auxiliary state and generates a dynamic mask function based on local private parameters; each agent uses the dynamic mask function to combine and encode the physical consensus state and the virtual auxiliary state to generate combined interaction information for neighbor interaction, and broadcasts the combined interaction information to neighbor agents. The specific process is as follows: Step 2.1: Construct a virtual assisted network model; The multi-agent consensus model based on virtual assisted networks is shown below: , , in, and Representing intelligent agents respectively At any moment The physical state and virtual auxiliary state; , , These are preset parameters, and they satisfy... and ; Consensus iteration step size; intelligent agent The physical consensus state and the virtual auxiliary state adopt the same communication topology, and transmit corresponding interaction information through independent information interaction channels respectively; Step 2.2: Construct the additive dynamic mask function as described below; , in, , For intelligent agents These are local private parameters that are not visible to other agents in the network. Step 2.3: Construct combined interactive information based on the dynamic mask function; To protect the privacy of interactive information, intelligent agents Do not directly broadcast physical consensus state and virtual assist status Instead, it broadcasts a combination of interactive information consisting of physical consensus states, virtual auxiliary states, and dynamic masking functions; in step 2, the intelligent agent Combined interactive information broadcast to the outside world and They are represented as follows: , , Due to privacy mask parameters and Visible only to the agent within the agent, neighboring agents cannot infer the true state information from the received combined information. and This ensures the privacy of the state information of each intelligent agent.

[0009] Preferably, in step 3, considering the possibility of false data injection attacks in the communication link, each agent performs elastic consensus updates of the physical consensus state and the virtual auxiliary state based on the received neighbor combination interaction information, the average consensus correction injection item, and the virtual auxiliary state. The specific process is as follows: Consider the existence of bounded attack injection signals in the communication link. intelligent agent From the neighboring intelligent agent The received interaction information is represented as follows: , Among them, bounded attack signals and Each condition is satisfied and ,and and This is a preset upper limit for attack. Based on the aforementioned combined interactive information, dynamic masking mechanism, and attack model, the intelligent agent... The physical consensus state has been updated to: , Virtual assistance status updated to: .

[0010] This invention also discloses a multi-agent distributed consensus system with attack resilience and privacy protection. Using the above method, the system is characterized by comprising multiple agent nodes, each agent node including: Left eigenvector distributed estimation module: used to perform distributed estimation of the left eigenvector corresponding to the zero eigenvalue of the Laplacian matrix under a directed communication topology, and generate a correction injection term to eliminate the average consensus bias based on the estimated left eigenvector. Virtual Auxiliary State Construction Module: Used for intelligent agents Constructing virtual auxiliary states and through physical consensus state The coupled update mechanism enables elastic consensus iteration under bounded spoofing injection attack environments; Privacy mask generation module: used to generate privacy masks based on the intelligent agent. Local private parameter generation dynamic mask function , wherein the local private parameters and Only for intelligent agents Visible from within; Combined Interaction Information Generation Module: Used to generate information based on physical consensus status. Virtual Assistance Status and dynamic masking function Generate combined interactive information and ; Broadcast communication module: Used to transmit the combined interaction information generated by the agents during each consensus iteration. and Broadcasts to its neighboring intelligent agents and receives combined interaction information broadcast by neighboring intelligent agents. and ; Elastic consensus iterative update module: used to update the consensus based on the agent's own state information. and Received interaction information from neighboring intelligent agents and Correcting injection items It performs iterative updates of the physical state and virtual auxiliary state to achieve a distributed average consensus with attack resilience and privacy protection.

[0011] The advantages of this invention compared to existing technologies are: 1. This invention enhances the robustness of the system against bounded spoofing attacks by coupling the physical consensus state and the virtual auxiliary state through an update mechanism, and protects the privacy of the interactive state by combining a dynamic mask combination encoding mechanism, thereby achieving a distributed average consensus with attack resilience and privacy protection capabilities. Compared with privacy protection methods that use fixed perturbations or static masks, the mask perturbation in the combined interactive information constructed in this invention gradually decays with the iteration process, thereby enhancing the privacy of state information while eliminating consensus bias. Attached Figure Description

[0012] Figure 1 This is the execution flowchart of the multi-agent distributed consensus method with attack resilience and privacy protection of the present invention.

[0013] Figure 2 This is a diagram of a multi-agent elastic consensus information interaction framework based on an auxiliary virtual network. Detailed Implementation

[0014] To make the content of this invention easier to understand, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Example 1

[0015] This embodiment provides a multi-agent distributed consensus method with attack resilience and privacy protection, the framework of which is as follows: Figure 1 As shown, it includes the following steps: Step 1: Each agent performs distributed estimation of the left eigenvector corresponding to the zero eigenvalue of the directed communication topology Laplace matrix based on neighbor interaction information, and calculates the injection term used to correct the average consensus bias of the directed graph. Step 2: Each agent constructs a virtual auxiliary state and generates a dynamic mask function based on local private parameters; each agent generates combined interaction information based on its own physical consensus state, virtual auxiliary state and dynamic mask function, and broadcasts the combined interaction information to neighboring agents; Step 3: Each agent performs elastic consensus iterative updates based on the combined interaction information of its neighbors, the average consensus correction injection item, and the virtual auxiliary state, in order to update the physical consensus state and virtual auxiliary state of each agent and obtain the distributed average consensus output.

[0016] Furthermore, in step 1, each agent obtains the estimated value of the left eigenvector corresponding to the zero eigenvalue of the Laplacian matrix by broadcasting its local left eigenvector estimate and performing neighborhood consensus iteration. The specific process is as follows: Step 1.1: Construct the topology and consistency iterative model of the strongly connected directed communication network; Consider a A strongly connected directed communication network consisting of several intelligent agents has the following network topology: ,in, A collection of intelligent agents, Let be a set of directed communication edges, and let its adjacency matrix be... When intelligent agents Able to receive intelligent agents When broadcasting information, ;otherwise The Laplace matrix of the network is: ,in, intelligent agent The set of incoming neighbors is: Intelligent agents The physical consensus state is initialized as follows: ,in, For intelligent agents The initial state value.

[0017] Step 1.2: Update the estimated left eigenvector according to the consensus iterative algorithm. And calculate the average consensus correction injection item. ; intelligent agent The left eigenvector estimation vector is represented as Its initial value is ,in, For the first One standard basis vector; agent Update the estimated vector according to the following rules. : , in, Let be the consensus iteration step size; ,in, Represents the estimated vector The Each component defines an intermediate correction vector. for: , Furthermore, the average consensus correction injection term is obtained: , in, This is the preset initial correction vector.

[0018] Further, in step 2, each agent constructs a virtual auxiliary state and generates a dynamic mask function based on local private parameters. Each agent uses the dynamic mask function to combine and encode the physical consensus state and the virtual auxiliary state, generating combined interaction information for neighbor interaction, and broadcasts this combined interaction information to neighboring agents. The information interaction framework between different agents is as follows: Figure 2As shown, the specific process is as follows: Step 2.1: Construct a virtual assisted network model; The multi-agent consensus model based on virtual assisted networks is shown below: , , in, and Representing intelligent agents respectively At any moment The physical state and virtual auxiliary state; , , These are preset parameters, and they satisfy... and ; Consensus iteration step size; intelligent agent The physical consensus state and the virtual auxiliary state adopt the same communication topology, and transmit corresponding interaction information through independent information interaction channels respectively; Step 2.2: Construct the following additive dynamic mask function; , in, , For intelligent agents These are local private parameters that are not visible to other agents in the network. Step 2.3: Construct combined interactive information based on the dynamic mask function; To protect the privacy of interactive information, intelligent agents Do not directly broadcast physical consensus state and virtual assist status Instead, it broadcasts a combination of interactive information consisting of physical consensus states, virtual auxiliary states, and dynamic masking functions; in step 2, the intelligent agent... Combined interactive information broadcast to the outside world and They are represented as follows: , , Due to privacy mask parameters and Visible only to the agent within the agent, neighboring agents cannot infer the true state information from the received combined information. and This ensures the privacy of the state information of each intelligent agent.

[0019] Furthermore, in step 3, considering the possibility of spoofed data injection attacks in the communication link, each agent performs elastic consensus updates of the physical consensus state and the virtual auxiliary state based on the received neighbor combination interaction information, the average consensus correction injection item, and the virtual auxiliary state. The specific process is as follows: Consider the existence of bounded attack injection signals in the communication link. intelligent agent From the neighboring intelligent agent The received interaction information is represented as follows: , Among them, bounded attack signals and Each condition is satisfied and ,and and This is a preset upper limit for attack. Based on the aforementioned combined interactive information, dynamic masking mechanism, and attack model, the intelligent agent... The physical consensus state has been updated to: , Virtual assistance status updated to: . Example 2

[0020] This embodiment provides a multi-agent distributed consensus system with attack resilience and privacy protection. The system includes multiple agent nodes, and each agent node includes: Left eigenvector distributed estimation module: used to perform distributed estimation of the left eigenvector corresponding to the zero eigenvalue of the Laplacian matrix under a directed communication topology, and generate a correction injection term to eliminate the average consensus bias based on the estimated left eigenvector. Virtual Auxiliary State Construction Module: Used for intelligent agents Constructing virtual auxiliary states and through physical consensus state The coupled update mechanism enables elastic consensus iteration under bounded spoofing injection attack environments; Privacy mask generation module: used to generate privacy masks based on the intelligent agent. Local private parameter generation dynamic mask function Among them, local private parameters and Only for intelligent agents Visible from within; Combined Interaction Information Generation Module: Used to generate information based on physical consensus status. Virtual Assistance Status and dynamic masking function Generate combined interactive information and ; Broadcast communication module: Used to transmit the combined interaction information generated by the agents during each consensus iteration. and Broadcasts to its neighboring intelligent agents and receives combined interaction information broadcast by neighboring intelligent agents. and ; Elastic consensus iterative update module: used to update the consensus based on the agent's own state information. and Received interaction information from neighboring intelligent agents and Correcting injection items It performs iterative updates of the physical state and virtual auxiliary state to achieve a distributed average consensus with attack resilience and privacy protection. Example 3

[0021] Embodiment 3 of the present invention provides an electronic device, including a processor, a memory, and a computer program stored in the memory and capable of running on the processor. When the processor executes the program, it implements a multi-agent distributed consensus method with attack resilience and privacy protection as described in Embodiment 1. Example 4

[0022] Embodiment 4 of the present invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements a multi-agent distributed consensus method with attack resilience and privacy protection as described in Embodiment 1.

[0023] The present invention and its embodiments have been described above, and such description is not restrictive. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A multi-agent distributed consensus method with attack resilience and privacy protection, characterized in that, The method includes the following steps: Step 1: Each agent performs distributed estimation of the left eigenvector corresponding to the zero eigenvalue of the directed communication topology Laplace matrix based on neighbor interaction information, and calculates the injection term used to correct the average consensus bias of the directed graph. Step 2: Each agent constructs a virtual auxiliary state and generates a dynamic mask function based on local private parameters; each agent generates combined interaction information based on its own physical consensus state, virtual auxiliary state and dynamic mask function, and broadcasts the combined interaction information to neighboring agents; Step 3: Each agent performs elastic consensus iteration based on the combined interaction information of its neighbors, the average consensus correction injection item, and the virtual auxiliary state, in order to update the physical consensus state and virtual auxiliary state of each agent and obtain the distributed average consensus output.

2. The multi-agent distributed consensus method with attack resilience and privacy protection according to claim 1, characterized in that, In step 1, each agent obtains the estimated value of the left eigenvector corresponding to the zero eigenvalue of the Laplacian matrix by broadcasting its local left eigenvector estimate and performing neighborhood consensus iteration. The specific process is as follows: Step 1.1: Construct a strongly connected directed communication network topology and a consistent iterative model; Consider a A strongly connected directed communication network consisting of several intelligent agents has the following network topology: ,in, A collection of intelligent agents, Let be a set of directed communication edges, and let its adjacency matrix be... When intelligent agents Able to receive intelligent agents When broadcasting information, ;otherwise The Laplace matrix of the network is: ,in, intelligent agent The set of incoming neighbors is: Intelligent agents The physical consensus state is initialized as follows: ,in, For intelligent agents The initial state value; Step 1.2: Update the estimated left eigenvector according to the consensus iterative algorithm. And calculate the average consensus correction injection item. ; intelligent agent The left eigenvector estimation vector is represented as Its initial value is ,in, For the first One standard basis vector; agent Update the estimated vector according to the following rules. : , in, Let be the consensus iteration step size; ,in, Represents the estimated vector The Each component defines an intermediate correction vector. for: , Furthermore, the average consensus correction injection term is obtained: , in, This is the preset initial correction vector.

3. The multi-agent distributed consensus method with attack resilience and privacy protection according to claim 1, characterized in that, In step 2, each agent constructs a virtual auxiliary state and generates a dynamic mask function based on local private parameters. Each agent uses the dynamic mask function to combine and encode the physical consensus state and the virtual auxiliary state to generate combined interaction information for neighbor interaction, and broadcasts the combined interaction information to neighbor agents. The specific process is as follows: Step 2.1: Construct a virtual assisted network model; The multi-agent consensus model based on virtual assisted networks is shown below: , , in, and Representing intelligent agents respectively At any moment The physical state and virtual auxiliary state; , , These are preset parameters, and they satisfy... and ; Consensus iteration step size; intelligent agent The physical consensus state and the virtual auxiliary state adopt the same communication topology, and transmit corresponding interaction information through independent information interaction channels respectively; Step 2.2: Construct the additive dynamic mask function as described below; , in, , For intelligent agents These are local private parameters that are not visible to other agents in the network. Step 2.3: Construct combined interactive information based on the dynamic mask function; To protect the privacy of interactive information, intelligent agents Do not directly broadcast physical consensus state and virtual assist status Instead, it broadcasts a combination of interactive information consisting of physical consensus states, virtual auxiliary states, and dynamic masking functions; in step 2, the intelligent agent Combined interactive information broadcast to the outside world and They are represented as follows: , , Due to privacy mask parameters and Visible only to the agent within the agent, neighboring agents cannot infer the true state information from the received combined information. and This ensures the privacy of the state information of each intelligent agent.

4. The multi-agent distributed consensus method with attack resilience and privacy protection according to claim 1, characterized in that, In step 3, considering the possibility of false data injection attacks in the communication link, each agent performs elastic consensus updates of the physical consensus state and the virtual auxiliary state based on the received neighbor combination interaction information, the average consensus correction injection item, and the virtual auxiliary state. The specific process is as follows: Consider the existence of bounded attack injection signals in the communication link. intelligent agent From the neighboring intelligent agent The received interaction information is represented as follows: , Among them, bounded attack signals and Each condition is satisfied and ,and and This is a preset upper limit for attack. Based on the aforementioned combined interactive information, dynamic masking mechanism, and attack model, the intelligent agent... The physical consensus state has been updated to: , Virtual assistance status updated to: 。 5. A multi-agent distributed consensus system with attack resilience and privacy protection, employing the method described in any one of claims 1-5, characterized in that... The system includes multiple agent nodes, and each agent node includes: Left eigenvector distributed estimation module: used to perform distributed estimation of the left eigenvector corresponding to the zero eigenvalue of the Laplacian matrix under a directed communication topology, and generate a correction injection term to eliminate the average consensus bias based on the estimated left eigenvector. Virtual Auxiliary State Construction Module: Used for intelligent agents Constructing virtual auxiliary states and through physical consensus state The coupled update mechanism enables elastic consensus iteration under bounded spoofing injection attack environments; Privacy mask generation module: used to generate privacy masks based on the intelligent agent. Local private parameter generation dynamic mask function , wherein the local private parameters and Only for intelligent agents Visible from within; Combined Interaction Information Generation Module: Used to generate information based on physical consensus status. Virtual Assistance Status and dynamic masking function Generate combined interactive information and ; Broadcast communication module: Used to transmit the combined interaction information generated by the agents during each consensus iteration. and Broadcasts to its neighboring intelligent agents and receives combined interaction information broadcast by neighboring intelligent agents. and ; Elastic consensus iterative update module: used to update the consensus based on the agent's own state information. and Received interaction information from neighboring intelligent agents and Correcting injection items It performs iterative updates of the physical state and virtual auxiliary state to achieve a distributed average consensus with attack resilience and privacy protection.

6. An electronic device comprising a processor, a memory, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the program, it implements a multi-agent distributed consensus method with attack resilience and privacy protection as described in any one of claims 1 to 4.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the multi-agent distributed consensus method with attack resilience and privacy protection as described in any one of claims 1 to 4.