Deep learning based multi-agent system consistency attack method and system
By identifying and attacking key nodes in a multi-agent system using a deep learning-based approach, and employing a consensus attack algorithm to split the system, this approach addresses the shortcomings of existing multi-agent systems under attack, achieving greater robustness and resistance to attacks.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG SCI-TECH UNIV
- Filing Date
- 2022-09-27
- Publication Date
- 2026-07-31
AI Technical Summary
Existing consensus algorithms for multi-agent systems are susceptible to negative impacts when facing external attacks or internal obstacles, and there is a lack of effective attack methods to understand and improve their shortcomings.
A deep learning-based approach is used to identify the two most suitable nodes through a CNN network, and a consistency attack algorithm is used to make their update directions different, thereby splitting the system and causing the entire system to fail to achieve consistency.
Effectively splitting multi-agent systems prevents them from converging and achieving consistency, thereby improving the system's robustness and resistance to attacks.
Smart Images
Figure CN115438350B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of multi-agent system technology and deep learning technology, and in particular to a consensus attack method and system for multi-agent systems based on deep learning. Background Technology
[0002] In recent years, with the continuous development of multi-agent technology, multi-agent systems have been widely applied in real life and scientific and technological fields. As a popular and mature research area, multi-agent systems involve arranging multiple agents, each with the ability to observe the environment and communicate. These agents can complete common tasks in a centralized or distributed manner. Due to the autonomy, robustness, and cooperation of multi-agent systems, they are widely used in many fields, such as aerospace, the Internet of Things, and distributed artificial intelligence. However, some agents may exhibit abnormal behavior due to external attacks or internal obstacles. These failure models can have many negative impacts on the system. Therefore, the fault-tolerant consensus problem has received widespread attention. Among them, the recently proposed Q-consensus algorithm has shown significant results, with low topology requirements and fast convergence speed.
[0003] While multi-agent consensus research has become a hot topic, the study of attack methods on the consensus of multi-agent systems is also particularly important due to the need to consider the universality of consensus methods. The application and verification of consensus attack methods help to understand the shortcomings of consensus methods and improve them. Based on this, this invention proposes a deep learning-based method and system for attacking the consensus of split multi-agent systems. This invention addresses this issue by targeting the Q-consensus algorithm, rendering it ineffective in this model. Summary of the Invention
[0004] The purpose of this invention is to propose a consensus attack method and system for multi-agent systems based on deep learning, which is aimed at existing consensus algorithms.
[0005] To achieve the above objectives, the present invention adopts the following technical solution:
[0006] A consensus attack method for multi-agent systems based on deep learning, comprising the following steps:
[0007] S1. Construct a multi-agent fault-tolerant and consistent system model;
[0008] S2. Use a CNN network to obtain two suitable nodes for attack on the constructed system model;
[0009] S3. Attack the two selected nodes and update them using a consistency attack algorithm;
[0010] S4. Perform interactive iterative processing on all agents after steps S2 and S3, causing the entire system model to be split and unable to reach a consensus.
[0011] Furthermore, in step S1, a multi-agent fault-tolerant consistency system model is constructed, where the topological structure of the multi-agent system is represented as G(V,E,A), where V={1,2,⋯,n} represents the set of agents, E⊆V×V represents the connection relationship between agents, and A represents the weighted adjacency matrix composed of non-negative elements. composition, Let be the weight of node i to node j at time k; an edge exists from node i to node j if and only if And such a node i is called a neighbor of node j, and the set of neighbors of node i is called the neighborhood of node j. ,and The state information of node i at time k is represented as follows: In this invention, if node i is a neighbor of node j, then node j is a neighbor of node i, and there must exist a node from which a path to any other node exists.
[0012] Furthermore, step S2 specifically includes:
[0013] S21. The dataset required for training, including a training set and a test set, preferably consists of: randomly generating m topologies and calculating the adjacency matrix of these topologies. Each topology is randomly assigned an initial value 1000 times, and the success rate of each graph and each node combination breaking the system's convergence consistency is calculated. The output of the neural network is a vector indicating whether the success rate of breaking the system is greater than 50%.
[0014] S22. Input the training set into the CNN model to train the CNN neural network;
[0015] S23. Input the test set into the trained CNN model to obtain the two nodes finally selected for attack. .
[0016] Furthermore, in step S3, the two selected nodes are attacked and updated using a consistency attack algorithm, which is represented as updating the two selected nodes. As faulty nodes, their connections are ignored, and state updates employ a consistency attack algorithm, specifically as follows:
[0017]
[0018]
[0019] in, Indicates an error node The average value of the neighboring nodes, similarly, Indicates an error node The average value of the neighboring nodes, This is represented as a perturbation, and k represents the k-th update.
[0020] Furthermore, in step S4, all nodes are processed iteratively. Normal nodes are updated in state according to the Q-consistency method, as follows:
[0021]
[0022] in, Represents a node , Represents a node For nodes exist Weight of time, Represented as the pair of nodes at time k The disturbance Represents a node exist The state at any given moment. The state of faulty nodes is updated according to a consistency attack algorithm.
[0023] This invention also provides a system based on the above-mentioned deep learning-based multi-agent system consensus attack method, comprising the following modules:
[0024] The building block is used to construct a multi-agent fault-tolerant and consistent system model;
[0025] The processing module is used to extract two suitable nodes for attack from the constructed system model using a CNN network;
[0026] The attack module is used to attack two selected nodes that are updating using a consistency attack algorithm.
[0027] The iteration module is used to perform interactive iterative processing on all the processed agents, which causes the entire system model to be split and unable to reach a consensus.
[0028] Compared to existing attack methods, this invention uses a consistency attack algorithm to attack the Q-consistency algorithm, which ultimately prevents the system from achieving convergence consistency. This invention increases the trust of neighboring nodes in the erroneous node by taking the average value of neighboring nodes, and by adding a certain amount of perturbation, it pulls the neighboring nodes of the erroneous node, thus ultimately preventing the entire system from achieving convergence consistency. Attached Figure Description
[0029] Figure 1 This is a flowchart of the consensus attack method for multi-agent systems based on deep learning, as described in this invention.
[0030] Figure 2 It is a multi-agent fault-tolerant and consistent system model that is constructed.
[0031] Figure 3 It is based on Figure 2 The results of the experiment on the change of agent state values conducted by Topology are shown in the figure.
[0032] Figure 4 This is a block diagram of a multi-agent system consistency attack system based on deep learning, as presented in this invention. Detailed Implementation
[0033] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in the specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.
[0034] The purpose of this invention is to propose a consensus attack method and system for multi-agent systems based on deep learning, which is aimed at existing consensus algorithms.
[0035] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0036] Example 1
[0037] This embodiment provides a deep learning-based method for splitting a multi-agent system's consistency attack, such as... Figure 1 As shown, the specific steps include the following:
[0038] S1. Construct a multi-agent fault-tolerant and consistent system model;
[0039] S2. Use a CNN network to obtain two suitable nodes for attack on the constructed system model;
[0040] S3. Attack the two selected nodes and update them using a consistency attack algorithm;
[0041] S4. Perform interactive iterative processing on all agents after steps S2 and S3, causing the entire system to be split up and unable to reach a consensus.
[0042] This embodiment provides a consensus attack method based on splitting a multi-agent system. The aim is to attack two faulty agents in a multi-agent system, each updating in a different direction, thereby splitting the system into multiple parts and preventing convergence. This strategy is validated by applying a deep learning-based consensus attack algorithm to the Q-consistency method.
[0043] Q-consistency is a method for solving the fault-tolerant consistency problem. In this method, each agent scores its neighbors at each time step by comparing its own value with that of its neighbors and updates its Q-value in real time. The Q-value is used to calculate the weight of each neighbor, and the agent updates its own state value according to the weight. The weight can also reflect the erroneous agents. However, Q-consistency focuses on each agent rather than the entire system. Therefore, this embodiment divides the system into multiple subsystems. Each agent may achieve convergence in a subsystem, but the entire system may not converge.
[0044] A consensus attack method for multi-agent systems based on deep learning is proposed. The main technical idea is to identify the two agents most suitable for attack in the Q-consensus system through a deep learning CNN network, and update these two agents in two different directions according to the consensus attack algorithm, thereby splitting the entire system and preventing it from converging.
[0045] Specifically, in step S1, a multi-agent fault-tolerant consistency system model is constructed, such as... Figure 2 As shown.
[0046] Construct a multi-agent fault-tolerant consistency system model, where the topology of the multi-agent system is represented by G(V,E,A), where V={1,2,⋯,n} represents the set of agents, E⊆V×V represents the connection relationship between agents, and A represents the weighted adjacency matrix consisting of non-negative elements. composition, Let be the weight of node i to node j at time k. An edge exists from node i to node j if and only if... And such a node i is called a neighbor of node j, and the set of neighbors of node i is called the neighborhood of node j. ,and The state information of node i at time k is represented as follows: In this invention, due to the use of an undirected graph quasi-strongly connected topology, if node i is a neighbor of node j, then node j is a neighbor of node i, and there must exist a node from which a path to any other node exists. This embodiment contains 12 agents. All agents are randomly connected, and no agents are isolated.
[0047] In step S2, the constructed system model is processed using a CNN network to obtain two nodes suitable for attack.
[0048] S21. The dataset required for training, including a training set and a test set, specifically: m topologies are randomly generated, and the adjacency matrix of these topologies is calculated. Each topology is randomly assigned an initial value 1000 times, and the success rate of each graph and each node combination breaking the system's convergence consistency is calculated. The output of the neural network is a vector indicating whether the success rate of breaking the system is greater than 50%.
[0049] S22. Input the training set into the CNN model to train the CNN neural network a;
[0050] S23. Input the test set into the trained CNN model to obtain the two nodes finally selected for attack. .
[0051] In step S3, the two selected nodes are attacked and updated using a consistency attack algorithm.
[0052] The two nodes selected in step S2 As faulty nodes, their connections are ignored, and state updates employ a consistency attack algorithm, specifically as follows:
[0053]
[0054]
[0055] in, Indicates an error node Status information, Indicates an error node Status information, Indicates an error node The average value of the neighboring nodes, similarly, Indicates an error node The average value of the neighboring nodes, This is represented as a perturbation, and k represents the k-th update.
[0056] In step S4, all processed agents undergo interactive iterative processing, causing the entire system to be fragmented and unable to reach consensus.
[0057] Normal nodes update their state according to the Q-consistency method, represented as follows:
[0058]
[0059] in, Represents a node , Represents a node For nodes exist Weight of time, Represented as the pair of nodes at time k The disturbance Represents a node exist The state at any given moment. The state of faulty nodes is updated according to a consistency attack algorithm.
[0060] Figure 3 The following is based on Figure 2 The results of the topological experiment on agent state value changes are shown, with the horizontal axis representing the time series and the vertical axis representing the state value of each agent. The graph records the state value of each agent at each time step. It can be seen that as time increases, the system is split into two subsystems, each biased towards its corresponding faulty node. Later, the agents within each subsystem reach convergence, but the entire system does not converge.
[0061] Table 1 shows the comparative experimental results between the node in the consistency attack model proposed in this invention and other faulty nodes.
[0062] Attack nodes of two deep learning-based multi-agent consensus methods 72.83% Two constant value fault nodes 45.43% Two random value fault nodes 12.22% One constant-value fault node, one random-value fault node 34.09%
[0063] Table 1 shows a comparative experiment between the nodes of the consistency attack model proposed in this invention and other faulty nodes. This experiment considered three different types of faulty nodes, including the multi-agent consistency attack method based on deep learning proposed in this invention, constant-value faulty nodes, and random-value faulty nodes. Ten topologies containing 12 nodes were randomly generated to meet the requirements of step S1. Using the three different types of faulty nodes, each experiment was repeated 1000 times, and the differences in success rate were observed. It can be seen that the method proposed in this invention is superior to the other two types of faulty nodes.
[0064] The experimental results are evaluated as follows:
[0065] 1. Aggressiveness
[0066] like Figure 3 As shown, in a multi-agent system, all normal agents achieve convergence in the subsystems, but the entire system does not converge. Therefore, this method is offensive to system consistency.
[0067] 2. Universality
[0068] As shown in Table 1, this method is capable of attacking any randomly generated topology.
[0069] 3. Excellent results
[0070] As shown in Table 1, this method has a higher success rate and better results compared to other fault types.
[0071] Example 2
[0072] like Figure 4As shown, this embodiment provides a system based on the multi-agent system consensus attack method described in Embodiment 1, which includes the following modules:
[0073] The building block is used to construct a multi-agent fault-tolerant and consistent system model;
[0074] The processing module is used to extract two suitable nodes for attack from the constructed system model using a CNN network;
[0075] The attack module is used to attack two selected nodes and update them using a consistency attack algorithm.
[0076] The iteration module is used to perform interactive iterative processing on all the processed agents, which causes the entire system model to be split and unable to reach a consensus.
[0077] The above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.
[0078] This plan will be used under legal circumstances.
Claims
1. A consensus attack method for multi-agent systems based on deep learning, characterized in that, Including the following steps: S1. Construct a multi-agent fault-tolerant and consistent system model; S2. The constructed system model is processed using a CNN network to obtain two suitable nodes for attack, specifically including: S21. The dataset required for training, including the training set and the test set; S22. Input the training set into the CNN network and train the CNN network; S23. Input the test set into the trained CNN network to obtain the two nodes finally selected for attack. ; S3. Attack the two selected nodes and update them using a consistency attack algorithm. This is represented as updating the two selected nodes... As faulty nodes, their connections are ignored, and state updates employ a consistency attack algorithm, specifically as follows: in, Indicates an error node Status information, Indicates an error node Status information, Indicates an error node The average value of the neighboring nodes, Indicates an error node The average value of the neighboring nodes, This is represented as a perturbation, and k represents the k-th update; S4. Perform interactive iterative processing on all agents after steps S2 and S3, causing the entire system model to be fragmented and unable to reach consensus; in step S4, iteratively process all nodes, and normal nodes update their states according to the Q-consistency method, as shown below: in, Represents a node. Represents a node For nodes exist Weight of time, Represented as the pair of nodes at time k The disturbance Represents a node exist The state at any given moment; the state update method for faulty nodes is based on the consistency attack algorithm.
2. The method for attacking consensus in a multi-agent system based on deep learning according to claim 1, characterized in that, In step S1, a multi-agent fault-tolerant consistency system model is constructed, where the topology of the multi-agent system is represented as G(V,E,A), V={1,2,⋯,n} represents the set of agents, E⊆V×V represents the connection relationship between agents, and A represents the weighted adjacency matrix composed of non-negative elements. Composition; An edge exists from node i to node j if and only if And such a node i is called a neighbor of node j, and the set of neighbors of node i is called the neighborhood of node j. ,and .
3. The method for attacking consensus in a multi-agent system based on deep learning according to claim 1, characterized in that, Step S21 is as follows: Randomly generate m topologies and calculate the adjacency matrix of these topologies. Each topology is randomly assigned an initial value 1000 times. The success rate of breaking the system convergence consistency for each graph and each node combination is calculated. The output of the network is a vector indicating whether the success rate of breaking the system is greater than 50%.
4. A system based on the multi-agent system consensus attack method according to any one of claims 1-3, characterized in that, Includes the following modules: The building block is used to construct a multi-agent fault-tolerant and consistent system model; The processing module is used to extract two suitable nodes for attack from the constructed system model through a CNN network. The attack module is used to attack two selected nodes and update them using a consistency attack algorithm; The iteration module is used to perform interactive iterative processing on all the processed agents, which causes the entire system model to be split and unable to reach a consensus.