Node protection method based on complex network task dependent dynamic attack strategy

Abstract

The invention discloses a node protection method based on a complex network task dependent dynamic attack strategy. The method comprises the following steps: S1, constructing a multi-level graph neural network; in the multilevel graph neural network, nodes are traffic hubs and intersections; s2, constructing a reinforcement learning model; s3, constructing a deep Q network model DQN based on the multi-level graph neural network and the reinforcement learning model; s4, acquiring a key node sequence by using a deep Q network model; s5, calculating a network failure scale when the key nodes are removed in the step S4 and a risk assessment result of each node, and obtaining priority nodes in combination with the key node sequence; and when a traffic fault occurs, carrying out protection strategy adjustment on the priority node. The problems that in an existing method, complex task dependency cannot be effectively handled, dynamic attack response is insufficient, and calculation complexity is high are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of traffic network, and in particular relates to a node protection method based on a complex network task-dependent dynamic attack strategy. Background Art

[0002] Most of the disruption methods in existing research are based on attacks, aiming to significantly reduce the size of the largest connected subgraph. Although in many cases the size of the largest connected subgraph is closely related to the effective operation of the network, this is not always true. For example, on some networks, disruption strategies that rely on specific tasks may cause more damage than strategies based solely on topological criteria. Taking the power grid as an example, a more effective attack strategy may be suggested based on the power grid rules: removing the nodes and / or edges that are most critical to the stability of the network. In this case, many methods mentioned in existing studies are not applicable, so new theoretical and computational techniques are urgently needed to functionally disrupt the network to protect the network and take into account various factors associated with dynamics.

[0003] At the core of complex networks are key players, also known as influential players, important players, or critical actors. These players represent certain nodes, edges, or substructures that, when removed, significantly degrade a network’s specific functionality. The importance of identifying these key players has far-reaching implications across a variety of fields.

[0004] To identify key players in complex networks, many techniques and methods have been developed over the years, each with its own strengths and weaknesses. As the size of the network increases, these methods quickly become computationally intensive, limiting their practical application on large-scale networks.

[0005] The existing methods of traffic network disruption are mostly based on simple topological structure attacks, ignoring the importance of different nodes and edges in the network in actual tasks or functions. For example, simply removing highly connected nodes or edges, especially in traffic networks with strong functional dependencies, cannot effectively disrupt traffic network functions in all cases. At the same time, most disruption strategies only focus on static structures, while ignoring the roles of nodes or edges in specific tasks or functions and the dependencies between them. Such strategies may lead to inefficient or erroneous disruption methods in complex traffic networks, and fail to accurately identify key nodes and edges, and thus fail to effectively protect these key nodes, making the traffic network vulnerable and wasting traffic resources. Summary of the invention

[0006] The present invention provides a node protection method based on a complex network task-dependent dynamic attack strategy to solve the problems existing in the existing methods of being unable to effectively identify key nodes in the traffic network, being unable to effectively protect key nodes, and wasting traffic resources.

[0007] The technical solution adopted by the present invention is:

[0008] The node protection method based on complex network task-dependent dynamic attack strategy includes the following steps:

[0009] S1: Construct a multi-level graph neural network; the nodes are transportation hubs and intersections;

[0010] S2: Build a reinforcement learning model;

[0011] S3: Construct a deep Q network model DQN based on a multi-level graph neural network and reinforcement learning model;

[0012] S4: Use the deep Q network model to obtain the key node sequence;

[0013] S5: Calculate the network failure scale and risk assessment results of each node when the key nodes are removed in S4, and obtain the priority nodes in combination with the key node sequence; when a traffic failure occurs, adjust the protection strategy of the priority nodes.

[0014] Further, the S1 includes:

[0015] S11: In a multi-level graph neural network, the calculation formula for node feature update is:

[0016]

[0017] in, is the feature vector of node i in layer l; represents the feature vector of node j at layer l, and nodes i and j represent transportation hubs and intersections; is the set of neighbor nodes of node i, and node j is Any node in d i ,d j are the degrees of node i and node j respectively; W l is the weight matrix of the lth layer of the multi-level graph neural network; σ is the activation function; Represents the feature vector of node i at layer l+1;

[0018] W l The calculation formula is:

[0019]

[0020] f l =φl ·δ l ·η l

[0021] Where N represents the total number of layers of the multi-level graph neural network, f l is the eigenvalue of the lth layer, φ l represents the characteristic factor; δ l represents the demand factor; η l Represents the characteristic factor.

[0022] Furthermore, the S1 further includes:

[0023] S12: In each subgraph neural network, the node feature calculation formula is:

[0024]

[0025] Among them, A ij Indicates the connection between node i and node j in the adjacency matrix; if there is a connection, A ij =1, otherwise A ij =0, b l is the bias term;

[0026] S13: The time dimension is introduced into the multi-level graph neural network, and the update formula of the node features at each moment is:

[0027]

[0028] Among them: RNN represents recurrent neural network, Represents the feature vector of node i in the l-th layer of the graph neural network at time t+1, The feature vector representing node i in the l-th layer of the graph neural network at time t is given by Obtained through RNN processing; A represents the adjacency matrix composed of the connection relationships between all nodes at the current time t.

[0029] Further, the S2 includes:

[0030] S21: In the reinforcement learning model, state G t The expression is as follows:

[0031]

[0032] Among them, G t is the state at time t, represents the feature vector of the first node in layer l at time t, Represents the feature vector of the second node in layer l at time t, Represents the feature vector of the i-th node in the l-th layer at time t, obtained by the formulas in S11-S13;

[0033] In state G t Add node degree, betweenness centrality, and closeness centrality to the state G t Transformed to enhanced state, enhanced state s t The expression is:

[0034] s t =(G t ,topological features)

[0035] Among them, s t is the enhanced state at time t, and topological features are the topological features of each node, that is, the comprehensive score of each node calculated based on node degree, betweenness centrality, and closeness centrality;

[0036] S22: Construct the state, action, and reward functions of the reinforcement learning model;

[0037] The state is the current state s t ; The action is to remove the node; The reward is the degree of decrease in road network efficiency after the removal;

[0038] The state update expression is as follows:

[0039] s t+1 =f(s t ,a t )

[0040] Where f is the state update function, s t+1 Indicates execution of action a t That is, the new state after removing the node;

[0041] The reward calculation formula is:

[0042] r t =μR(s t ,a t )=-μΔE net(t)

[0043] Where, ΔE net(t) is the change in road network efficiency after removing the node at time t, μ is the time decay factor, and R is the immediate reward function, which means that the execution of action a t Impact on road network efficiency, r t is the reward obtained at time t.

[0044] Further, the S3 includes:

[0045] S31: In the deep Q network model DQN, the Q value update formula is as follows:

[0046] Q1(s t ,a t ; θ) = r t +γmaxQ(s t+1 ,a′;θ)

[0047] Where γ is the discount factor; Q1 is the state s t Next, perform action a t The Q value of s represents the updated Q value, and a′ is the new state s t+1 One of all possible actions, maxQ(s t+1 ,a′;θ) means to select action a′ in the new state so that Q(s t+1 ,a′;θ) has the largest value, θ is the parameter of the deep Q network model; r t is in t Execute a t Immediate rewards after

[0048] S32: According to the current state s t , select the optimal removal strategy a t , the calculation formula is:

[0049] π(s t )=argmax(ζQ1(s t ,a t ;θ))

[0050] Among them, π represents the selection operation, and argmax represents selecting the action a that can maximize the Q1 value among all possible actions. t ;ζ represents the adjustment coefficient;

[0051] S33: For a t , use greedy selection ε, that is, there is a probability of ε to randomly select a new action, and there is a probability of 1-ε to select the action with the largest Q1 value;

[0052] S34: The loss function of the deep Q network model DQN is as follows:

[0053]

[0054] Among them, (s t ,a t ,r t ,s t+1 ) refers to the current state s in each iteration process t , perform action a t , Instant Rewards t , new state t+1 The samples composed of; D is the sample set in the experience replay pool; Refers to the sample set drawn from the experience replay pool; Q(st ,a t ; θ) represents the current Q function.

[0055] Further, the S4 includes:

[0056] S41: Initialize GNN, Q value, and experience replay pool; preset the maximum number of iterations;

[0057] S42: Through S3, at each time step t, according to the current state s t Select action a t , perform action a t , instant reward t and transfer to the new state s t+1 ; will (s t ,a t ,r t ,s t+1 ) is stored as a sample in the experience replay pool;

[0058] S43: Repeat S42 continuously;

[0059] S44: Randomly select samples from the experience replay pool and use the loss function Update the parameters θ of the deep Q network model DQN to obtain the trained DQN model;

[0060] S45: Use the trained deep Q network model DQN and iterate continuously using S31-S33 until the maximum number of iterations is reached; obtain the optimal collapse operation and obtain the key node sequence.

[0061] Further, the S5 includes:

[0062] S51: Select the relative size LCC of the largest connected subgraph as a measure of the network failure scale;

[0063] The relative size of the largest connected subgraph LCC is defined as follows:

[0064]

[0065] Among them, n′ is the number of nodes in the largest connected subgraph after the network is attacked, and n represents the total number of nodes when it is not attacked;

[0066] S52: Calculate the network failure scale in each iteration in S45;

[0067] S53: Calculate the state vector of each node in the deep Q network model DQN and its risk assessment results; including:

[0068] S531: Calculate the state of each node using the following formula:

[0069] S i (t) = w i (t)·M i (t)

[0070]

[0071] Among them, S i (t) is the state vector of node i at time t; w i (t) represents the weight of node i at time t; a ij Indicates the connection between nodes i and j; if there is a connection, a ij =1, otherwise a ij =0;d i is the degree of node i; is the set of neighbor nodes of node i; M i (t) is the performance index of node i at time t, i.e., node flow;

[0072] S532: Calculate the risk assessment result of the node using a logistic regression algorithm, which is expressed by the following formula:

[0073]

[0074] Among them, R i (t) is the risk assessment result of node i at time t, g is the logistic regression algorithm, β0,β1,…,β n+1 is the parameter of the logistic regression model; S i (t-1) represents the state vector of node i at time t-1, S i (tn) represents the state vector of node i at time tn.

[0075] Furthermore, the S5 further includes:

[0076] S54: Obtaining priority nodes and dynamically adjusting protection strategies for the priority nodes; including:

[0077] S541: Determine the order of nodes according to the size of the network failure scale in S52, order the nodes according to the node risk assessment results obtained in S532, and combine the key node sequences obtained in S4 to find the nodes in the first 10% to 15% of the three sequences as priority nodes.

[0078] Furthermore, the S5 further includes:

[0079] S542: Obtain a protection strategy, which is expressed by the following formula:

[0080]

[0081] Among them, Pi (t) is the protection strategy of node i at time t, is the set of neighbor nodes of node i, j represents Any node in R j (t) represents the risk assessment result of node j at time t, represents the protection strategy of node j at time t, that is, setting a backup path at node j, s j (t), d j (t) represents the state and node degree of node j, and ε is the weight coefficient;

[0082] S543: Dynamically adjust the protection strategy for the priority node, as expressed by the following formula:

[0083]

[0084] Among them, ξ is the protection level, k represents the policy dynamic adjustment function, P i (t) represents the risk assessment result at the current time node i, S i (t+1) represents the state of the node at time t+1; P i (t+1) represents the dynamically adjusted protection strategy of node i at time t+1, and ΔP represents the protection action adjustment, that is: S i (t+1) When the protection level is exceeded, that is, when a traffic failure occurs, the traffic lights and signal controllers at the priority nodes are set as priority protection, the traffic data collection equipment at the priority nodes are set as priority protection, the traffic signal cycle at the priority nodes is optimized, and the vehicles at the priority nodes are guided to detour.

[0085] Beneficial effects of the present invention:

[0086] 1. Identification advantage based on task dependencies. Most existing technologies only evaluate the importance of nodes based on their topological properties (such as degree, betweenness, clustering coefficient, etc.), ignoring the dependencies of each node in the network on actual tasks or functions. This strategy can easily lead to inefficient identification of key nodes in transportation networks. The present invention takes into account the specific tasks undertaken by nodes and edges in the network and their dependencies, and can more accurately identify nodes and edges that are critical to the overall transportation network function during disruption or attack, thereby providing a more effective protection solution.

[0087] 2. Real-time detection and response capabilities of dynamic attack strategies. Existing methods mainly detect and protect against attacks on static network structures, and cannot adapt to dynamic attacks in the network, especially when the attacks change over time. Such attacks are often highly concealed. The present invention can monitor traffic failures in the traffic network in real time. By introducing task-dependent strategies, the system can dynamically adjust the detection and protection schemes and respond quickly to traffic failures, thereby improving network security.

[0088] 3. Improve network disruption effect and protection efficiency. Traditional methods focus more on disrupting the topological connectivity of the network, ignoring the dependencies between different tasks in the functional network, resulting in incomplete disruption of network functions or insufficient protection effect. In addition, the computational cost of existing technologies in large-scale networks is high and their application is limited. By considering functional task dependencies, the present invention can not only achieve accurate disruption of large-scale networks at a lower computing cost, but also effectively protect key functional nodes in the transportation network, improve protection efficiency, and is applicable to a wider range of network environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0090] Figure 1 It is a flow chart of the present invention.

[0091] Figure 2 It is a comparison diagram of the change in relative size of the largest connected subgraph during the failure process of the Sioux Falls network using the method of the present invention and other methods.

[0092] Figure 3 This is a comparison diagram of the changes in the relative sizes of the largest connected subgraphs during the failure process of the London multi-way traffic network using the method of the present invention and other methods. DETAILED DESCRIPTION

[0093] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0094] See also Figure 1This embodiment uses the London multi-way traffic network as an example to propose a protection method based on a complex network task-dependent dynamic attack strategy, with special attention paid to the task or function dependency in the network. Specifically, it includes the following steps:

[0095] S1: Based on Graph Neural Networks (GNN), a multi-level graph neural network is constructed; the node feature update with time series weighting is introduced to capture the dynamic node features of complex networks by iteratively aggregating the local neighborhood information of the nodes. The specific steps are described as follows:

[0096] S11: Formula for updating the node features of each layer of graph neural network (GNN):

[0097]

[0098] in, is the feature vector of node i at layer l, where nodes i and j represent transportation hubs and intersections;

[0099] The vector is obtained by aggregating the information of direct neighbor nodes and updating it with its own features, capturing the dynamic characteristics of the node and its surrounding neighborhood. represents the feature vector of node j in layer l, is the set of neighbor nodes of node i, and node j is Any node in d i and d j are the degrees of node i and node j respectively; W l is the weight matrix of the lth layer; σ is the activation function, namely the linear activation function ReLU.

[0100] represents the feature vector of node i at layer l+1, which is obtained by further aggregating the features of neighboring nodes based on the feature vector at layer l. The feature vector update of each layer aggregates information from more distant neighboring nodes. Therefore: Only contains information about node i and its direct neighbor nodes, It not only contains the information of direct neighbors, but also aggregates the information of the neighboring nodes of direct neighbor nodes, so the information scope is wider.

[0101] Based on the characteristic index f of each layer l To allocate layer weights, the weight of the lth layer can be expressed as:

[0102]

[0103] f l =φ l ·δ l ·ηl

[0104] Among them, N represents the total number of layers of multi-level GNN, f l is the eigenvalue of the lth layer, respectively composed of the characteristic factors φ of layer l l , demand factor δ l , utility factor η l OK. l Specific characteristics of the reaction layer such as flow rate, traffic capacity, etc.;δ l Reflects the importance of the layer, such as traffic demand; η l Reflects the utility and contribution of layer l in actual use.

[0105] The output of ReLU can only take two values: 0 or 1. When the input is less than 0, the output is 0; when the input is greater than 0, the output is 1.

[0106] The activation function (ReLU) is expressed as follows:

[0107]

[0108] S12: A multi-level graph neural network contains multiple sub-GNN models. Each sub-GNN extracts features of one level. The calculation formula of each sub-GNN is:

[0109]

[0110] A ij Indicates the connection between node i and node j in the adjacency matrix. If there is a connection, A ij =1, otherwise A ij = 0, which is used to determine which neighbor node j’s features node i should aggregate; The features of neighbor node j are After the weight matrix W l Perform linear transformation to extract features, W l is a trainable parameter. l is a bias term and a trainable parameter used to adjust the linear combination of node features.

[0111] Therefore, each layer of the sub-GNN is connected by the adjacency matrix A ij Determine the connections between nodes using trainable parameters W l and b l To update the feature representation of each node. Each layer of sub-GNN is only responsible for extracting feature information at a specific level. This structure enables the entire network to extract features at different levels in a hierarchical manner.

[0112] S13: The time dimension is introduced into the multi-level GNN, and the node features are dynamically updated using the timing information. The node feature update formula at each moment is:

[0113]

[0114] Among them: RNN (Recurrent Neural Network) is mainly used to process time series data. RNN updates node features at each moment, introduces the time dimension, and captures the dependencies in the time dimension by memorizing the previous state, so it is used here to update the time features of the node. GNN (Graph Neural Network) is mainly used to capture spatial features in data. The features of each node are updated by aggregating the information of the node and its neighbors, thereby learning the representation of the node. Therefore, GNN is used to update the spatial features at the current moment. Represents the feature vector of node i in the lth layer of GNN at time t+1, It represents the feature vector of node i in the l-th layer of GNN at time t, which is obtained by RNN processing; A represents the adjacency matrix, which represents the adjacency matrix composed of the connection relationships between all nodes at the current time t.

[0115] S2: Build a reinforcement learning model, optimize the collapse strategy based on reinforcement learning (RL), dynamically adjust the reward function of reinforcement learning based on network topology properties, gradually find the optimal node or edge removal order through trial and error and reward mechanism, and build a reinforcement learning environment. The specific steps are as follows:

[0116] S21: The reinforcement learning expression constructed is as follows:

[0117]

[0118] Among them, G t is the network state at time t, where each element Represents the feature vector of node i in layer l at time t, and the network state is obtained by aggregating the features of each node through GNN features.

[0119] In the state representation of reinforcement learning, network topology characteristics such as node degree, betweenness centrality, closeness centrality, etc. are added, and the enhanced state is used to train the reinforcement learning model. Its expression is updated as follows:

[0120] s t =(G t ,topological features)

[0121] Among them, s tIt is the network status after integrating topological indicators at time t. Topological features are the topological features of network nodes. According to different features such as node degree, betweenness centrality, and closeness centrality, the entropy weight method is used to calculate the comprehensive score of each node.

[0122] S22: Build a reinforcement learning environment, including states, actions, and reward mechanisms.

[0123] The state is the feature vector of the node in the current network; the action is to remove a node; the reward is the degree of decline in network function after removal.

[0124] The constructed status update expression is as follows:

[0125] s t+1 =f(s t ,a t )

[0126] Where f is the state update function, s t+1 Indicates execution of action a t That is, the new state after the node is removed.

[0127] Time-weighted reward function update.

[0128] In order to better describe the impact of time on rewards, the time decay factor μ is introduced, and the time-weighted reward function is:

[0129] r t =μR(s t ,a t )=-μΔE net(t)

[0130] Among them, ΔE net(t) is the change in network efficiency after removing the node at time t, μ is the time decay factor, where μ is 1. R is the immediate reward function, indicating the execution of action a t That is, the impact of removing a node on network function. t is the reward obtained at time t.

[0131] S3: Build a deep Q network (DQN) model, build a dynamic, multi-level DQN, each layer of DQN corresponds to a layer of the network, each layer of DQN updates the Q value of its layer and passes the result to the previous layer for collaborative optimization; add a time-weighted Q value update formula to ensure that the feedback after each attack can adjust the next strategy in real time; introduce a relational network in the Q value update process to learn the optimal disruption strategy.

[0132] S31: The Q value update formula is as follows:

[0133] Q1(s t ,at ; θ) = r t +γmaxQ(s t+1 ,a′;θ)

[0134] Where: Q1 is the state s t Next, perform action a t Q value, θ represents the weight parameter of the deep Q network (DQN) model. t ,a t ; θ) is the current state s t Execute action a t Finally, combine the immediate reward and the future state s t+1 The expected return is calculated by the formula.

[0135] maxQ(s t+1 ,a′;θ) means in the next state s t+1 Select the maximum Q value corresponding to the optimal action a′ in the next state s t+1 The maximum value that can be brought by choosing the best action a′ is the “greedy” part of the Q-value iteration algorithm, that is, always choosing the best action to update the Q-value.

[0136] a′ represents the state s t+1 The best action among all possible actions is the action that the model considers to be the best in the future. When the Q value is updated, the Q value corresponding to this best action is used to update the Q value at the current moment.

[0137] γ is a discount factor, which indicates the importance of future rewards;

[0138] s t is the state at the current moment, that is, the state at time t; s t+1 Is to perform action a t After that, the system enters the next state, that is, the state at time t+1.

[0139] r t is in t Execute a t Instant rewards after.

[0140] a t is the state s at time t t The specific actions performed under the model are used in the decision-making process. At time t, DQN makes a decision based on the current state s. t and the network's strategy selects action a tThis action is selected by the ∈-greedy strategy, using the greedy selection of the next action ε, that is, there is a probability of ε to randomly select a new action, and there is a probability of 1-ε to select the action with the largest Q value in the current state; that is, in most cases, the action with the largest Q value is selected, but occasionally an action is randomly selected to maintain exploration. The execution of this action affects the immediate reward r t and future states t+1 .

[0141] S32: Dynamic adjustment strategy:

[0142] According to the current network status t , select the optimal removal strategy a t :

[0143] π(s t )=argmax(ζQ1(s t ,a t ;θ))

[0144] Among them, π represents the selection of the optimal removal strategy, according to the current network state s t Select Removal Policya t argmax means choosing the action that maximizes the subsequent function value Q1. t ,a t ; θ)) represents the state s t Next select action a t The expected cumulative reward after . ζ represents the adjustment coefficient; ζ·Q1(s t ,a;θ) indicates that ζ is used to adjust the output value of the function. For example, when ζ>1, the Q value will be amplified, making the reward associated with the action more important; when ζ<1, the Q value will be reduced, reducing the weight of the action.

[0145] S33: Use greedy selection of the next action ε, that is, a new action is randomly selected with a probability of ε, and the action with the largest Q1 value in the current state is selected with a probability of 1-ε;

[0146] S34: Construct the loss function as follows:

[0147]

[0148] Among them, γ is the discount factor, and D is the sample set in the experience replay pool. Refers to the set of samples drawn from the experience replay pool for model training. s t Indicates the current state, a t Indicates action, r t Indicates reward, s t+1represents the next state, and these four elements together constitute a basic interaction loop in reinforcement learning. t ,a t ; θ) is the current Q value. Q1 is the target Q value, obtained according to S31; these two Q functions are used to calculate the loss, that is, the error between the current Q function and the target Q function, which is used to update the model's parameter θ, so that the current Q function is getting closer and closer to the real future return.

[0149] S4: Real-time monitoring and feedback, monitor changes in network functions, and adjust the DQN model based on feedback to improve the disruption effect. The specific steps are as follows:

[0150] S41: Initialize GNN, Q value, and experience replay pool; preset the maximum number of iterations;

[0151] S42: Through S3, at each time step t, according to the current state s t Select action a t , perform action a t , observe the reward r t and transfer to the new state s t+1 ; will (s t ,a t ,r t ,s t+1 ) is stored as a sample in the experience replay pool;

[0152] S43: Repeat S42 continuously;

[0153] S44: Randomly select samples from the experience replay pool and use the loss function Update the parameters θ of the deep Q network model DQN to obtain the trained DQN model;

[0154] S45: Use the trained deep Q network model DQN and iterate S31-S33 to obtain the optimal collapse operation and key node sequence.

[0155] S5: Sequentially protect nodes according to the size of network failures, adopt adaptive protection strategies and collaborative protection mechanisms, dynamically adjust protection strategies according to network status and attack trends, and introduce collaborative protection mechanisms between different levels and regions of the network to enhance the robustness of the overall network. The specific steps are described as follows:

[0156] S51: Select measurement indicators.

[0157] Select a metric, usually the largest connected component (LCC) or average path length or network robustness is selected as the metric. In this embodiment, LCC is selected as the metric;

[0158] S52: According to S45, the network failure scale in each iteration is calculated.

[0159] In a topological network, the relative size of the largest connected subgraph LCC is defined as follows:

[0160]

[0161] Among them, n′ is the number of nodes in the largest connected subgraph after the network is attacked, and n represents the total number of nodes when it is not attacked.

[0162] S53: Implement protection strategy: adaptive protection mechanism.

[0163] S531: Real-time monitoring of the status and performance indicators of each node in the network. It is expressed by the following formula:

[0164] S i (t) = w i (t)·M i (t)

[0165]

[0166] Among them, S i (t) is the state vector of node i at time t; a ij is the edge between nodes i and j; d i is the degree of node i; is the set of neighbor nodes of node i; M i (t) is the performance index of node i at time t, i.e., node flow, w i (t) represents the weight of node i at time t.

[0167] S532: Perform risk assessment, use machine learning algorithms to analyze monitoring data, and predict the attack or failure risks that the node may face. This is expressed by the following formula:

[0168]

[0169] Among them, R i (t) is the risk assessment result of node i at time t, g uses logistic regression to predict the risk that the node may face, β0,β1,…,β n+1 is the parameter of the logistic regression model. i (t-1) represents the state vector of node i at time t-1, S i (tn) represents the state vector of node i at time tn, which is used to track the risk changes of the node before time t.

[0170] S54: Find the priority node. For the priority node, dynamically allocate network resources, such as bandwidth, computing power, etc., to enhance the protection capability of the node.

[0171] S541: Determine the order of nodes according to the size of network failure in each iteration of S52, combine the key node sequence obtained in S4, and order the nodes according to the node risk assessment results obtained in S532, and find out the nodes in the first 10% to 15% of the three sequences as priority nodes.

[0172] S542: Obtain a protection strategy, which is expressed by the following formula:

[0173]

[0174] Among them, P i (t) is the protection strategy of node i at time t, j represents Any node in R j (t) represents the risk assessment result of node j at time t, represents the protection strategy of node j at time t, i.e. setting a backup path, s j (t), d j (t) represents the state and node degree of node j, ε is the weight coefficient, which adjusts the importance between the two and takes (0, 1). i (t) measures the overall protection effect or risk distribution of node i at time t.

[0175] S543: To adjust the strategy, the protection strategy needs to be automatically adjusted according to the real-time monitoring and risk assessment results. For example, when a node is detected to be abnormal, the node protection strategy can be automatically increased. It can be expressed by the following formula:

[0176]

[0177] Among them, ξ is the protection level, which can be set according to the actual situation. k in the formula represents the function of strategy adjustment, which is used to calculate the basis for adjusting the protection strategy. i (t),S i (t+1)) is a strategy adjustment function, which means that the risk assessment result P of the current time node i i (t) and the state S at the next time step i (t+1) to adjust the protection strategy. i (t+1) represents the protection strategy of node i at time t+1, which is optimized based on the prediction of the current state and future state. ΔP represents the protection action, that is: S i(t+1) When the protection level is exceeded, that is, when a traffic failure occurs, the traffic lights and signal controllers at the priority nodes are set as priority protection, the traffic data collection equipment at the priority nodes are set as priority protection, the traffic signal cycle at the priority nodes is optimized, and the vehicles at the priority nodes are guided to detour.

[0178] like Figure 2 and Figure 3 The process shown shows the change in the relative size of the largest connected subgraph of the network as the number of attacks increases. The method of the present invention is compared with the random attack method or the deliberate attack method, where Random is a random attack method and the method and are both deliberate attack methods. Compared with the attacks carried out by these common topological indicators, the LCC of the method of the present invention decreases fastest, so it can more effectively disintegrate the network. And in Figure 3 In the experiment, the method of the present invention not only has the fastest LCC drop rate, but also requires the least number of attacks when the network is completely disconnected. It also proves that the method of the present invention can more efficiently identify and remove key nodes when the network is attacked, thereby destroying the core functions of the network. Compared with the traditional topology-based disruption strategy, the present invention can more accurately destroy the task dependency of the network, resulting in greater connectivity loss, which demonstrates the superiority of the present invention in practical applications.

[0179] The present invention adopts multi-level feature aggregation technology, which enables the system to obtain feature information from different levels (such as local network structure and global topology), fully capture the dependencies between nodes, and effectively improve the network's representation ability. Combined with the reinforcement learning model optimization strategy, through continuous interaction with the environment, the optimal strategy is gradually learned, and the optimal protection strategy is adaptively selected according to the dynamic changes of the network, without relying on static fixed rules; its protection strategy has better adaptability and scalability, and can adapt to both small-scale local networks and large wide-area networks.

[0180] At the same time, by evaluating the health status and potential risks of nodes in real time, and calculating the risks based on the node status and performance, the protection strategy can be adjusted dynamically, which can effectively avoid resource waste and quickly adjust the protection strength when the network is attacked or under other adverse conditions, providing more efficient protection measures. The risk assessment model based on time series data can predict the possible failures or risks of nodes in the future by comparing the historical status of the node with the current status, which can effectively improve the system's prediction ability and response speed, and respond quickly when attacks or adverse conditions occur.

[0181] In addition, through the risk assessment function, the protection strategy can select the optimal strategy for each node's risk level. The system dynamically allocates resources based on the node's risk level and the overall vulnerability of the network. Priority nodes will receive more protection resources, while the remaining nodes will receive correspondingly reduced resource allocations, ensuring the rational allocation of resources and reducing the overall vulnerability of the network.

[0182] Therefore, by adopting multi-level feature aggregation technology and reinforcement learning model optimization strategy, the adaptability and scalability of the system are improved to cope with network environments of different scales. By evaluating the health status and risks of nodes in real time, the system can dynamically adjust the protection strategy, effectively avoid waste of resources, and provide more efficient protection measures when the network is attacked or under other adverse conditions. In addition, the optimization decision based on the risk assessment function can achieve a reasonable allocation of resources, thereby reducing the vulnerability of the overall network. In summary, based on this efficient process of identifying key nodes, the method of the present invention can more effectively protect priority nodes.

[0183] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0184] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A node protection method based on a complex network task-dependent dynamic attack strategy, characterized in that: The following steps are involved: S1: Construct a multi-level graph neural network; in the multi-level graph neural network, the nodes are transportation hubs and intersections; S2: Build a reinforcement learning model; S3: Construct a deep Q network model DQN based on a multi-level graph neural network and reinforcement learning model; S4: Use the deep Q network model to obtain the key node sequence; S5: Calculate the network failure scale and risk assessment results of each node when the key nodes are removed in S4, and obtain the priority nodes in combination with the key node sequence; when a traffic failure occurs, adjust the protection strategy of the priority nodes.

2. The node protection method based on complex network task-dependent dynamic attack strategy according to claim 1 is characterized in that: The S1 includes: S11: In a multi-level graph neural network, the calculation formula for node feature update is: in, is the feature vector of node i in layer l; represents the feature vector of node j at layer l, and nodes i and j represent transportation hubs and intersections; is the set of neighbor nodes of node i, and node j is Any node in d i d j are the degrees of node i and node j respectively; W l is the weight matrix of the lth layer of the multi-level graph neural network; σ is the activation function; Represents the feature vector of node i at layer l+1; W l The calculation formula is: f l =φ l ·d l ·or l Where N represents the total number of layers of the multi-level graph neural network, f l is the eigenvalue of the lth layer, φ l represents the characteristic factor; δ l represents the demand factor; η l Represents the characteristic factor.

3. The node protection method based on complex network task-dependent dynamic attack strategy according to claim 2 is characterized in that: The S1 further comprises: S12: In each subgraph neural network, the node feature calculation formula is: Among them, A ij Indicates the connection between node i and node j in the adjacency matrix; if there is a connection, A ij =1, otherwise A ij =0, b l is the bias term; S13: The time dimension is introduced into the multi-level graph neural network, and the update formula of the node features at each moment is: Among them: RNN represents recurrent neural network, Represents the feature vector of node i in the l-th layer of the graph neural network at time t+1, The feature vector representing the node i in the l-th layer of the graph neural network at time t is given by Obtained through RNN processing; A represents the adjacency matrix composed of the connection relationships between all nodes at the current time t.

4. The node protection method based on complex network task-dependent dynamic attack strategy according to claim 2 or 3, characterized in that: The S2 includes: S21: In the reinforcement learning model, state G t The expression is as follows: Among them, G t is the state at time t, represents the feature vector of the first node in layer l at time t, Represents the feature vector of the second node in layer l at time t, Represents the feature vector of the i-th node in the l-th layer at time t, obtained by the formulas in S11-S13; In state G t Add node degree, betweenness centrality, and closeness centrality to the state G t Transformed to enhanced state, enhanced state s t The expression is: s t =(G t ,topological features) Among them, s t is the enhanced state at time t, and topological features are the topological features of each node, that is, the comprehensive score of each node calculated based on node degree, betweenness centrality, and closeness centrality; S22: Construct the state, action, and reward functions of the reinforcement learning model; The state is the current state s t ; The action is to remove the node; The reward is the degree of decrease in road network efficiency after the removal; The state update expression is as follows: s t+1 =f(s t ,a t ) Where f is the state update function, s t+1 Indicates execution of action a t That is, the new state after removing the node; The reward calculation formula is: r t =μR(s t ,a t )=-μΔE net(t) Where, ΔE ne t(t) is the change in road network efficiency after removing the node at time t, μ is the time decay factor, and R is the immediate reward function, which means that the execution of action a t Impact on road network efficiency, r t is the reward obtained at time t.

5. The node protection method based on complex network task-dependent dynamic attack strategy according to claim 4 is characterized in that: The S3 includes: S31: In the deep Q network model DQN, the Q value update formula is as follows: Q1(s t ,a t ;θ)=r t +γmaxQ(s t+1 ,a′;θ) Where γ is the discount factor; Q1 is the state s t Next, perform action a t The Q value of s represents the updated Q value, and a′ is the new state s t+1 One of all possible actions, max Q(s t+1 ,a′;θ) means to select action a′ in the new state so that Q(s t+1 ,a′;θ) has the largest value, θ is the parameter of the deep Q network model; r t is in t Execute a t Immediate rewards after S32: According to the current state s t , select the optimal removal strategy a t , the calculation formula is: π(s t )=argmax(ζQ1(s t ,a t ;i)) Among them, π represents the selection operation, and arg max represents selecting the action a that can maximize the Q1 value among all possible actions. t ;ζ represents the adjustment coefficient; S33: For a t , use greedy selection ε, that is, there is a probability of ε to randomly select a new action, and there is a probability of 1-ε to select the action with the largest Q1 value; S34: The loss function of the deep Q network model DQN is as follows: Among them, (s t ,a t ,r t ,s t+1 ) refers to the current state s in each iteration process t , perform action a t , Instant Rewards t , new state t+1 The samples composed of; D is the sample set in the experience replay pool; Refers to the sample set drawn from the experience replay pool; Q(s t ,a t ; θ) represents the current Q function.

6. The node protection method based on complex network task-dependent dynamic attack strategy according to claim 5 is characterized in that: The S4 includes: S41: Initialize GNN, Q value, and experience replay pool; preset the maximum number of iterations; S42: Through S3, at each time step t, according to the current state s t Select action a t , perform action a t , instant reward t and transfer to the new state s t+1 ; will (s t ,a t ,r t ,s t+1 ) is stored as a sample in the experience replay pool; S43: Repeat S42 continuously; S44: Randomly select samples from the experience replay pool and use the loss function Update the parameters θ of the deep Q network model DQN to obtain the trained DQN model; S45: Use the trained deep Q network model DQN and iterate continuously using S31-S33 until the maximum number of iterations is reached; obtain the optimal collapse operation and obtain the key node sequence.

7. The node protection method based on complex network task-dependent dynamic attack strategy according to claim 6 is characterized in that: The S5 includes: S51: Select the relative size LCC of the largest connected subgraph as a measure of the network failure scale; The relative size of the largest connected subgraph LCC is defined as follows: Among them, n′ is the number of nodes in the largest connected subgraph after the network is attacked, and n represents the total number of nodes when it is not attacked; S52: Calculate the network failure scale in each iteration in S45; S53: Calculate the state vector of each node in the deep Q network model DQN and its risk assessment results; including: S531: Calculate the state of each node using the following formula: S i (t)=w i (t)·M i (t) Among them, S i (t) is the state vector of node i at time t; w i (t) represents the weight of node i at time t; a ij Indicates the connection between nodes i and j; if there is a connection, a ij =1, otherwise a ij =0;d i is the degree of node i; is the set of neighbor nodes of node i; M i (t) is the performance index of node i at time t, i.e., node flow; S532: Calculate the risk assessment result of the node using a logistic regression algorithm, which is expressed by the following formula: Among them, R i (t) is the risk assessment result of node i at time t, g is the logistic regression algorithm, β0,β1,…,β n+1 is the parameter of the logistic regression model; S i (t-1) represents the state vector of node i at time t-1, S i (tn) represents the state vector of node i at time tn.

8. The node protection method based on complex network task-dependent dynamic attack strategy according to claim 7 is characterized in that: The S5 further includes: S54: Obtaining priority nodes and dynamically adjusting protection strategies for the priority nodes; including: S541: Determine the order of nodes according to the size of the network failure scale in S52, order the nodes according to the node risk assessment results obtained in S532, and combine the key node sequences obtained in S4 to find the nodes in the first 10% to 15% of the three sequences as priority nodes.

9. The node protection method based on complex network task-dependent dynamic attack strategy according to claim 8 is characterized in that: The S5 further includes: S542: Obtain a protection strategy, which is expressed by the following formula: Among them, P i (t) is the protection strategy of node i at time t, is the set of neighbor nodes of node i, j represents Any node in R j (t) represents the risk assessment result of node j at time t, represents the protection strategy of node j at time t, that is, setting a backup path at node j, s j (t), d j (t) represents the state and node degree of node j, and ε is the weight coefficient; S543: Dynamically adjust the protection strategy for the priority node, as expressed by the following formula: Among them, ξ is the protection level, k represents the policy dynamic adjustment function, P i (t) represents the risk assessment result based on the current time node i, S i (t+1) represents the state of the node at time t+1; P i (t+1) represents the dynamically adjusted protection strategy of node i at time t+1, and ΔP represents the protection action adjustment, that is: S i (t+1) When the protection level is exceeded, that is, when a traffic failure occurs, the traffic lights and signal controllers at the priority nodes are set as priority protection, the traffic data collection equipment at the priority nodes are set as priority protection, the traffic signal cycle at the priority nodes is optimized, and the vehicles at the priority nodes are guided to detour.

Images