A resilient distributed parameter estimation method in multi-attack adversarial networks

Abstract

The present application belongs to the field of signal processing, and specifically provides an elastic distributed parameter estimation method in a multi-attack adversarial network, to solve the problem that the existing elastic distributed parameter estimation method cannot normally detect FDI attacks and deteriorate the performance of the attacked nodes in the multi-attack adversarial network. The present application proposes a node classification method and a diffusion state identification method based on correlation entropy in a distributed framework, to realize the detection of FDI attacks; at the same time, an elastic strategy based on the dLMS algorithm is proposed, and the selection of the reference node is optimized by minimizing the network steady-state mean square error, to improve the distributed parameter estimation performance in the multi-attack adversarial network; compared with the existing elastic security distributed parameter estimation method, the present application can achieve higher estimation accuracy in the network with relatively more attacked nodes.

Description

Technical Field

[0001] This invention belongs to the field of signal processing, and relates to the parameter estimation problem in signal processing, and further to the problem of resilient distributed parameter estimation in multi-attack adversarial networks. Specifically, it provides a method for resilient distributed parameter estimation in multi-attack adversarial networks. Background Technology

[0002] In sensor networks, distributed signal processing has become a research hotspot in the field of signal processing. Compared with traditional centralized methods, distributed strategies do not require a specific fusion center, which can improve the robustness and scalability of the network. They have been widely used in parameter estimation, target tracking, decision making and distributed clustering.

[0003] Distributed parameter estimation is one of the important research contents of distributed signal processing. The Diffusion Least Mean Square (dLMS) algorithm proposed in the literature "Cattivelli FS, Sayed AH. Diffusion LMS Strategies for Distributed Estimation[J].IEEE Transactions on Signal Processing,2010,58(3):1035-1048." is a commonly used distributed parameter estimation method. In this method, each node in the network can communicate with its directly connected nodes (i.e., neighbors) to cooperate in completing the estimation task. dLMS uses the ATC diffusion strategy. Each node uses its own measurement data to update the intermediate estimate, and then processes the intermediate estimates of its neighbors according to some combination rules (such as the Metropolis criterion) to further update the local estimate, thereby improving the performance of distributed parameter estimation. However, dLMS can only obtain the desired estimate in a secure network, i.e., a network free from attacks during measurement processing and network communication. In reality, since sensor nodes are placed in public environments, wirelessly transmitted data is easily captured and eavesdropped on, thus the network may be attacked, forming an adversarial network. Nodes that are attacked are called attacked nodes, and nodes that are not attacked are called secure nodes. In adversarial networks for distributed parameter estimation, attackers may inject malicious observation data into some attacked nodes during the measurement process, causing the network to converge to an incorrect estimate. This type of attack is called False Data Injection (FDI) attack.

[0004] To reduce the negative impact of FDI attacks on the performance of distributed parameter estimation, some literature has studied corresponding resilient (i.e., resistant to attacks, adaptable, and capable of rapid recovery) secure distributed parameter estimation methods to improve the accuracy of distributed parameter estimation in adversarial networks. Based on the data used to detect attacks, existing methods can be divided into two categories. The first category uses intermediate estimates to detect FDI attacks. The paper "Liu Y, Li C. Secure Distributed Estimation Over Wireless Sensor Networks Under Attacks[J].IEEE Transactions on Aerospace and Electronic Systems, 2018." proposes a secure dLMS (Secure dLMS, S-dLMS) method. Each node sorts the intermediate estimates obtained by neighboring nodes from non-cooperative estimation, selects the intermediate estimate in the middle position as the reference estimate, and constructs a threshold to detect trusted neighbors (i.e., secure neighbors). The dLMS algorithm is then executed in combination with the intermediate estimates received from trusted neighbors. The second category uses the measurement data obtained by each node to directly detect FDI attacks. The paper "Y.Hua, F.Chen, S.Deng, S.Duan, L.Wang, Secure distributed estimation against false data injection" proposes a method to detect FDI attacks directly using the measurement data obtained by each node. The paper "Attack, Inf. Sci. 515 (2020) 248-262" proposes a dLMS (DLMSKL) method based on Kullback-Leibler divergence. Each node stores a set of measurement data with a fixed window size and uses these historical measurement data to approximate the distribution of the measurement data. Then, each node calculates the KL divergence between its own and its neighboring nodes and sorts them. By comparing the ratio of the KL divergences with a threshold, the attacked nodes are identified and trusted neighbors are obtained. For each attacked node, the node located in the middle of its trusted neighbors is used as a reference neighbor, and the malicious data of the attacked node is replaced with the data of the reference neighbor.

[0005] Both of the aforementioned methods rely on the crucial assumption that "the number of attacked neighbors for each node does not exceed half of its degree (i.e., the number of neighbors including itself)." However, for nodes with few neighbors in the network, the number of attacked neighbors can easily exceed half of its degree. In this case, neither of the above methods can properly detect FDI attacks and attacked nodes, leading to a deterioration in their estimation performance. Therefore, this invention refers to adversarial networks where the number of attacked neighbors for a node exceeds half of its degree as multi-attack adversarial networks, and further provides a resilient distributed parameter estimation method for multi-attack adversarial networks. Summary of the Invention

[0006] The purpose of this invention is to propose a resilient distributed parameter estimation method in multi-attack adversarial networks (MAVs) to solve the problem that existing resilient distributed parameter estimation methods cannot properly detect FDI attacks and attacked nodes in MAVs, leading to performance degradation.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A resilient distributed parameter estimation method in multi-attack adversarial networks, characterized by the following steps:

[0009] Step 1: Construct a linear measurement model using a distributed wireless sensor network:

[0010]

[0011] Where i represents the current time, k represents the node, and d k,i For scalar observations, w k,i For unknown parameters that have been tampered with by a fake data injection attack, u k,i Let v be a spatiotemporally independent and identically distributed regression vector with zero mean. k,i Zero-mean, spatiotemporally independent and identically distributed Gaussian white noise; Let w represent the set of nodes subjected to FDI attack at time i. o Let ε represent the true unknown parameter to be estimated. k,i The estimation error introduced by the attack;

[0012] Step 2: Calculate the estimated correlation entropy based on the historical measurements of node k and its neighbor node l:

[0013]

[0014] in, This indicates that node k does not include its own neighborhood, K represents the number of historical measurements, and λ k κ represents the forgetting factor of node k. σ (·) represents the Gaussian kernel;

[0015] The estimated correlation entropies of node k and its neighbor node l are sorted in ascending order to obtain the correlation entropy set:

[0016]

[0017] Among them, l (n) express The node index of the nth item in the array, where n = 1, 2, ..., n k -1;

[0018] Step 3: Use variance Λ k,i,α Evaluate the relevant entropy set The degree of fluctuation is determined, and a threshold θ is set based on the detection statistic I. k,i,α Divide the neighboring nodes of node k into two clusters; detect the statistics I. k,i,α for:

[0019]

[0020] Where H0 indicates that the first to the αth nodes belong to the first cluster. The assumption is that H1 represents the neighboring nodes of node k being divided into and The assumptions;

[0021] Step 4: Based on each neighbor node at the previous time i-1 Status identifier τ l,i-1 The computation node k considers the assumed state symbol of node l at the current time i.

[0022]

[0023] Where sgn(·) represents the sign function;

[0024] Node k and its neighboring nodes Perform the exchange of hypothetical state identifiers, and based on the received hypothetical state identifier... Update its own state identifier τ k,i :

[0025]

[0026] Step 5: Node k and its neighboring nodes By exchanging state identifiers, node k considers neighboring nodes with a state identifier of 1 as trusted neighbors, thus constructing a set of trusted neighbors:

[0027]

[0028] And obtain the number of trusted neighbors t k,i :

[0029] Step 6: Based on the measured value d at the current moment k,i Compared to the previous moment w k,i-1 Local estimate x k,i-1 Update intermediate estimate ψ k,i :

[0030]

[0031] Where, μ k This represents the step size at node k;

[0032] Step 7: If the state identifier of node k is -1 and its number of trusted neighbors is t k,i If the value is greater than 0, then a reference neighbor is selected from its set of trusted neighbors. Update the intermediate estimate of node k to the intermediate estimate of the reference neighbor node:

[0033] ψ k,i =ψ rk,i,i

[0034] Step 8: Node k and its neighboring nodes By interacting with intermediate estimates, the parameter w at the current time can be calculated. k,i Local estimate x k,i :

[0035]

[0036] Among them, c lk,i This represents the combined weight of neighbor node l at node k.

[0037] Furthermore, in step 3, the variance Λ k,i,η Specifically:

[0038]

[0039] in, express The average value of the first η-1 terms is removed from the total, specifically:

[0040]

[0041] Furthermore, in step 3, the threshold θ is selected as an appropriate value based on experience.

[0042] Furthermore, in step 7, the process for selecting reference neighbors is as follows:

[0043] Connect node k with its trusted neighbors The estimated related entropies are arranged in ascending order, and the related entropy set is...

[0044]

[0045] Select relevant entropy set Node l in the middle position (p) As a reference neighbor node r k,i :r k,i =l (q) ,

[0046] Furthermore, in step 7, the process for selecting reference neighbors is as follows:

[0047] Calculate the power parameter γ of node k k : R is the variance of zero-mean, spatiotemporally independent and identically distributed Gaussian white noise. u,k Let tr{·} be the covariance of spatiotemporally independent and identically distributed regression vectors with zero mean, and let tr{·} denote the trace of the matrix; node k and its trusted neighbors are nodes. Perform power parameter γ k The interaction, and the received Arrange in ascending order to obtain the power parameter set:

[0048]

[0049] Select the first node l in the power parameter set. (1) As a reference neighbor r k,i :r k,i =l (1) .

[0050] Furthermore, in step 8, the combined weight c of neighbor node l at node k... lk,i Specifically:

[0051]

[0052] The beneficial effects of this invention are as follows:

[0053] This invention proposes a resilient distributed parameter estimation method in multi-attack adversarial networks, which has the following advantages:

[0054] 1. This invention proposes a node classification method based on correlation entropy. Based on a distributed framework, this method can accurately classify the neighbors of each node into attacked nodes and safe nodes according to the correlation entropy. Furthermore, this invention proposes a state recognition method based on a diffusion strategy. This method allows each node in the network to combine state information from multi-hop neighbors to determine whether each node is under attack. Thus, this invention can realize the detection of FDI attacks and the identification of trusted neighbors in multi-attack adversarial networks.

[0055] 2. This invention proposes a resilient strategy based on the dLMS algorithm, which develops a new reference neighbor selection scheme by minimizing the steady-state mean square error of the network; in multi-attack adversarial networks, this strategy can further improve the estimation performance of the system under FDI attacks.

[0056] 3. Compared with other currently researched flexible and secure distributed parameter estimation methods, the method proposed in this invention can achieve higher estimation accuracy in multi-attack adversarial networks. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the resilient distributed parameter estimation method for multi-attack adversarial networks according to the present invention.

[0058] Figure 2 This is a schematic diagram of the topology of a distributed network in an embodiment of the present invention.

[0059] Figures 3-6 The figure shows the simulation results in an embodiment of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0061] This embodiment provides a resilient distributed parameter estimation method in multi-attack adversarial networks, the process of which is as follows: Figure 1 As shown, the specific steps include:

[0062] Step 1: Employ a distributed wireless sensor network with N nodes, where node k can only exchange information with its neighboring nodes; define the set of nodes k, including its own neighbors, as its neighborhood. and Let |k| represent the degree of node k (|·| represents the cardinality of the set); define the set. The subset that does not contain node k is

[0063] Construct a measurement model that considers attacks that include fake data:

[0064]

[0065]

[0066] Where i represents the current time, k represents the node, and d k,i Represents a scalar observation; w k,i This indicates unknown parameters that have been altered by a fake data injection attack; The regression vectors are spatiotemporally independent and identically distributed with zero mean, and follow a dynamic regression model u. k,i =[u k,i ,u k,i-1 ,...,u k,i-M+1 ] T Covariance is v k,i It is zero-mean, spatiotemporally independent and identically distributed Gaussian white noise with variance . Assume that the regression vector of all nodes at each time step is independent of the Gaussian white noise;

[0067] This represents the set of nodes that are attacked by FDI at time i. This represents the true unknown parameter to be estimated. The estimation error introduced by the attack, Let M represent an M-dimensional column vector in the real number field, where M represents the dimension of the vector;

[0068] Step 2: Since there are differences in the distribution of historical measurements between each secure node and the attacked node, these differences can be determined based on node k and its neighboring nodes l. The assessment is based on the estimated correlation entropy between historical measurements, specifically:

[0069]

[0070] in, Let K represent the estimated correlation entropy between node k and its neighbor node l, where K represents the number of historical measurements, and λ represents the number of historical measurements. k κ represents the forgetting factor of node k. σ (·) represents the Gaussian kernel:

[0071] σ represents the width of the Gaussian kernel;

[0072] Furthermore, by sorting the estimated correlation entropies between node k and its neighboring node l in ascending order, we obtain the correlation entropy set for node k:

[0073]

[0074] in, l (n)Indicates corresponding to The node index of the nth item in the array, where n = 1, 2, ..., n k -1;

[0075] Step 3: For each node, if there are attacked neighbors, then the relevant entropy set is... There will be significant fluctuations; therefore, the variance Λ is used. k,i,η To evaluate The degree of fluctuation is as follows:

[0076]

[0077] in, express The average value of the first η-1 terms is removed from the total, specifically:

[0078]

[0079] Furthermore, a threshold θ is constructed for detection. Are there significant fluctuations? Further, distinguish between attacked neighbor clusters and secure neighbor clusters; detection statistics I. k,i,α The expression is:

[0080]

[0081] Where H0 indicates that the first to the αth nodes belong to the first cluster. The assumption is that H1 represents the neighboring nodes of node k being divided into and The assumptions are as follows; the threshold θ is selected based on experience to find an appropriate value.

[0082] Step 4: For each node, set the status identifier τ k,i ∈{-1,0,1} represents its detection state at time i, specifically:

[0083]

[0084] It should be noted that the symbol τ k,i This is not necessarily the true state of node k; that is, the attacked node may be mistakenly identified as a safe node, i.e., the state symbol τ. k,i =1; Initialize the state identifier of all nodes in the network to τ. k,i =0;

[0085] Each node uses a majority voting mechanism to determine the cluster. The state identifiers for β = 1 and 2 are as follows:

[0086] Each neighbor node obtained in the previous time step i-1 The state identifiers are aggregated, and the assumed state identifier of node l at the current time i is obtained by node k according to the following formula.

[0087]

[0088] Where sgn(·) represents the sign function;

[0089] Node k and its neighboring nodes Perform the exchange of hypothetical state identifiers, and based on the received hypothetical state identifier... Update its own state identifier τ k,i :

[0090]

[0091] Step 5: Node k and its neighboring nodes By exchanging state identifiers, node k considers neighboring nodes with a state identifier of 1 as trusted neighbors, thus constructing a set of trusted neighbors:

[0092]

[0093] and, This represents the number of trusted neighbors of node k at time i;

[0094] Step 6: Based on the measured value d at the current moment k,i Compared to the previous moment w k,i-1 Local estimate x k,i-1 Update intermediate estimate ψ k,i :

[0095]

[0096] Where, μ k This represents the step size at node k;

[0097] Step 7: If the state identifier of node k is -1 and its number of trusted neighbors is t k,i If the value is greater than 0, then a reference neighbor is selected from its instantaneous trusted neighbor set. To provide a reliable estimate, that is, the intermediate estimate of node k is updated to the intermediate estimate of the corresponding reference neighbor node, i.e.:

[0098] ψ k,i =ψ rk,i,i

[0099] The selection of reference neighbors can be done in the following two ways:

[0100] 1) Heuristic selection method: Match the attacked node k with its trusted neighbors The relevant entropies are arranged in ascending order, and the relevant entropy set is...

[0101]

[0102] Select relevant entropy set Node l in the middle position (p) As a reference neighbor node r k,i :r k,i =l (q) ,

[0103] 2) Optimization selection method: Calculate the power parameter γ of the attacked node k. k : tr{·} denotes the trace of the matrix; node k and its trusted neighbors are nodes. Perform power parameter γ k The interaction, and the received Arrange in ascending order to obtain the power parameter set:

[0104]

[0105] Select the first node l in the power parameter set. (1) As a reference neighbor r k,i :r k,i =l (1) ;

[0106] Step 8: Node k and its neighboring nodes By interacting with intermediate estimates, the parameter w at the current time can be calculated. k,i Local estimate x k,i :

[0107]

[0108] Among them, c lk,i The combined weight of neighbor node l at node k is represented as follows:

[0109]

[0110] The following constructs a system model with an attack scenario, comparing the method of this invention (CDSD-T) with a classic dLMS algorithm proposed in the literature "Cattive lli FS, Sayed AH. Diffusion LMS Strategies for Distributed Estimation[J].IEEE Transactions on Signal Processing,2010,58(3):1035-1048.", the S-dLMS algorithm proposed in the literature "Liu Y, Li C. Secure Distributed Estimation Over Wireless Sensor Networks Under Attacks[J].IEEE Transactions on Aerospace and Electronic Systems,2018.", and the DLMSKL proposed in the literature "Jia X C. Resource-efficient and secure distributed state estimation over wireless sensor networks: a survey[J]. International Journal of Systems Science,2021,52.", to illustrate the feasibility and superiority of the method of this invention.

[0111] The simulation test conditions are as follows:

[0112] 1) Signal Model: A distributed network with N nodes is adopted, and the network topology is as follows: Figure 2 As shown; consider the measurement model that includes fake data in the attack:

[0113]

[0114] in,

[0115]

[0116] Furthermore, the regression vector u k,i It originates from a zero-mean Gaussian function with a variance of 1;

[0117] 2) Parameter settings: There are 21 sensors in the network. The initial sensor estimate is x0 = [0,0,0,0]. T Unknown parameter w o =[0.5,1,1.5,0.1] T Noise variance of each node Attack strength Ξk,i =||ε k,i || / ||w o || = 4, and the step size μ is set uniformly throughout the network. k =0.02 and forgetting factor λ k =0.97, the kernel width used in the Gaussian kernel function is σ=4, the threshold is set to θ=4, the dimension of the historical data used by each node is K=50, and all simulation experiments are the average results of Z=200 independent repeated experiments;

[0118] To evaluate the estimation performance, the transient network MSD is obtained by averaging the mean squared deviation (MSD) of all nodes at time i, defined as:

[0119]

[0120]

[0121] Where, x k,i,ζ It is the independent experiment of node k at time i with respect to the unknown parameter w. o The estimation results are as follows. The steady-state MSD performance is obtained by averaging the last 200 transient MSD samples at each node k.

[0122] When the network is subjected to relatively few attacks, that is, the number of attacked neighbors is less than the degree n of each node k. k When the transient network MSD of the above algorithm is half, the simulation results are as follows: Figure 3 As shown, the steady-state MSD performance is as follows: Figure 4 As shown;

[0123] When a network is subjected to relatively many attacks, that is, when the number of attacked neighbors in the network is greater than the degree n of node k. k When the transient network MSD of the above algorithm is half, the simulation results are as follows: Figure 5 As shown, the steady-state MSD performance is as follows: Figure 6 As shown;

[0124] Figure 3 , 4 This demonstrates that the methods of the present invention (i.e., CDSD-H and CDSD-F, where CDSD-H represents selecting reference neighbors based on a heuristic selection method, and CDSD-F represents selecting reference neighbors based on an optimized method) can improve the estimation accuracy in networks with relatively few attacked nodes compared to the inelastic dLMS, and approach the estimation accuracy of existing resilient secure distributed parameter estimation algorithms. Figure 5 , 6This demonstrates that the methods of this invention (i.e., CDSD-H and CDS DF) exhibit lower MSD in steady state compared to other algorithms mentioned above in networks with a relatively large number of attacked nodes. This implies that the methods of this invention significantly outperform other algorithms mentioned above in estimating unknown parameters in networks with a relatively large number of attacked nodes; furthermore, in networks with a relatively large number of attacked nodes, CDSD-F exhibits a lower MSD in steady state compared to CDSD-H.

[0125] Furthermore, such as Figure 3 , 4 The results shown in 5 and 6 indicate that the method of the present invention can effectively estimate unknown parameters in adversarial networks, whether the number of attacked nodes is relatively small or large.

[0126] The above description is merely a specific embodiment of the present invention. Any feature disclosed in this specification may be replaced by other equivalent or similar features unless otherwise specified. All disclosed features, or steps in all methods or processes, may be combined in any way except for mutually exclusive features and / or steps.

Claims

1. A resilient distributed parameter estimation method in multi-attack adversarial networks, characterized in that, Includes the following steps: Step 1: Construct a linear measurement model using a distributed wireless sensor network: Where i represents the current time, k represents the node, and d k,i For scalar observations with real values, w k,i For unknown parameters that have been tampered with by a fake data injection attack, u k,i Let v be a spatiotemporally independent and identically distributed regression vector with zero mean. k,i Zero-mean, spatiotemporally independent and identically distributed Gaussian white noise; Let w represent the set of nodes subjected to FDI attack at time i. o Let ε represent the true unknown parameter to be estimated. k,i The estimation error introduced by the attack; Step 2: Calculate the estimated correlation entropy based on the historical measurements of node k and its neighbor node l: in, This indicates that node k does not include its own neighborhood, K represents the number of historical measurements, and λ k κ represents the forgetting factor of node k. σ (·) represents the Gaussian kernel; The estimated correlation entropies of node k and its neighbor node l are sorted in ascending order to obtain the correlation entropy set: Among them, l (n) express The node index of the nth item in the array, where n = 1, 2, ..., n k -1; Step 3: Use variance Λ k,i,α Evaluate the relevant entropy set The degree of fluctuation is determined, and a threshold θ is set based on the detection statistic I. k,i,α Divide the neighboring nodes of node k into two clusters; detect the statistics I. k,i,α for: Where H0 indicates that the first to the αth nodes belong to the first cluster. The assumption is that H1 represents the neighboring nodes of node k being divided into and The assumptions; Step 4: Based on each neighbor node at the previous time i-1 Status identifier τ l,i-1 The computation node k considers the assumed state symbol of node l at the current time i. Where sgn(·) represents the sign function; Node k and its neighboring nodes Perform the exchange of hypothetical state identifiers, and based on the received hypothetical state identifier... Update its own state identifier τ k,i : Step 5: Node k and its neighboring nodes By exchanging state identifiers, node k considers neighboring nodes with a state identifier of 1 as trusted neighbors, thus constructing a set of trusted neighbors: And obtain the number of trusted neighbors t k,i : Step 6: Based on the measured value d at the current moment k,i Compared to the previous moment w k,i-1 Local estimate x k,i-1 Update intermediate estimate ψ k,i : Where, μ k This represents the step size at node k; Step 7: If the state identifier of node k is -1 and its number of trusted neighbors is t k,i If the value is greater than 0, then a reference neighbor is selected from its set of trusted neighbors. Update the intermediate estimate of node k to the intermediate estimate of the reference neighbor node: ψ k,i =ψ rk,i,i Step 8: Node k and its neighboring nodes By interacting with intermediate estimates, the parameter w at the current time can be calculated. k,i Local estimate x k,i : Among them, c lk,i This represents the weight used at node k with respect to its neighbor node l.

2. The resilient distributed parameter estimation method in multi-attack adversarial networks as described in claim 1, characterized in that, In step 3, the variance Λ k,i,η Specifically: in, express The average value of the first η-1 terms is removed from the total, specifically:

3. The resilient distributed parameter estimation method in multi-attack adversarial networks as described in claim 1, characterized in that, In step 7, the process of selecting reference neighbors is as follows: Connect node k with its trusted neighbors The estimated related entropies are arranged in ascending order, and the related entropy set is... Select relevant entropy set Node l in the middle position (p) As a reference neighbor node r k,i :r k,i =l (q) , 4. The resilient distributed parameter estimation method in multi-attack adversarial networks as described in claim 1, characterized in that, In step 7, the process of selecting reference neighbors is as follows: Calculate the power parameter γ of node k k : R is the variance of zero-mean, spatiotemporally independent and identically distributed Gaussian white noise. u,k Let tr{·} be the covariance of spatiotemporally independent and identically distributed regression vectors with zero mean, and let tr{·} denote the trace of the matrix; node k and its trusted neighbors are nodes. Perform power parameter γ k The interaction, and the received Arrange in ascending order to obtain the power parameter set: Select the first node l in the power parameter set. (1) As a reference neighbor r k,i :r k,i =l (1) .

5. The resilient distributed parameter estimation method in multi-attack adversarial networks as described in claim 1, characterized in that, In step 8, the combined weight c of neighbor node l at node k lk,i Specifically:

Images