CN116346641A - A Control Plane Policy Verification Method Based on GNN Fault Model Learning - Google Patents

Abstract

The invention develops a control plane strategy verification method based on GNN fault model learning, which comprises the following steps: obtaining network topology information, dividing equivalence classes by using a K-Means clustering method, and generating a corresponding fault model; designing a GNN fault learning model based on a message passing mechanism; training a GNN fault learning model based on a message passing mechanism for each equivalence class; constructing a network policy diagram conforming to the network topology characteristics; preprocessing a strategy directed graph, extracting characteristic data, inputting the characteristic data into a trained GNN model, and generating an accurate network environment; and respectively carrying out network policy verification according to the network policy category. According to the invention, the situation that the actual network has faults is considered, the actual network environment corresponding to the network strategy to be verified is inferred by using the GNN auxiliary fault model, the equivalent class division is carried out by the K-Means clustering algorithm, the network scale is reduced, the verification speed is improved, and the data enhancement module is added, so that the trained GNN has certain generalization capability, and the multi-type network strategy verification is accurately and rapidly realized.

Description

Control plane strategy verification method based on GNN fault model learning

Technical Field

The invention relates to the technical field of network policy verification, in particular to a control plane policy verification method based on GNN fault model learning.

Background

The increase in network complexity may lead to various failures in the network, such as misconfigurations, software vulnerabilities, and malicious attacks. These faults may deviate the network from the policy objectives, thereby affecting its availability. However, the conventional manual failure removal method, for example, using Ping to detect the delay and the packet loss rate, or using Traceroute to detect the delay, the reachability and the packet loss rate, is not only inefficient, but also usually causes the network to stop for several hours, thereby causing serious loss. Therefore, how to prevent the occurrence of network faults and ensure the correct operation of the network becomes a fundamental problem faced by network operators and researchers.

Network policy validation techniques are used to solve this problem. Compared with the traditional manual fault removal mode, the network strategy verification technology not only can realize quick, real-time and automatic detection and improve the efficiency and accuracy of network strategy verification, but also improves the reliability and accuracy of verification results by combining a formal verification method and utilizing a mathematical model and algorithm. Currently, there are two main types of network policy verification technologies: data plane analysis and control plane verification. Data plane analysis is primarily used to infer policies that apply to a particular data plane, which can infer that all policies under that data plane are expected within seconds or minutes. Whereas control plane authentication is the authentication of a single policy under a particular network configuration. In order to enable the network policy validation technique to be applied to an actual network, it is necessary to model faults that may occur in the network with a fault model. However, if existing data plane analysis and control plane verification tools are deployed in a network that considers failure models, it is necessary to traverse all possible failure environments to determine an accurate environment that complies with network policies. Therefore, with the exponential increase of the network scale, the accurate network environment conforming to the network policy is difficult to find, the direct use of the data plane analysis or control plane verification tool can result in high computational complexity, slow verification speed, and the accuracy of the verification result can be reduced.

The network specifications with high accuracy and recall rate can be generated In a large network by analyzing historical iteration information and fault models, combining the advantages of both data plane verification methods and control plane verification methods, and switching back and forth In both data plane analysis and control plane verification methods using a phase predictor to obtain the optimal verification method, as a typical network policy verification tool taking into account the fault models, as mentioned In paper Birkner, r, draghsler-Cohen, d., vanbever, l., and vechev, m.t. (2020, february), "Config2Spec: mining Network Specifications from Network Configurations," In NSDI (pp. 969-984). However, the tool is still limited by the verification capabilities of existing data plane analysis and control plane verification methods, which are difficult to meet with the need for fast verification in large scale network topologies where network conditions are poor. Furthermore, config2Spec may produce false specifications because the network profile may contain inconsistent or erroneous information, which may affect the accuracy of the specifications generated by Config2 Spec.

Disclosure of Invention

Therefore, the present invention is directed to a control plane policy verification method based on GNN fault model learning, which is used for solving the technical problems mentioned in the background art. For large-scale complex dynamic networks, it is apparent that the reliability of the network policy verification results is reduced regardless of the existence of faults in the network. Therefore, the invention fully utilizes the graph representation learning capability of the GNN neural network and fully learns the network structure when faults exist by combining the fault model, thereby positioning an accurate fault environment for the network strategy and carrying out more accurate network strategy verification. In addition, since the GNN is an efficient neural network model, large-scale network data can be processed, which greatly improves the speed of network policy verification and helps to quickly discover network faults.

In order to achieve the above purpose, the present invention adopts the following technical scheme:

a control plane strategy verification method based on GNN fault model learning comprises the following steps:

s1, obtaining network topology information, dividing equivalence classes by using a K-Means clustering method, reducing the network scale, setting the maximum fault number in the network according to requirements, and generating a corresponding fault model;

s2, designing a GNN fault learning model based on a message passing mechanism, wherein the model comprises a linear transformation layer, a graph roll lamination layer and a full connection layer, inputting an adjacent matrix of network topology and a corresponding fault model, and outputting probability distribution of fault probability among feature nodes;

step S3, training a GNN fault learning model based on a message transmission mechanism for each equivalence class, and using the trained model for network policy verification;

step S4, collecting network policy information to be verified by an administrator, generating a network policy set according to requirements, and constructing a network policy diagram conforming to network topology characteristics;

s5, preprocessing a strategy directed graph, extracting feature data, inputting the feature data into a trained GNN model, acquiring probability distribution of fault probability among feature nodes, and generating an accurate network environment;

and step S6, respectively carrying out network policy verification according to the network policy category under the accurate network environment to obtain a final verification result.

Preferably, step S1 comprises the steps of:

step S1.1: obtaining network topology information, calculating node degree and degree centrality, realizing the division of equivalent classes by training a K-Means clustering model of the network topology through the node degree and degree centrality, selecting the optimal cluster number as the optimal class number of the equivalent classes by using an elbow rule, and generating a unique label and a topological graph corresponding to the equivalent class for each equivalent class;

step S1.2: setting the maximum fault number k in the network topology according to the requirement _max GeneratingMontecalo Monte Carlo simulation-based fault model = { env ₁ ,...env _q }, wherein env _q Representing each possible fault environment in which the fault probability P is calculated by Monte Carlo simulation _fail ；

Assuming that the probability of failure of one edge (u, v) of the topology in step S1.1 is p, then at k _max In the step, the fault probability P _fail Expressed in binomial distribution:

wherein the method comprises the steps of

Represented at k _max In the step, q (q is more than or equal to 0 and less than or equal to k) _max ) The number of combinations of secondary faults; probability of failure P _fail Expressed in the maximum failure number k _max In, one edge (u, v) fails.

Preferably, in step S2:

constructing a graph neural network GNN fault learning model, wherein the graph neural network GNN fault learning model comprises a linear transformation layer, a graph roll lamination layer and a full connection layer:

the linear transformation layer:

h _i ⁽¹⁾ ＝σ(W ⁽¹⁾ x _i +b ⁽¹⁾ ) (2)

wherein: h is a _i ⁽¹⁾ Is the feature vector of the hidden layer, σ is the ReLU activation function, W ⁽¹⁾ ∈R ^din×dout Is the weight matrix of the input layer, x _i ∈R ^din Is the input eigenvector of node i, b ⁽¹⁾ Is the bias vector of the linear transformation layer;

the picture volume is laminated:

wherein:

is the set of neighbor nodes of node i, AGGREGATE represents an aggregation function, +.>

Is the weight matrix of the convolutionally layer of the k-th layer diagram,>

feature vector representing node j of the layer-1 graph roll stack, +.>

A feature vector representing node i of the convolutionally layer of the k-th layer graph;

the full connection layer:

wherein: w (W) ^(L) ∈R ^dout×dout And

the weight and bias matrix of the full connection layer, respectively,>

representing the feature vector output by the node i at the full connection layer;

and then the feature vector is used for

The final probability distribution of failure probability is obtained by a Softmax activation function:

wherein: h is a _ij ^(L) Representing probability of failure between full connection layer node i and node jCharacteristic value, h _im ^(L) Representing the mth feature, p, of the ith node in the full connectivity layer output _ij A probability distribution representing the probability of failure between the slave node i and the node j.

Preferably, in step S3:

the graph neural network GNN fault learning model adopts MSE mean square error loss function during training

Where output is the output of the graph neural network GNN fault learning model and adj is the adjacency matrix of the actual network topology.

Preferably, step S4 comprises the steps of:

step S4.1: generating a network policy set according to requirements, wherein the network policy set comprises: policy sequence number, policy type, source device, destination device and latency constraint, wherein the policy type comprises: policy reachability, policy Isolation, and policy load balancing are represented by reach (src, dst, delay), isolation (src, dst), and Loadbalancing (src, dst), respectively, where src represents a source device, dst represents a destination device, and delay represents a latency constraint; the reach (src, dst, delay) indicates that network traffic reaches dst from src under the premise of delay constraint delay; isolation (src, dst) means that network traffic does not reach dst from src; loadbalancing (src, dst) means that network traffic is load balanced on at least two paths from src to dst;

step S4.2: constructing a strategy directed graph G according to a network strategy set, wherein G= (V, E); wherein v= { d|d=1,..n } is the set of all physical devices in the network, where d e V is the physical device with sequence number d in the strategy graph; e= { E (a, b) |a, b E V } is a set of policy state transitions between physical devices, where a represents a source device number, b represents a destination device number, E (a, b) ∈e represents a link exists between a device with device number a and a device with device number b so that network traffic is transferred from device a to device b.

Preferably, step S5 comprises the steps of:

step S5.1: preprocessing the strategy directed graph G by combining clipping and filling modes: if the number of the nodes or edges of the strategy directed graph G is larger than that of the nodes or edges of the network topology, cutting operation is carried out, and the nodes or edges which do not exist in the original network topology in the strategy directed graph G are deleted until the scale of the strategy directed graph G is the same as that of the original network topology; if the number of nodes or edges of the strategy directed graph G is smaller than that of the nodes or edges of the original network topology, filling operation is carried out, and virtual nodes or edges are added until the scale of the strategy directed graph G is the same as that of the original network topology;

step S5.2: inputting a trained K-Means model into the preprocessed strategy directed graph G, dividing the strategy directed graph G to obtain strategy subgraphs, wherein each strategy subgraph corresponds to a network topology subgraph under a label, and extracting the characteristics of each strategy subgraph;

step S5.3: inputting the features of the extracted strategy subgraph into a trained graph neural network GNN fault learning model under the corresponding label, and carrying out output results

Is transformed and binary classification is carried out; introducing a threshold value theta, judging the probability of being larger than the threshold value as 1, and judging the probability of being smaller than or equal to the threshold value as 0, namely:

where Edge represents the state of each Edge, i and j represent the source and destination devices of the Edge, respectively, and values 0 and 1 represent the failure and failure of the Edge, respectively,

representing connection probability between node i and node j learned by graph neural network GNN fault learning modelProbability distribution of (2); and obtaining the fault environment corresponding to the strategy subgraph according to the obtained link states among the nodes of the accurate fault environment.

Preferably, step S6 comprises the steps of:

step S6.1: classifying network policies according to requirements to obtain network policy sets under different policy types;

step S6.2: if the reachability of the network policy needs to be verified, a depth-first search algorithm is used to determine whether an reachable path exists from the source node to the target node in the given policy subgraph, and the maximum distance of the search is limited;

step S6.3: if the isolation of the network policy needs to be verified, a depth-first search algorithm is used for judging whether an reachable path from a source node to a target node exists in a given policy subgraph, the maximum distance of search is limited, and the result is inverted;

step S6.4: if the load balance of the network strategy needs to be verified, ensuring the load balance by checking whether a node arrival distance does not exceed a certain multiple of the average shortest path length; wherein, the average shortest path length is calculated by calculating the shortest path length from the source node to all other nodes, and then averaging.

Compared with the prior art, the invention has the beneficial effects that:

1. the network policy verification method based on the GNN fault model learning utilizes the GNN learning network environment, and can better learn the performance and the characteristics of the network compared with the common verification method, thereby determining the accurate fault environment for the policy to be verified and realizing accurate network policy verification.

2. According to the network policy verification method based on GNN fault model learning, the fault model is introduced during GNN learning, and reinforcement learning is combined, so that the influence of noise and error data on a network policy verification result is reduced, the robustness of decision is improved, and the network policy verification result is more accurate.

3. The network policy verification method based on the GNN fault model learning uses the K-Means clustering method to assist the GNN in carrying out network policy verification, reduces the network scale, ensures that the verification process is faster, and effectively relieves the state explosion problem of the reference tool.

Drawings

FIG. 1 is a schematic flow chart of the present invention;

FIG. 2 is a schematic diagram of a message passing mechanism used in the present invention;

FIG. 3 is a graph of model structure calculations based on GNN fault model learning as used in the present invention;

FIG. 4 is a schematic representation of a network policy expression defined in the present invention;

FIG. 5 is a comparison of verification times for the present invention and baseline methods when different maximum numbers of faults are considered.

Detailed Description

For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments of the present invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

As shown in fig. 1, the present invention provides a control plane policy verification method based on GNN fault model learning, which includes the following steps:

step S1, generating a corresponding fault model according to network topology information and the maximum fault number in the network set according to the requirement:

specifically, in this embodiment, the step S1 includes:

firstly, network topology information is obtained from public configuration data set of 5 months in 2015 of Internet2 network, the network topology comprises 33 nodes and 96 edges, and the calculated node degree D E R of the network topology is calculated ⁿ Degree of sum centrality node centrality C epsilon R ⁿ Wherein n is represented as the number 33 of devices in the network topology, the two features are taken as the standard to realize the division of equivalence classes for the training K-Means cluster model of the network topology, and the method is utilizedThe elbow rule selects the optimal equivalence class number as 8, and then generates a unique label and a topological graph corresponding to the equivalence class for each equivalence class.

Then, setting the maximum fault number k in the network topology according to the requirement _max . In the present embodiment, let k be _max =1, at maximum failure number k _max Under the limitation of (1), enumerating all possible failure environments. In order to avoid the explosion of the number caused by enumeration, the number num_combinations=100 of Montecello simulation is set, the number num_combinations=10 of fault combinations to be simulated is required, and the fault probability of each edge is calculated by Montecello simulation and added into the corresponding fault environment feature vector.

Step S2, designing a GNN fault learning model based on a message transmission mechanism, adopting a three-layer network structure, and comprising the following steps: linear transformation layer, picture convolution layer and full connection layer. The model is input into an original network topology adjacency matrix A epsilon R ^n×n Fault model F e R ^l×m (l represents the number of fault environments in the fault model, m represents the number of edges in the network), and the node centrality C E R ⁿ And node degree D epsilon R ⁿ Node characteristics expressed together, probability distribution p of failure probability between output nodes _ij 。

Specifically, as shown in FIG. 2, in the present embodiment, the input dimension is N _in =96, hidden layer dimension is N _hid =16, the output dimension is N _out =96, the activation function employs a nonlinear activation function ReLU. First, the input features undergo linear transformation and a ReLU activation function to obtain hidden layer features.

h _i ⁽¹⁾ ＝ReLU(W ⁽¹⁾ x _i +b ⁽¹⁾ ) (1)

Wherein W is ⁽¹⁾ Is a weight matrix of the input layer, with dimension [ N ] _in ,N _hid ]，x _i Is of dimension [ N ] _in ,N _in ]Input features of b) ⁽¹⁾ Is the bias vector of the linear transformation layer, and is output as the feature vector h of the hidden layer _i ⁽¹⁾ 。

Then, taking the first layer hidden layer characteristic as input, and obtaining a new hidden layer characteristic through a graph convolution layer:

wherein the method comprises the steps of

Is the neighbor node set of node i, c _ij The normalization coefficient between the node i and the node j is adopted, and the aggregation function adopts addition, W ^(k) Is the weight matrix of the convolutionally layer of the k-th layer diagram,>

feature vector representing node j of the layer-1 graph roll stack, +.>

A feature vector representing node i of the convolution layer of the k-th layer graph. As shown in fig. 2, the neighboring nodes of the destination node add the features element by element according to a certain rule, collect the features to the destination node for information transfer, and connect the nodes themselves by self-connection in order to consider the information quantity of the source node itself.

The third layer is a multi-layer perceptron (MLP) layer consisting of three linear layers and two ReLU activation functions, converts the output result of the graph convolution layer into a final output result, and obtains probability distribution of fault probability among nodes through an activation function Softmax:

p _ij ＝Softmax(W ^(L) h _i ^(L-1) +b ^(L) ) (3)

wherein,,

and->

The weight and bias matrix of the full connection layer are respectively output p _ij A probability distribution representing the probability of failure between the slave node i and the node j.

And S3, training the GNN model after the GNN model is designed, wherein the whole calculation diagram is shown in fig. 3, training data come from the original network topology, a Mean Square Error (MSE) loss function is adopted to measure the difference between the output of the GNN model and the adjacency matrix adj of the actual network topology of the real result, and an Adam optimizer is used to update model parameters.

And S4, generating a network policy map according to the network policy information which needs to be verified by the administrator.

Specifically, as shown in fig. 3, each network policy is represented by a five-tuple, including: policy sequence number, policy type, source device, destination device and latency constraints. Wherein the policy types include: policy reachability (src, dst, delay): indicating that network traffic can reach dst from src under the premise of time delay limitation delay; policy Isolation (src, dst): indicating that network traffic may not reach dst from src; load balancing Loadbalancing (src, dst) of the policy: indicating that network traffic is load balanced on at least two paths from src to dst. Where src denotes a source device, dst denotes a destination device, and delay denotes a delay constraint.

Then, constructing a strategy directed graph G according to the network strategy set, wherein G= (V, E); wherein v= { d|d=1,..n } is the set of all physical devices in the network, where d e V is the physical device with sequence number d in the strategy graph; e= { E (a, b) |a, b E V } is a set of policy state transitions between physical devices, where a represents a source device number, b represents a destination device number, E (a, b) ∈e represents a link exists between a device with device number a and a device with device number b so that network traffic is transferred from device a to device b.

S5, preprocessing a strategy directed graph, extracting feature data, inputting the feature data into a trained GNN model, acquiring probability distribution of fault probability among feature nodes, and generating an accurate network environment;

specifically, if the number of nodes or edges of the policy map is greater than the number of nodes or edges of the network topology, performing cutting operation, and deleting nodes or edges which do not exist in the original topology in the policy map; if the number of nodes or edges of the policy graph is smaller than that of the nodes or edges of the original topology, filling operation is carried out, and virtual nodes or edges are added.

And then inputting the preprocessed strategy diagram into a trained K-Means model to realize the division of strategy subgraphs, wherein each strategy subgraph corresponds to a network topology subgraph under one label, and extracting the characteristics of each strategy subgraph.

Finally, inputting the features of the extracted strategy subgraph into the trained GNN model under the corresponding label, and carrying out output result

The transformation of (a) to obtain the connection probability between nodes +.>

Then, binary classification is performed, a threshold value theta is introduced, the probability of the probability being larger than the threshold value is judged to be 1, and the probability of the probability being smaller than or equal to the threshold value is judged to be 0, namely:

where Edge represents the state of each Edge, i and j represent the source and destination devices of the Edge, respectively, the values 0 and 1 represent whether the Edge has a fault, 0 represents a fault, 1 represents no fault,

the probability distribution indicating the probability of connection between the slave node i and the node j obtained by GNN learning. And obtaining the fault environment corresponding to the strategy subgraph according to the obtained link states among the nodes of the accurate fault environment.

Step S6, respectively verifying the network policies of different types in an accurate fault environment;

specifically, network policies are classified according to requirements, and policy sets under different policy types are obtained.

If the reachability of the policy consideration delay needs to be verified, a depth-first search algorithm is used to determine if there is a reachable path from the source node to the target node in a given policy sub-graph, and to limit the maximum distance of the search.

If the isolation of the strategy needs to be verified, the method is consistent with the accessibility, and the final result is reversed.

If the policy's load balancing needs to be verified, load balancing is ensured by checking if there is a node arriving at a distance not exceeding a certain multiple of the average shortest path length. The average shortest path length is obtained by calculating the shortest path length from the source node to all other nodes and then averaging.

In summary, the present invention proposes a control plane policy verification method based on GNN fault model learning, as shown in fig. 4, with the increase of the maximum fault number, the present invention can quickly locate the accurate environment where the policy to be verified is located in the complex fault model, and realize accurate and quick network policy verification compared with the reference method. In addition, the method has certain generalization capability and is suitable for verification of different network strategies under various network topologies.

The present invention is not described in detail in the present application, and is well known to those skilled in the art.

The foregoing describes in detail preferred embodiments of the present invention. It should be understood that numerous modifications and variations can be made in accordance with the concepts of the invention by one of ordinary skill in the art without undue burden. Therefore, all technical solutions which can be obtained by logic analysis, reasoning or limited experiments based on the prior art by the person skilled in the art according to the inventive concept shall be within the scope of protection defined by the claims.

Claims (7)

1. The control plane strategy verification method based on GNN fault model learning is characterized by comprising the following steps:

s1, obtaining network topology information, dividing equivalence classes by using a K-Means clustering method, reducing the network scale, setting the maximum fault number in the network according to requirements, and generating a corresponding fault model;

s2, designing a GNN fault learning model based on a message passing mechanism, wherein the model comprises a linear transformation layer, a graph roll lamination layer and a full connection layer, inputting an adjacent matrix of network topology and a corresponding fault model, and outputting probability distribution of fault probability among feature nodes;

step S3, training a GNN fault learning model based on a message transmission mechanism for each equivalence class, and using the trained model for network policy verification;

step S4, collecting network policy information to be verified by an administrator, generating a network policy set according to requirements, and constructing a network policy diagram conforming to network topology characteristics;

s5, preprocessing a strategy directed graph, extracting feature data, inputting the feature data into a trained GNN model, acquiring probability distribution of fault probability among feature nodes, and generating an accurate network environment;

and step S6, respectively carrying out network policy verification according to the network policy category under the accurate network environment to obtain a final verification result.

2. The control plane policy verification method based on GNN fault model learning of claim 1, wherein step S1 comprises the steps of:

step S1.1: obtaining network topology information, calculating node degree and degree centrality, realizing the division of equivalent classes by training a K-Means clustering model of the network topology through the node degree and degree centrality, selecting the optimal cluster number as the optimal class number of the equivalent classes by using an elbow rule, and generating a unique label and a topological graph corresponding to the equivalent class for each equivalent class;

step S1.2: setting the maximum fault number k in the network topology according to the requirement _max Generating a fault model fault_model= { env based on Monte Carlo simulation ₁ ,...env _q }, wherein env _q Representing each possible fault environment in which the fault probability P is calculated by Monte Carlo simulation _fail ；

Assuming that the probability of failure of one edge (u, v) of the topology in step S1.1 is p, then at k _max In the step, the fault probability P _fail Expressed in binomial distribution:

wherein the method comprises the steps of

Represented at k _max In the step, q (q is more than or equal to 0 and less than or equal to k) _max ) The number of combinations of secondary faults; probability of failure P _fail Expressed in the maximum failure number k _max In, one edge (u, v) fails.

3. The control plane policy verification method based on GNN fault model learning according to claim 2, wherein in step S2:

constructing a graph neural network GNN fault learning model, wherein the graph neural network GNN fault learning model comprises a linear transformation layer, a graph roll lamination layer and a full connection layer:

the linear transformation layer:

h _i ⁽¹⁾ ＝σ(W ⁽¹⁾ x _i +b ⁽¹⁾ ) (2)

wherein: h is a _i ⁽¹⁾ Is the feature vector of the hidden layer, σ is the ReLU activation function, W ⁽¹⁾ ∈R ^din×dout Is the weight matrix of the input layer, x _i ∈R ^din Is the input eigenvector of node i, b ⁽¹⁾ Is the bias vector of the linear transformation layer;

the picture volume is laminated:

wherein:

is the set of neighbor nodes of node i, AGGREGATE represents an aggregation function, +.>

Is the weight matrix of the convolutionally layer of the k-th layer diagram,>

feature vector representing node j of the layer-1 graph roll stack, +.>

A feature vector representing node i of the convolutionally layer of the k-th layer graph;

the full connection layer:

h _i ^(L) ＝W ^(L) h _i ^(L-1) +b ^(L) (4)

wherein: w (W) ^(L) ∈R ^dout×dout And

the weight and bias matrix of the full connection layer, respectively,>

representing the feature vector output by the node i at the full connection layer;

and then the feature vector is used for

The final probability distribution of failure probability is obtained by a Softmax activation function:

wherein: h is a _ij ^(L) Characteristic value h representing failure probability between full connection layer node i and node j _im ^(L) Representing the mth feature, p, of the ith node in the full connectivity layer output _ij A probability distribution representing the probability of failure between the slave node i and the node j.

4. A control plane policy verification method based on GNN fault model learning according to claim 3, wherein in step S3:

the graph neural network GNN fault learning model adopts MSE mean square error loss function during training

Where output is the output of the graph neural network GNN fault learning model and adj is the adjacency matrix of the actual network topology.

5. The control plane policy verification method based on GNN fault model learning of claim 4 wherein step S4 comprises the steps of:

step S4.1: generating a network policy set according to requirements, wherein the network policy set comprises: policy sequence number, policy type, source device, destination device and latency constraint, wherein the policy type comprises: policy reachability, policy Isolation, and policy load balancing are represented by reach (src, dst, delay), isolation (src, dst), and Loadbalancing (src, dst), respectively, where src represents a source device, dst represents a destination device, and delay represents a latency constraint; the reach (src, dst, delay) indicates that network traffic reaches dst from src under the premise of delay constraint delay; isolation (src, dst) means that network traffic does not reach dst from src; loadbalancing (src, dst) means that network traffic is load balanced on at least two paths from src to dst;

step S4.2: constructing a strategy directed graph G according to a network strategy set, wherein G= (V, E); wherein v= { d|d=1,..n } is the set of all physical devices in the network, where d e V is the physical device with sequence number d in the strategy graph; e= { E (a, b) |a, b E V } is a set of policy state transitions between physical devices, where a represents a source device number, b represents a destination device number, E (a, b) ∈e represents a link exists between a device with device number a and a device with device number b so that network traffic is transferred from device a to device b.

6. The control plane policy verification method based on GNN fault model learning of claim 5, wherein step S5 comprises the steps of:

step S5.1: preprocessing the strategy directed graph G by combining clipping and filling modes: if the number of the nodes or edges of the strategy directed graph G is larger than that of the nodes or edges of the network topology, cutting operation is carried out, and the nodes or edges which do not exist in the original network topology in the strategy directed graph G are deleted until the scale of the strategy directed graph G is the same as that of the original network topology; if the number of nodes or edges of the strategy directed graph G is smaller than that of the nodes or edges of the original network topology, filling operation is carried out, and virtual nodes or edges are added until the scale of the strategy directed graph G is the same as that of the original network topology;

step S5.2: inputting a trained K-Means model into the preprocessed strategy directed graph G, dividing the strategy directed graph G to obtain strategy subgraphs, wherein each strategy subgraph corresponds to a network topology subgraph under a label, and extracting the characteristics of each strategy subgraph;

step S5.3: inputting the features of the extracted strategy subgraph into a trained graph neural network GNN fault learning model under the corresponding label, and carrying out output results

Is transformed and binary classification is carried out; introducing a threshold value theta, judging the probability of being larger than the threshold value as 1, and judging the probability of being smaller than or equal to the threshold value as 0, namely:

where Edge represents the state of each Edge, i and j represent the source and destination devices of the Edge, respectively, and values 0 and 1 represent the failure and failure of the Edge, respectively,

the probability distribution of connection probability between the node i and the node j obtained through the learning of the graph neural network GNN fault learning model is represented; and obtaining the fault environment corresponding to the strategy subgraph according to the obtained link states among the nodes of the accurate fault environment.

7. The control plane policy verification method based on GNN fault model learning of claim 6 wherein step S6 includes the steps of:

step S6.1: classifying network policies according to requirements to obtain network policy sets under different policy types;

step S6.2: if the reachability of the network policy needs to be verified, a depth-first search algorithm is used to determine whether an reachable path exists from the source node to the target node in the given policy subgraph, and the maximum distance of the search is limited;

step S6.3: if the isolation of the network policy needs to be verified, a depth-first search algorithm is used for judging whether an reachable path from a source node to a target node exists in a given policy subgraph, the maximum distance of search is limited, and the result is inverted;

step S6.4: if the load balance of the network strategy needs to be verified, ensuring the load balance by checking whether a node arrival distance does not exceed a certain multiple of the average shortest path length; wherein, the average shortest path length is calculated by calculating the shortest path length from the source node to all other nodes, and then averaging.

Images