Service identification method and system based on SDN shipborne network
By applying collaborative learning of SDN architecture and multi-view models in shipboard networks, it is possible to quickly identify and manage business traffic in complex maritime environments, solve the problem of insufficient flexibility and real-time in the existing technology, and improve network performance and reliability.
Patent Information
- Application Number
- CN202411051932.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-01
- Publication Date
- 2024-12-03
AI Technical Summary
The existing ship-borne network service identification technology lacks flexibility and real-time in complex maritime environments. Traditional algorithms cannot be differentiated according to business needs. It relies on fixed rules based on port and deep packet detection. Machine learning training time is long and difficult to meet. Real-time adjustment.
The online service identification method based on the SDN architecture is adopted, and data is quickly forwarded at edge nodes using flow table matching. The unmatched data is processed by the central controller, and the multi-view model and collaborative learning dynamic update model is used to adapt to network changes.
It improves the flexibility of network management and business orchestration efficiency, improves network performance and reliability, realizes optimized resource allocation, avoids the problems of long training time and dependence on fixed rules, and meets the needs of real-time adjustment.
Smart Images

Figure CN119071240A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of network communication technology, and in particular to a method and system for identifying online services in a shipborne network based on a software defined network (SDN). Background Art
[0002] As an important part of modern maritime communications, shipborne networks play a key role in ensuring maritime operations, navigation safety and ship management. The characteristics of shipborne networks are complex communication environment, uneven signal coverage at sea, limited bandwidth and susceptibility to factors such as weather. In a relatively harsh communication environment, a technology that can manage network traffic according to the network status is particularly important, which can enhance the stability and flexibility of shipborne networks and improve the efficiency and quality of mission execution.
[0003] With the development of digitalization of ships, the network carries data transmission services such as navigation, weather health, real-time monitoring, emergency communications, etc. These services are of various types, and the required network resources, priorities and service quality requirements are different. However, traditional network traffic management algorithms can usually only manage the overall traffic in a coarse-grained manner, lack the ability to identify services, and cannot differentiate management according to business needs. In addition, the scheduling flexibility in a more complex network environment is insufficient. In order to ensure the reliability and real-time performance of these key services, there is an urgent need for an algorithm that can identify services in the marine network environment and realize automated service orchestration and resource scheduling.
[0004] Existing service identification technologies are mainly divided into the following three types: 1. Port-based service identification technology, which determines the type of application layer protocol by capturing the port number, and then identifies the corresponding network service traffic. 2. Service identification technology based on deep packet inspection technology, which checks the header information and payload of the data packet, identifies specific application signatures or keywords, and thus realizes accurate identification of network services. 3. Service identification technology based on machine learning, Gomez et al. use decision trees to realize the classification and management of traffic.
[0005] The existing technology has the following technical problems: 1. Port-based business identification and deep packet inspection-based business identification methods rely too much on fixed rules, require regular manual adjustments, and have poor flexibility; 2. Business identification methods based on machine learning take a long time to train, their accuracy depends on the size of the training data set, and it is difficult to meet the needs of real-time adjustment. Summary of the invention
[0006] In view of this, the present invention provides an SDN-based shipboard network online service identification method and system, which applies the SDN architecture to the service identification task of the shipboard network. At the nodes at the edge of the network, some data can be forwarded according to the flow table matching method, and the characteristics of the unmatched data can be quickly identified, which can greatly improve the flexibility of network management and the efficiency of service arrangement. The solution also conducts collaborative learning of the model during the identification process, thereby dynamically updating it so that the model adapts to changes in the network, thereby improving the overall performance and reliability of the network.
[0007] In order to solve the above technical problems, the present invention is implemented as follows.
[0008] A method for identifying online services in a shipborne network based on SDN, wherein the shipborne network includes a control plane including a central controller and a data plane composed of edge switches; the method includes the following steps:
[0009] Step 1: When the data flow in the network reaches the edge switch, the data is forwarded using the flow table matching method, and the unmatched data flow is sent to the central controller for processing;
[0010] Step 2: A multi-view model is configured in the central controller, and the characteristics of the data stream from the edge switch are input into the multi-view model to identify the service type, and then returned to the edge switch for forwarding;
[0011] Step 3: Based on the data flow and business type identification results in the network, the multi-view model performs collaborative learning and updates the multi-view model to adapt to changes in network behavior.
[0012] Preferably, the multi-view model includes a node view model and a flow view model; the input of the node view model is the node feature, and the input of the flow view model is the network flow feature; the node view model is constructed using a K-means clustering algorithm.
[0013] Preferably, the multi-view model is constructed as follows:
[0014] The initial multi-view model is constructed using secondary clustering;
[0015] During the first clustering, the node features and network flow features of each piece of historical data are extracted, and the K-means++ algorithm is used for clustering to obtain the first K-means clustering result. The known business types of the traffic data in each cluster are counted as the dominant business type of the cluster.
[0016] During the second clustering, the K-means++ algorithm is used for clustering, and the distance from the data points that do not belong to the dominant business type of the cluster to the cluster centroid is negative to obtain the second K-means clustering result;
[0017] The distance metric in the K-means++ algorithm uses Manhattan distance.
[0018] Preferably, the node characteristics include network quintuple, maximum packet size, minimum packet size, average packet size, and interval time; the network flow characteristics include total flow data packet size, flow duration, and number of packets sent per second.
[0019] Preferably, in step S3, the multi-view model performs collaborative learning based on the new data flow and recognition results in the network, and the multi-view model is updated as follows: the credibility of the business type recognition results of the network incoming data flow is determined, each view model in the multi-view model screens the unlabeled data that is identified as high credibility, and affixes the recognition label of the view model to form a sample, which is provided to other view models for collaborative learning, and multiple negative samples identified as low credibility are deleted; each view model uses the samples provided by other view models to add to the cluster, recalculate the radius and center of the cluster, and update the multi-view model.
[0020] Preferably, the credibility is obtained by: recording the feature extracted from the network flow as x i , x i After clustering calculation with multi-view model, we can get i The closest cluster C, the radius of cluster C divided by the data x i The distance to the centroid of cluster C is x i The corresponding credibility.
[0021] Preferably, after calculating the credibility, the credibility level is further qualitatively determined: by default, the business type recognition results are all low credibility; if the business type of the nearest cluster C is the same as that of the second nearest cluster, the business type recognition result is medium credibility; if the feature x i The distance to cluster C is less than the radius of cluster C. At this time, the credibility is greater than 1, and the service type recognition result is highly reliable.
[0022] The present invention also provides an SDN-based shipborne network online service identification system, which includes a control plane and a data plane; the control plane includes a central controller, and the data plane is composed of multiple edge switches;
[0023] The edge switch is used to forward data by flow table matching when receiving data flow in the network, and send unmatched data flow to the central controller for service type identification, and forward data according to the identification result;
[0024] The central controller is configured with a multi-view model and a model update module; the features of the data flow from the edge switch are input into the multi-view model to identify the business type, and then returned to the edge switch for forwarding; the model update module constructs samples based on the data flow and business type identification results in the network, provides them to the multi-view model for collaborative learning, and updates the multi-view model to adapt to changes in network behavior.
[0025] Preferably, the central controller includes a model building module for building an initial multi-view model by adopting a secondary clustering method;
[0026] During the first clustering, the node features and network flow features of each historical data are extracted to obtain the input feature x i ,Use K-means++ algorithm to cluster, get the first K-means clustering result, count the known business types of traffic data in each cluster as the dominant business type of the cluster;
[0027] During the second clustering, the K-means++ algorithm is used for clustering, and the distance from the data points that do not belong to the dominant business type of the cluster to the cluster centroid is negative to obtain the second K-means clustering result;
[0028] The distance metric in the K-means++ algorithm uses Manhattan distance.
[0029] Preferably, when the model updating module updates the multi-view model, it selects the unlabeled data identified as high-credibility according to the credibility of the business type recognition result, and affixes the recognition label of the current view model to form a sample, which is provided to other view models for collaborative learning, and deletes multiple negative samples identified as low-credibility; each view model uses the samples provided by other view models to add to the cluster, recalculates the radius and center of the cluster, and updates the multi-view model;
[0030] The credibility is obtained by: the feature extracted from the network flow is recorded as x i , x i After clustering calculation with multi-view model, we can get i The closest cluster C; if feature x i If the distance to cluster C is less than the radius of cluster C, the service type recognition result is highly reliable.
[0031] Beneficial effects:
[0032] Aiming at the complex communication environment of shipboard networks, uneven signal coverage at sea, limited bandwidth and susceptible to weather and other factors, the present invention applies SDN to shipboard networks, takes advantage of the SDN architecture, performs functional segmentation on shipboard networks at different levels, retains fast identification based on flow tables at edge nodes, and only uploads data that cannot be matched to the central controller for deep identification, organically allocates fast identification and deep identification, reduces the computing and storage overhead of a single shipboard network, and thus ensures the timeliness and accuracy of identification. Moreover, separating the control plane from the data plane can greatly improve the flexibility of network management and the efficiency of service orchestration, making service identification and management in complex network environments more efficient.
[0033] The present invention installs the recognition model in the SDN controller and initializes the recognition model, extracts multi-view features of the network flow, and collaboratively learns and dynamically updates the model during the recognition process, so that the recognition model can dynamically adapt to network changes. Under this architecture, business recognition technology is particularly critical, and dynamic traffic engineering strategies and network management measures are implemented through business recognition technology to achieve optimal resource configuration, thereby improving the overall performance and reliability of the network.
[0034] In actual recognition applications, the present invention also selects highly credible data and recognition results as labeled samples based on the credibility of the samples. The two models use each other's labeled samples to add to the cluster to recalculate the radius and center of the cluster, and update the model to adapt to changes in network behavior.
[0035] The present invention adopts a secondary clustering method when performing initial training of the recognition model. In the objective function of the second clustering, for those data points that do not belong to the dominant business type of the cluster, their distance to the cluster centroid will be taken as a negative value, focusing on optimizing the cohesion of the dominant business type within the cluster while reducing the interference of non-dominant business types.
[0036] The solution of the present invention uses semi-supervised learning to initialize the model training, uses unlabeled data sets to update the model, and uses multi-view data features to achieve collaborative learning of two recognition models. The model is optimized in real time through the recognition results. Compared with the existing business identification schemes that use ports alone or business identification schemes based on deep packet detection, it does not rely too much on fixed rules, and adopts a complementary approach of rules + deep learning, which has better flexibility. Moreover, online learning based on high-confidence samples not only avoids the problem of long training time for business identification methods, but also meets the needs of real-time adjustment, and high-confidence samples improve the accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 1 It is a schematic diagram of the SDN-based shipborne network online service identification method of the present invention.
[0038] Figure 2 It is a schematic diagram of an application scenario of the SDN-based shipborne network online service identification method of the present invention.
[0039] Figure 3 The present invention is a block diagram of the composition of the SDN-based shipborne network online service identification system. DETAILED DESCRIPTION
[0040] The present invention is described in detail below with reference to the accompanying drawings and embodiments.
[0041] The present invention provides an SDN-based shipboard network online service identification solution, the core idea of which is: applying the SDN architecture to the service identification task of the shipboard network, the shipboard network includes a control plane and a data plane; when the data flow in the network reaches the edge switch of the data plane, the edge switch uses a flow table matching method to forward the data, and the unmatched data flow is sent to the central controller for rapid identification. This method of distinguishing edge nodes from central controllers reduces the requirements for storage and computing power of edge devices, and can greatly improve the flexibility of network management and the efficiency of service orchestration.
[0042] At the same time, the central controller is configured with a multi-view model to identify the business type of the data stream from the edge switch, and build samples based on the new data stream and business type identification results, so that the multi-view model can learn collaboratively, thereby dynamically updating the multi-view model and adapting it to network changes, thereby improving the overall performance and reliability of the network.
[0043] Figure 1 FIG. 1 is a schematic diagram showing a method for identifying online services in a shipborne network based on SDN according to the present invention. Figure 1 As shown, the method comprises the following steps:
[0044] Step S1: Use the labeled offline data set to perform multi-view feature extraction, build models of different views, and configure them in the central controller of the shipboard network.
[0045] In this embodiment, two feature sets are selected, both of which can be used to classify network data flows, and each feature set is called a view. In this embodiment, the two views are the node view model and the flow view model. The input of the node view model is the node feature, and the input of the flow view model is the network flow feature. The specific content of the input features of the two views can be, for example, that the node features include the five-tuple of the cluster, the maximum packet size, the minimum packet size, the average packet size, and the interval time; the network flow features include the total data packet size of the flow, the duration of the flow, and the number of packets sent per second, as shown in Table 1 below.
[0046] Table 1 Input features of two views
[0047]
[0048] All data are converted to decimal numbers for normalization. The node feature is a five-dimensional vector X1, and the network flow feature is a three-dimensional vector X2. Two vectors are extracted for each piece of historical data. Combined with the original service type tag of the data packet, the service type is marked for the historical data. Figure 2 The actual business application scenarios shown can be marked as the following five business types: large-scale data transmission, remote voice calls, remote video calls, interactive signals, and intensive general business.
[0049] Based on the division of views, the multi-view model of the present invention includes a node view model and a flow view model. Based on these two view models, the business flow is analyzed.
[0050] First, the multi-view model needs to be initialized. The data set is divided into two independent views. Based on these two views, a multi-view model is constructed using K-means clustering, including selecting the initial cluster center, iteratively updating the cluster center until convergence, and assigning a representative business type to each generated cluster. The business type is determined based on the most common business type in the cluster. In addition, secondary clustering and model adjustment are performed to further improve the accuracy of recognition.
[0051] In this embodiment, the process of constructing the initial multi-view model by means of secondary clustering is as follows:
[0052] Step S101: first clustering.
[0053] The set of all data streams is defined as X, and the vector of data stream X in d-dimensional space is defined as x = (x0, x1, ..., x Nj ),in Between any two data streams (x i ,x j ) is:
[0054]
[0055] Each cluster j in the view is denoted as C j =N j ,u j ,R j ,T j ,y j , N j Indicates that cluster j has N j d-dimensional data x i , where i = 1, 2, ..., N j , represents the centroid of cluster j, represents the radius of cluster j, T j represents the timestamp of cluster j, y j Indicates the actual type of the flow represented by cluster j.
[0056] Node features and network flow features are extracted for each piece of historical data, and clustered using the k-means++ algorithm to obtain the first K-means clustering result, i.e., K initial clustering centers, where K is the predetermined number of clusters.
[0057] Formula (2) is the objective function of the first clustering. For each data point x i For each point in the data set, calculate the square of its distance to the nearest selected center, denoted by D 2 (x), the probability of selecting the next cluster center and D 2 The objective function of clustering is formula (2). For each data point, the distance between it and all cluster centers is calculated, and it is classified into the cluster represented by the nearest cluster center:
[0058]
[0059] Formula (3) is the optimization process of the first clustering. Once all data points are classified, the new cluster center of each cluster is calculated, and the mean position of all points in the cluster is calculated. This process is repeated until the change of the cluster center converges.
[0060]
[0061] The first line in the above formula represents the calculation of the distance data point x i The nearest centroid is denoted by c i , and assign the point to the cluster. The second line represents the recalculation of the position mean of all points in the cluster as the new centroid. This process is repeated until the centroid of the cluster does not change, which means convergence.
[0062] The result of the first clustering is obtained, i.e., k clusters. Then, the known business types of the traffic data in the clusters are counted, and the one with the largest number is selected as the dominant business type of each cluster.
[0063] Step S102: Second clustering.
[0064] The K-means++ algorithm is used for clustering, and the distance from the data points that do not belong to the dominant business type of the cluster to the cluster centroid is negative to obtain the second K-means clustering result.
[0065] For the second clustering, the objective function is adjusted to formula (4):
[0066]
[0067] In the above formula, y represents the business type corresponding to the data point. The distance metric function at this time is: i The sum of the distances to the data points with the same business type as the i-th cluster minus the sum of the distances to the data sets with different business types. In this objective function, for those data points that do not belong to the dominant business type of the cluster, their distance to the cluster centroid will be negative, focusing on optimizing the cohesion of the dominant business type within the cluster while reducing the interference of non-dominant business types.
[0068] Preferably, based on the application scenario of the present invention, the distance metric in the K-means++ algorithm adopts Manhattan distance.
[0069] This completes the construction of the multi-feature view.
[0070] Step S2: When the data flow in the network reaches the edge switch, the characteristics of each data packet are extracted and matched with the flow table entry. The matched data flow is forwarded; the unmatched data flow is sent to the central controller for processing.
[0071] In this step, matching with flow table entries is as follows: the edge switch extracts the features of each data packet, which may include source IP address, destination IP address, source port, destination port, protocol type, input port, VLAN ID, MAC address, IP protocol field, Ethernet type, and matches with the existing flow table entries. If a match is found, it means that the data flow has been identified, and the data packet is processed according to the action specified in the flow table, such as forwarding or discarding; otherwise, there is no matching rule, and the data packet is marked as unidentified and uploaded to the central controller for further processing.
[0072] Step S3: The central controller checks the received data stream, identifies the service type using a multi-view model based on the characteristics of the data stream, and then returns it to the edge switch for forwarding.
[0073] In this step, the features of the data flow can be extracted by the edge switch. The edge switch collects the initial data packets of the network flow and extracts features from the initial data packets of each flow. After feature extraction, if the data plane flow table fails to match, the feature data set is sent to the control plane, which performs the first rapid identification.
[0074] Step S4: Based on the business type identification result of the data stream by the central controller, the data stream information and the identification result are combined into a new sample, and the multi-view model is updated to adapt to the changes in network behavior.
[0075] The business type identification operation in this step is:
[0076] First, heuristic rules are applied to perform preliminary analysis on the collected features, including obtaining specific ports or DNS fields in the packet header through table lookup, connection establishment time, active time period, fixed-size and periodic data packets, etc. Classification can be performed based on heuristic rules.
[0077] Secondly, construct node features / network flow features, denoted as x i , the feature x i As input, use the semi-supervised learning model trained in step S1 clustering for recognition and calculate the centroids u of all clusters j With x i The distance D(x i ,u j ), find the closest cluster C. Among them, D(x i ,u j ) refers to the feature x i The distance to the centroid of the j-th cluster.
[0078] Finally, the output of the heuristic rule and the recognition result of the semi-supervised learning model are combined to determine the service type of each data flow. That is, if a flow is successfully classified by the heuristic rule, the result is directly adopted, otherwise the recognition result of the semi-supervised learning model is adopted.
[0079] The operations for updating the multi-view model in this step are:
[0080] First calculate the credibility:
[0081]
[0082] In the above formula, R is the characteristic x i The radius of the nearest cluster C, D(x i ,u c ) represents the feature xi to the nearest cluster C centroid u c distance.
[0083] According to the credibility p, a ratio or threshold can be set to filter high credibility data. In a preferred embodiment, the credibility level can be further characterized by the following method:
[0084] By default, the classification result is of low confidence. If the business type of the nearest cluster C and the second nearest cluster is the same, the confidence of the classification result is increased to medium confidence. i and cluster centroid u c If the distance is less than the radius R of cluster C, the data point is considered to be in the core area of the cluster. At this time, the credibility value is greater than 1, and the credibility of the classification result is improved to high credibility.
[0085] During the operation, the models of the two views are asked to screen the unlabeled flows with high credibility and attach the identification labels of the models. The labeled samples with high reliability selected by the two models are provided to each other for collaborative learning. The top p samples with high credibility are selected for subsequent model updates. At the same time, the n negative samples with the lowest credibility are deleted. The two models use the samples labeled by each other to add to the cluster, recalculate the radius and center of the cluster, and update the model.
[0086] Based on the above method, the present invention also provides a shipborne network online service identification system based on SDN, such as Figure 3 As shown, the system includes a control plane and a data plane; the control plane contains a central controller, and the data plane is composed of multiple edge switches.
[0087] The edge switch is used to forward data by flow table matching when receiving data flow from the network. Unmatched data flow is sent to the central controller for service type identification, and data is forwarded according to the identification result.
[0088] The central controller is configured with a multi-view model, a model building module, and a model updating module; the features of the data stream from the edge switch are input into the multi-view model for business type identification, and then returned to the edge switch for forwarding; the model updating module constructs samples based on the data stream and business type identification results in the network, provides them to the multi-view model for collaborative learning, and updates the multi-view model to adapt to changes in network behavior.
[0089] The model building module is used to build the initial multi-view model using secondary clustering:
[0090] During the first clustering, the node features and network flow features of each historical data are extracted to obtain the input feature x i ,Use K-means++ algorithm to cluster, get the first K-means clustering result, count the known business types of traffic data in each cluster as the dominant business type of the cluster;
[0091] During the second clustering, the K-means++ algorithm is used for clustering, and the distance from the data points that do not belong to the dominant business type of the cluster to the cluster centroid is negative to obtain the second K-means clustering result;
[0092] The distance metric in the K-means++ algorithm uses Manhattan distance.
[0093] When the model update module updates the multi-view model, it selects the unlabeled data with high credibility according to the credibility of the business type recognition results, and affixes the recognition label of this view model to form samples, which are provided to other view models for collaborative learning; each view model uses the samples provided by other view models to join the cluster, recalculate the radius and center of the cluster, and update the multi-view model;
[0094] The credibility is obtained by: the feature extracted from the network flow is recorded as x i , x i After clustering calculation with multi-view model, we can get i The closest cluster C; if feature x i If the distance to cluster C is less than the radius of cluster C, the service type recognition result is highly reliable.
[0095] The present invention aims to overcome the limitations of the prior art and provide a method that can accurately identify the service type and service quality requirements of the entire data stream by analyzing the information of some groups in the data stream in a software-defined shipboard network. The method uses online incremental learning and a shipboard network model to adaptively determine the service type and network environment requirements of each data stream, without the need for manual adjustment of identification parameters, significantly improving the accuracy of identification and the applicability of the method, and is suitable for shipboard networks, SDN networks and other marine communication fields.
[0096] The above specific embodiments only describe the design principle of the present invention. The shapes and names of the components in the description may be different and are not limited. Therefore, those skilled in the art in the field of the present invention may modify or replace the technical solutions recorded in the above embodiments; and these modifications and replacements do not deviate from the creative purpose and technical solutions of the present invention and should all fall within the protection scope of the present invention.
Claims
1. A method for identifying online services in a shipborne network based on SDN, characterized in that: The onboard network includes a control plane including a central controller and a data plane composed of edge switches; the method includes the following steps: Step 1: When the data flow in the network reaches the edge switch, the data is forwarded using the flow table matching method, and the unmatched data flow is sent to the central controller for processing; Step 2: A multi-view model is configured in the central controller, and the characteristics of the data stream from the edge switch are input into the multi-view model to identify the service type, and then returned to the edge switch for forwarding; Step 3: Based on the data flow and business type identification results in the network, the multi-view model performs collaborative learning and updates the multi-view model to adapt to changes in network behavior.
2. The method according to claim 1, characterized in that The multi-view model includes a node view model and a flow view model; the input of the node view model is the node feature, and the input of the flow view model is the network flow feature; Use K-means clustering algorithm to build node view model.
3. The method according to claim 2, characterized in that The multi-view model is constructed as follows: The initial multi-view model is constructed using secondary clustering; During the first clustering, the node features and network flow features of each piece of historical data are extracted, and the K-means++ algorithm is used for clustering to obtain the first K-means clustering result. The known business types of the traffic data in each cluster are counted as the dominant business type of the cluster. During the second clustering, the K-means++ algorithm is used for clustering, and the distance from the data points that do not belong to the dominant business type of the cluster to the cluster centroid is negative to obtain the second K-means clustering result; The distance metric in the K-means++ algorithm uses Manhattan distance.
4. The method according to claim 2, characterized in that The node characteristics include network quintuple, maximum packet size, minimum packet size, average packet size, and interval time; the network flow characteristics include total flow data packet size, flow duration, and number of packets sent per second.
5. The method according to claim 1, characterized in that In step S3, the multi-view model performs collaborative learning according to the new data flow and recognition results in the network, and updates the multi-view model by: determining the credibility of the business type recognition results of the network incoming data flow, each view model in the multi-view model screens the unlabeled data identified as high credibility, and affixes the recognition label of the view model to form a sample, which is provided to other view models for collaborative learning, and deletes multiple negative samples identified as low credibility; Each view model uses the samples provided by other view models to join the cluster, recalculate the radius and center of the cluster, and update the multi-view model.
6. The method according to claim 5, characterized in that The credibility is obtained by: the feature extracted from the network flow is recorded as x i , x i After clustering calculation with multi-view model, we can get i The closest cluster C is the radius of cluster C divided by the data x i The distance to the centroid of cluster C is x i The corresponding credibility.
7. The method according to claim 6, characterized in that After calculating the credibility, the credibility level is further qualitatively determined: by default, the business type recognition results are all low credibility; if the business type of the nearest cluster C is the same as that of the second nearest cluster, the business type recognition result is medium credibility; if the feature x i The distance to cluster C is less than the radius of cluster C. At this time, the credibility is greater than 1, and the service type recognition result is highly reliable.
8. A shipborne network online service identification system based on SDN, characterized in that: The system includes a control plane and a data plane; the control plane contains a central controller, and the data plane consists of multiple edge switches; The edge switch is used to forward data by flow table matching when receiving data flow in the network, and send unmatched data flow to the central controller for service type identification, and forward data according to the identification result; The central controller is configured with a multi-view model and a model update module; the features of the data flow from the edge switch are input into the multi-view model to identify the business type, and then returned to the edge switch for forwarding; the model update module constructs samples based on the data flow and business type identification results in the network, provides them to the multi-view model for collaborative learning, and updates the multi-view model to adapt to changes in network behavior.
9. The method according to claim 8, characterized in that The central controller includes a model building module for building an initial multi-view model by using a secondary clustering method; During the first clustering, the node features and network flow features of each historical data are extracted to obtain the input feature x i ,Use K-means++ algorithm to cluster, get the first K-means clustering result, count the known business types of traffic data in each cluster as the dominant business type of the cluster; During the second clustering, the K-means++ algorithm is used for clustering, and the distance from the data points that do not belong to the dominant business type of the cluster to the cluster centroid is negative to obtain the second K-means clustering result; The distance metric in the K-means++ algorithm uses Manhattan distance.
10. The method according to claim 8, characterized in that When the model update module updates the multi-view model, it selects the unlabeled data identified as high-credibility according to the credibility of the business type recognition result, and affixes the recognition label of the current view model to form a sample, which is provided to other view models for collaborative learning, and deletes multiple negative samples identified as low-credibility; each view model uses the samples provided by other view models to add to the cluster, recalculates the radius and center of the cluster, and updates the multi-view model; The credibility is obtained by: the feature extracted from the network flow is recorded as x i , x i After clustering calculation with multi-view model, we can get i The closest cluster C; If the feature x i If the distance to cluster C is less than the radius of cluster C, the service type recognition result is highly reliable.
Citation Information
Cited By
Low-delay network scheduling method and system based on traffic type identification
CN122437815A
