A street-level IP address location method based on graph neural network

Through the method based on graph neural network, the topological graph structure is constructed and graph convolutional neural network and regular flow technology is used to solve the accuracy and stability of IP address positioning in the existing technology, and high-precision street-level IP address is achieved Positioning and anti-interference capability.

CN115242868BActive Publication Date: 2024-11-12郑州埃文科技有限公司
View PDF 7 Cites 0 Cited by

Patent Information

Application Number
CN202210838595.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-07-13
Publication Date
2024-11-12
Estimated Expiration
2042-07-13

AI Technical Summary

Technical Problem

Existing IP address positioning technologies are difficult to achieve fine-grained and reliable geographical location positioning, especially in complex and unstable network environments, and cannot cope with the growing demand for high-precision street-level positioning.

Method used

Using a graph neural network-based method, the topological graph structure is constructed through IP address clustering, topological spatial relationship establishment, graph convolutional neural network aggregation features and regular flow technology, and a multi-layer perceptron is used to perform linear affine changes to predict the target. The latitude and longitude of IP.

Benefits of technology

It realizes high-precision street-level IP address positioning, enhances the interpretability and anti-interference ability of the model, can more accurately handle noise in complex network environments, and provides more valuable prediction results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure SMS_100
    Figure SMS_100
  • Figure QLYQS_16
    Figure QLYQS_16
Patent Text Reader

Abstract

The present invention provides a street-level IP address positioning method based on a graph neural network, comprising the following steps: first, screening landmark IPs matching a target IP; second, establishing a topological spatial relationship between the target IP and the landmark IPs in the area from two different perspectives, namely, IP features and network measurements, to obtain edge weights between the target IP and the associated landmark IPs, and forming a complete topological graph structure; third, utilizing the edge weights and topological graph structure between the target IP and the landmark IP, using a graph convolutional neural network to aggregate neighborhood landmark features and geographic coordinates, and obtaining a probabilistic representation of the target IP through a regularized flow to resist the negative impact of network fluctuations, and obtaining a robust hidden state of the IP; finally, using a multi-layer perceptron to perform linear affine transformation, decoding the longitude and latitude coordinates of the target IP from the probabilistic representation, i.e., the hidden state of the target IP, to complete the IP positioning task; the present invention has the advantages of accurate IP address positioning and strong anti-interference ability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of IP address positioning, and specifically relates to a street-level IP address positioning method based on graph neural network. Background Art

[0002] An IP address is a unique identifier assigned to each host connected to the Internet, used for personal information identification and virtual location addressing. Actively exploring the characteristics of the IP address itself and extracting the knowledge behind it from multiple dimensions (network latency, port openness, and WHOIS registration information, etc.) can be considered IP address analysis. Through IP address analysis, risky behaviors of virtual addresses can be identified and users' personal information can be mined. This information is widely used in many downstream software (such as payment risk control, DDos attack prevention, and life recommendation platforms). In IP address analysis, the speculation of geographic location information has attracted much attention in recent years. It refers to the process of assigning a real-world geographic location (usually represented by longitude and latitude) to a given target IP address. This is of great significance to a wide range of Internet applications customized by many users, including targeted marketing, fraud prevention, etc. With the user's accurate geographic location, online platforms can also provide a variety of personalized services. For example, online targeted advertising can recommend restaurants near users. Ubiquitous location-based services have always been considered an important challenge facing the Internet.

[0003] In the past, IP address positioning was mainly carried out by collecting geographic location-related fields stored on public routers and establishing a mapping relationship library. However, the number of public routers containing geographic location information is extremely small, and the passive collection method cannot guarantee the reliability of geographic location. Relying on such methods, it is almost impossible to obtain fine-grained and reliable geographic location. Some methods use probes in various places to send ping commands to the target IP and observe the delay to infer its geographic location. However, the network environment is complex and unstable, and the topology structure is often uncertain, so this method often finds it difficult to obtain accurate results. In recent years, the explosive growth of network information and the rapid development of multimedia information processing technology have enabled IP address positioning to use emerging technologies for data collection and positioning. These methods use natural language processing, multimodal and other technologies to search and store login records on social networks, geographic clues in web multimedia information, and real-time webcam data, search for their server IP addresses, and establish a mapping relationship between IP addresses and geographic locations.

[0004] With the rapid development of machine learning technology, deep learning has achieved excellent results in many application fields. In the field of IP address positioning, the current methods use clustering methods (such as KNN) and multi-layer perceptron (MLP) technology, combined with the characteristics of IP addresses, first gradually narrow the range of IP addresses, and then infer their longitude and latitude, so that the IP address positioning task can be basically completed. However, this method cannot consider the topological association between IPs, so the model has poor interpretability, incomplete data utilization, and difficulty in coping with widespread network fluctuations, and cannot meet the growing needs of high-precision street-level positioning. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide an effective street-level IP address positioning method based on graph neural network.

[0006] The technical solution of the present invention is as follows:

[0007] A street-level IP address location method based on graph neural network, comprising the following steps:

[0008] Step S1, IP address clustering, screening landmark IPs matching the target IP, and obtaining an area containing the target IP and several landmark IPs as graph nodes;

[0009] Step S2, IP address connection, from two different perspectives of IP features and network measurement, establish the topological spatial relationship between the target IP and the landmark IP in the area, so as to obtain the edge weight between the target IP and the associated landmark IP, and form a complete topological graph structure;

[0010] Step S3: IP feature aggregation: using the edge weights and topological graph structure between the target IP and the landmark IP, a graph convolutional neural network is used to aggregate the neighborhood landmark features and geographic coordinates, and a probabilistic representation of the target IP is obtained through a regularized flow to resist the negative impact of network fluctuations, thereby obtaining a robust hidden state of the IP.

[0011] Step S4: IP address location. Use a multi-layer perceptron to perform linear affine transformation, decode the longitude and latitude coordinates of the target IP from the probability representation, that is, the hidden state of the target IP, and complete the IP location task.

[0012] Further, the specific process of step S1 is as follows:

[0013] Use the tracert tool on probe machines located in different cities to search for routing paths and record the last-hop router; store the obtained last-hop router in the traceroute list, and then cluster the IP hosts with the same last-hop router in the traceroute list into a graph, that is, obtain an area containing the target IP and several landmark IPs as graph nodes.

[0014] Furthermore, the specific process of obtaining the edge weight between the target IP and the associated landmark IP in step S2 is as follows:

[0015] Establishing a first edge weight using an existing topology obtained by network measurement between IP addresses in the same area and public routers of the IP addresses in the area;

[0016] Landmark IPs with similar radius to the target IP are selected to construct the first edge weight which is inversely proportional to the distance difference between the target IP and the landmark IP, expressed as:

[0017] Formula 1,

[0018] in, , and They represent the target IP, landmark IP and public router respectively;

[0019] Through the distance mapping function Map network measurements to geographic distances and calculate and The radial distance difference between

[0020] The expression of the distance mapping function is as follows:

[0021] Formula 2,

[0022] in, and Respectively represent the delay and routing hops between the IP host and the public router, , and is a hyperparameter, which is obtained by learning using a fully connected layer network with network measurement as input.

[0023] The second edge weight is constructed by using the feature similarity between IP addresses. Given the target IP and landmark IP, the feature vectors of any two IPs are concatenated and their dot product is calculated to obtain the similarity between the two IPs as the second edge weight. The expression is as follows:

[0024] Formula 3,

[0025] Where T represents the matrix transpose, Represents the concatenation of the target IP and landmark IP feature vectors, and are the parameters learned through the neural network.

[0026] Further, the first edge weight and the second edge weight After fusing and controlling the sparsity of the final graph, the topological graph structure is described as follows:

[0027] Formula 4,

[0028] in, is a hyperparameter that balances the weight of the first edge with the weight of the second edge, It is the threshold that controls the sparsity of the adjacency matrix of the topological graph structure.

[0029] Further, the specific process of step S3 is as follows:

[0030] First, a simple Gaussian distribution is obtained using a graph convolutional neural network (GCN):

[0031] Formula 5,

[0032] in, The concatenation of the feature vector representing the IP and the longitude and latitude of the landmark. For the target IP to be predicted, (0,0) is used instead. and are two graph convolutional neural networks with different parameters, is a random variable that obeys Gaussian distribution, representing the hidden state of all IP nodes;

[0033] Then, a continuous regularization flow is used to transform the hidden state Continue to derive a more flexible distribution form to cope with the ever-changing network environment. and is the initial random variable and probability distribution, then the distribution change can be expressed as:

[0034] Formula 6,

[0035] Formula 7,

[0036] Formula 8,

[0037] in, is a hyperparameter representing the span of the distribution transformation, So is a fully connected layer with learnable parameters, represents the trace of a matrix;

[0038] Finally, the final posterior distribution is optimized by maximizing the evidence lower bound (ELBO), as follows:

[0039] Formula 9,

[0040] in, Represents the reconstruction loss between the reconstructed topological graph structure and the original graph structure, which is calculated as follows:

[0041] Formula 10,

[0042] In the formula is the Heaviside step function, Represents a fully connected neural network used to reconstruct the topology of a graph from its hidden states.

[0043] Furthermore, in step S4, a multi-layer perceptron is used to perform linear affine transformation to predict the longitude and latitude of the target IP address to obtain the target IP address, and the expression is:

[0044] Formula 11,

[0045] in, is a latitude and longitude tuple representing geographic location information, Represents the parameters obtained through training.

[0046] Compared with the prior art, the street-level IP address positioning method based on graph neural network provided by the present invention has the following beneficial effects:

[0047] 1. Based on graph neural network, the present invention provides a complete set of processing flow and framework for street-level IP positioning; the present invention uses IP features and network measurements to predict its high-precision geographic location; compared with traditional methods that learn linear features interactively or independently process each IP address, the method provided by the present invention integrates the features of multiple IP addresses and the various neighborhood relationships between them, and extracts public knowledge and topological connections from the topological graph structure, making IP positioning more accurate, thereby realizing street-level IP address positioning;

[0048] 2. The present invention is based on the generative model regularized flow technology to build a hidden state related to the target IP address derived in a probabilistic form; through continuous and flexible reasoning in the probability space, the uncertainty in the data is modeled, thus paying attention to the noise caused by phenomena such as network congestion and jitter in complex network environments; compared with deterministic modeling, the IP positioning framework of the present invention is more robust and has stronger anti-interference ability, so the prediction results are also more valuable. DETAILED DESCRIPTION

[0049] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] A street-level IP address location method based on graph neural network, comprising the following steps:

[0051] Step S1, IP address clustering, screening landmark IPs matching the target IP, and obtaining an area containing the target IP and several landmark IPs as graph nodes;

[0052] Step S2, IP address connection, from two different perspectives of IP features and network measurement, establish the topological spatial relationship between the target IP and the landmark IP in the area, so as to obtain the edge weight between the target IP and the associated landmark IP, and form a complete topological graph structure;

[0053] Step S3: IP feature aggregation: using the edge weights and topological graph structure between the target IP and the landmark IP, a graph convolutional neural network is used to aggregate the neighborhood landmark features and geographic coordinates, and a probabilistic representation of the target IP is obtained through a regularized flow to resist the negative impact of network fluctuations, thereby obtaining a robust hidden state of the IP.

[0054] Step S4: IP address location. Use a multi-layer perceptron to perform linear affine transformation, decode the longitude and latitude coordinates of the target IP from the probability representation, that is, the hidden state of the target IP, and complete the IP location task.

[0055] In the present invention, the specific process of step S1 is as follows:

[0056] Use the tracert tool on probe machines located in different cities to search for routing paths and record the last-hop router; store the obtained last-hop router in the traceroute list, and then cluster the IP hosts with the same last-hop router in the traceroute list into a graph, that is, obtain an area containing the target IP and several landmark IPs as graph nodes;

[0057] In specific implementation, there may be multiple last-hop routers in different routing paths, and the router with the smallest delay is selected; if the network's firewall policy hides the last-hop router, the last visible router will be used at this time; due to the Internet's regional management policy, the IP addresses with a public last-hop router are usually geographically close, so the public router can be used as an intermediary to connect the target IP and the landmark IP, so that an accurate network topology relationship can be established between the target IP and the landmark; so far, a rough area containing the target IP and the associated landmark IP as graph nodes is provided, and the feasibility of analyzing the precise geographic location through the graph structure using the characteristics of the IP address and the topological relationship is guaranteed.

[0058] In the present invention, the purpose of step S2 is to obtain the edge weights, i.e., the topological connections, between the target IP and the landmark IP nodes, so as to obtain a complete graph structure. The present invention establishes a geolocation graph based on two assumptions: (1) In the graph structure, a larger edge weight reflects a stronger correlation between two nodes; (2) The distance between the geographical coordinates of IP hosts with stronger correlation is closer; in short, the closer the geographical locations, the greater the edge weights between IP addresses; in order to achieve this purpose, the present invention defines edge weights from two perspectives, i.e., network topology and feature similarity;

[0059] In specific implementation, the specific process of obtaining the edge weight between the target IP and the associated landmark IP in step S2 is as follows:

[0060] Establishing a first edge weight using an existing topology obtained by network measurement between IP addresses in the same area and public routers of the IP addresses in the area;

[0061] IP hosts in the same area surround a common router and select landmarks with a similar radius to the target IP; select landmark IPs with a similar radius to the target IP to construct the first edge weight that is inversely proportional to the distance difference between the target IP and the landmark IP, expressed as:

[0062] Formula 1,

[0063] in, , and They represent the target IP, landmark IP and public router respectively;

[0064] Through the distance mapping function Map network measurements to geographic distances and calculate and The radial distance difference between

[0065] The expression of the distance mapping function is as follows:

[0066] Formula 2,

[0067] in, and Respectively represent the delay and routing hops between the IP host and the public router, , and is a hyperparameter, which is obtained by learning and using a fully connected layer network with network measurement as input; for or .

[0068] In step S2 of the present invention, the feature similarity between IP addresses is used to construct the second edge weight to better connect the graph nodes and describe the relationship between them; specifically, the target IP ( ) and landmark IP ( ), by concatenating the feature vectors of any two IPs and calculating their dot product, we can get the similarity between the two IPs as the second edge weight, which is expressed as follows:

[0069] Formula 3,

[0070] Where T represents the matrix transpose, Represents the concatenation of the target IP and landmark IP feature vectors, and are the parameters obtained through neural network learning; Formula 3 introduces the original IP features into the metric space and extracts similarities interactively to obtain better semantic relevance.

[0071] In the present invention, the edge weights of two different perspectives are used and , both the topological relationship and IP host knowledge are incorporated into the structured information; however, the current structure established is a complete graph; this not only consumes a lot of computing resources, but also considers some irrelevant nodes, resulting in inefficiency and inaccuracy; therefore, the present invention introduces a mechanism to fuse the edge weights of the two perspectives and control the sparsity of the final graph; specifically, the first edge weight and the second edge weight After fusing and controlling the sparsity of the final graph, the topological graph structure is described as follows:

[0072] Formula 4,

[0073] in, is a hyperparameter that balances the weight of the first edge with the weight of the second edge, It is the threshold that controls the sparsity of the adjacency matrix of the topological graph structure.

[0074] The purpose of step S3 of the present invention is to cope with the complex, changeable and unstable characteristics of the network environment; the present invention adopts the more flexible and efficient continuous normalizing flow (CNF) technology in the generative model (Generative Model) regularization flow (NF), combined with the obtained graph structure, to aggregate the graph nodes in a probabilistic form; the specific process of step S3 is as follows:

[0075] First, a simple Gaussian distribution is obtained using a graph convolutional neural network (GCN):

[0076] Formula 5,

[0077] in, The concatenation of the feature vector representing the IP and the longitude and latitude of the landmark. For the target IP to be predicted, (0,0) is used instead. and are two graph convolutional neural networks with different parameters, is a random variable that obeys Gaussian distribution, representing the hidden state of all IP nodes;

[0078] Then, a continuous regularization flow is used to transform the hidden state Continue to derive a more flexible distribution form to cope with the ever-changing network environment. and is the initial random variable and probability distribution, then the distribution change can be expressed as:

[0079] Formula 6,

[0080] Formula 7,

[0081] Formula 8,

[0082] in, is a hyperparameter representing the span of the distribution transformation, So is a fully connected layer with learnable parameters, represents the trace of a matrix;

[0083] Finally, the final posterior distribution is optimized by maximizing the evidence lower bound (ELBO), as follows:

[0084] Formula 9,

[0085] in, Represents the reconstruction loss between the reconstructed topological graph structure and the original graph structure, which is calculated as follows:

[0086] Formula 10,

[0087] In the formula is the Heaviside step function, Represents a fully connected neural network used to reconstruct the topology of a graph from its hidden states.

[0088] In the present invention, in step S4, a multi-layer perceptron is used to perform linear affine transformation to predict the longitude and latitude of the target IP address to obtain the target IP address, and the expression is:

[0089] Formula 11,

[0090] in, is a latitude and longitude tuple representing geographic location information, Represents the parameters obtained through training.

[0091] The network and framework for street-level IP address location based on the aforementioned graph neural network have been constructed. The present invention randomly samples enough numbers from a normal distribution with a mean of 0 and a variance of 0.1 as the initial parameters of the network; the purpose of the present invention is to obtain a system with excellent performance in the IP address location task; therefore, the goal of the above training is to minimize the geographical distance between the predicted longitude and latitude and the true label. In order to improve the optimization efficiency, the present invention uses the mean square error (MSE) of the distance as the loss function, and uses a small batch stochastic gradient descent method to perform gradient descent to update the parameters and obtain parameters with good prediction effects. At the same time, we also use this indicator, combined with the mean absolute error (MAE) and the median to evaluate the model IP location effect.

[0092] So far, the present invention has realized a street-level IP location framework based on graph neural network. It should be noted that the framework provided by the present invention does not impose any restrictions on the graph neural network itself, so it can be used in combination with any advanced graph neural network and has good scalability. Since simple graph convolutional neural networks are efficient and perform well in regression problems such as IP classification, our framework is usually used in combination with convolutional neural networks.

[0093] Experimental Description

[0094] In order to verify the effectiveness of the street-level IP address positioning based on graph neural network in this invention, this embodiment uses three large-scale real IP geolocation data sets collected from three metropolitan areas, New York, Los Angeles and Shanghai, consisting of 91808, 92804 and 126258 IP addresses respectively. At the same time, multiple IP positioning methods, tabular data learning methods, and graph neural network models are used as baseline comparisons to evaluate the effects.

[0095] This application example uses Root Mean Squared Error (RMSE), Mean Absolute Error (MAE) and Median Error as evaluation indicators (the lower the value, the better the model method). Among them, Ours represents the test results of the positioning system provided in this embodiment, and the best results are represented in bold:

[0096] Table 1: IP address location experiment results

[0097]

[0098] The remaining methods in the table are described as follows:

[0099] CBG: Establish a continuous space and use multilateral positioning with distance constraints to infer the geographic location of Internet hosts. The implementation process is referenced in [Bamba Gueye, Artur Ziviani, Mark Crovella, and Serge Fdida. 2006. Constraint-Based Geolocation of Internet Hosts. IEEE / ACM Trans. Netw. (2006), 1219–1232];

[0100] TBG: is a topology-based geolocation method that converts topology and communication delay into a set of constraints to geolocate routers and Internet hosts simultaneously. Its implementation process is referenced in [Ethan Katz-Bassett, John P John, Arvind Krishnamurthy, David Wetherall, Thomas Anderson, and Yatin Chawathe. 2006. Towards IP geolocation using delay and topology measurements. In SIGCOMM. 71–84];

[0101] TLP: proposed the concept of IP range interpolation and combined it with traceroute delay neighbors for IP geolocation. Its method and process references [Ovidiu Dan, Vaibhav Parikh, and Brian D Davison. 2021. IP Geolocation Using Traceroute Location Propagation and IP RangeLocation Interpolation. In WWW. 332–338];

[0102] SVR: A regression model based on support vector machines. It plots each data item as a point in n-dimensional space, and we can draw a hyperplane so that our target approximates the plane. For the kernel function of the support vector machine, we choose a linear kernel. It can also be used in IP address location tasks. Its implementation process is referenced in [Qian Zhao, Fei Wang, Can Huang, and Chuan Yu. 2020. Improving IP geolocation databases based on multi-method classification. In ASID. 44–48.];

[0103] AutoInt: is a model that automatically learns interactive features for tabular data. It uses the self-attention mechanism to establish connections between multiple features and can be used for multiple tasks such as prediction and missing value filling. Its method and process reference [Sercan Ö. Arik and Tomas Pfister. 2021. TabNet: Attentive InterpretableTabular Learning. In AAAI. 6679–6687];

[0104] GAE: is a graph neural network model. It learns latent variables and performs variational reasoning in a probabilistic form to process graph structure data, and uses a simple Gaussian distribution to model uncertainty on the graph structure. The implementation process can be referred to the references [Thomas N. Kipf and Max Welling. 2016. Variational Graph Auto-Encoders. arXiv:1611.07308];

[0105] GAT: It is a graph neural network model. When aggregating features, it uses the attention mechanism to build the relationship between nodes for the construction and supplement of the graph structure. Its implementation process and details refer to [Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. 2018. Graph Attention Networks. In ICLR];

[0106] Graphomer: It combines graph neural networks with the Transformer framework. It is built on the standard Transformer architecture and can achieve excellent results in a wide range of graph learning tasks. Its implementation process and details can be found in [Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, DiHe, Yanming Shen, and Tie-Yan Liu. 2021. Do Transformers Really Perform Badlyfor Graph Representation?. In NeurIPS];

[0107] It can be seen from the experimental results in Table 1 that the street-level IP address location framework based on graph neural network provided in this embodiment can be used for street-level IP address location tasks. Compared with other most advanced IP location, machine learning methods and graph neural networks, the performance of the present invention in terms of root mean square error, mean absolute error and median error is greatly improved.

[0108] It can be seen that the present invention uses the street-level IP address positioning framework based on graph neural network to predict the latitude and longitude of the IP address for positioning. By inputting the characteristics of the IP address (such as operator information and WHOIS registration information, etc.) and network measurements (such as ping value and traceroute routing path), the location information of the target IP address can be given more accurately. At the same time, the present invention also provides a general framework for IP address positioning based on graph neural network.

[0109] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A street-level IP address location method based on graph neural network, characterized in that: The following steps are involved: Step S1, IP address clustering, screening landmark IPs matching the target IP, and obtaining an area containing the target IP and several landmark IPs as graph nodes; Step S2, IP address connection, from two different perspectives of IP features and network measurement, establish the topological spatial relationship between the target IP and the landmark IP in the area, so as to obtain the edge weight between the target IP and the associated landmark IP, and form a complete topological graph structure; Step S3: IP feature aggregation: using the edge weights and topological graph structure between the target IP and the landmark IP, a graph convolutional neural network is used to aggregate the neighborhood landmark features and geographic coordinates, and a probabilistic representation of the target IP is obtained through a regularized flow to resist the negative impact of network fluctuations, thereby obtaining a robust hidden state of the IP. The specific process of step S3 is as follows: First, a simple Gaussian distribution is obtained using a graph convolutional neural network (GCN): Formula 5, in, The concatenation of the feature vector representing the IP and the longitude and latitude of the landmark. For the target IP to be predicted, (0,0) is used instead. and are two graph convolutional neural networks with different parameters, is a random variable that obeys Gaussian distribution, representing the hidden state of all IP nodes; Then, a continuous regularization flow is used to transform the hidden state Continue to derive a more flexible distribution form to cope with the ever-changing network environment. and is the initial random variable and probability distribution, then the distribution change is expressed as: Formula 6, Formula 7, Formula 8, in, is a hyperparameter representing the span of the distribution transformation, So is a fully connected layer with learnable parameters, represents the trace of a matrix; Finally, the final posterior distribution is optimized by maximizing the evidence lower bound (ELBO), as follows: Formula 9, in, Represents the reconstruction loss between the reconstructed topological graph structure and the original graph structure, which is calculated as follows: Formula 10, In the formula is the Heaviside step function, Represents a fully connected neural network used to reconstruct the topology of the graph from the hidden state; Step S4: IP address location. Use a multi-layer perceptron to perform linear affine transformation, decode the longitude and latitude coordinates of the target IP from the probability representation, that is, the hidden state of the target IP, and complete the IP location task.

2. The street-level IP address location method based on graph neural network according to claim 1 is characterized in that: The specific process of step S1 is as follows: Use the tracert tool on probe machines located in different cities to search for routing paths and record the last-hop router; store the obtained last-hop router in the traceroute list, and then cluster the IP hosts with the same last-hop router in the traceroute list into a graph, that is, obtain an area containing the target IP and several landmark IPs as graph nodes.

3. The street-level IP address positioning method based on graph neural network according to claim 1 is characterized in that: The specific process of obtaining the edge weight between the target IP and the associated landmark IP in step S2 is as follows: Establishing a first edge weight using an existing topology obtained by network measurement between IP addresses in the same area and public routers of the IP addresses in the area; Landmark IPs with similar radius to the target IP are selected to construct the first edge weight which is inversely proportional to the distance difference between the target IP and the landmark IP, expressed as: Formula 1, in, , and They represent the target IP, landmark IP and public router respectively; Through the distance mapping function Map network measurements to geographic distances and calculate and The radial distance difference between The expression of the distance mapping function is as follows: Formula 2, in, and Respectively represent the delay and routing hops between the IP host and the public router, , and is a hyperparameter, which is obtained by learning using a fully connected layer network with network measurement as input.

4. The street-level IP address positioning method based on graph neural network according to claim 3 is characterized in that: The second edge weight is constructed by using the feature similarity between IP addresses. Given the target IP and landmark IP, the feature vectors of any two IPs are concatenated and their dot product is calculated to obtain the similarity between the two IPs as the second edge weight. The expression is as follows: Formula 3, Where T represents the matrix transpose, Represents the concatenation of the target IP and landmark IP feature vectors, and are the parameters learned through the neural network.

5. The street-level IP address positioning method based on graph neural network according to claim 4 is characterized in that: Set the first edge weight and the second edge weight After fusing and controlling the sparsity of the final graph, the topological graph structure is described as follows: Formula 4, in, is a hyperparameter that balances the weight of the first edge with the weight of the second edge, It is the threshold that controls the sparsity of the adjacency matrix of the topological graph structure.

6. The street-level IP address location method based on graph neural network according to claim 1 is characterized in that: In step S4, a multi-layer perceptron (MLP) is used to perform linear affine transformation to predict the longitude and latitude of the target IP address to obtain the target IP address, and the expression is: Formula 11, in, is a latitude and longitude tuple representing geographic location information, Represents the parameters obtained through training.

Citation Information

Patent Citations

  • IP positioning method based on Bayes and deep neural network

    CN109858508A

  • Interaction method and device, electronic equipment and storage medium

    CN113014638A

  • Improved method for forming street-level positioning library

    CN113242332A

  • Map Automation - Lane Classification

    US20190138024A1

  • Graph convolutional networks with motif-based attention

    US20200285944A1