Network fraud identification method and system based on hierarchical semantic perception

By employing causal attention and gating update mechanisms, semantic features of the network transaction graph are extracted hierarchically, addressing the issues of semantic aliasing and risk clue dilution in existing methods, thereby improving the accuracy and robustness of network fraud detection.

CN121907574APending Publication Date: 2026-04-21CHONGQING POLICE VOCATIONAL COLLEGE
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610095154.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-01-23
Publication Date
2026-04-21

AI Technical Summary

Technical Problem

Existing network fraud detection methods based on graph neural networks struggle to effectively distinguish between different relationship types and different semantic levels, leading to semantic aliasing and dilution of key risk clues in heterogeneous fraud networks.

Method used

A hierarchical semantic perception-based approach is adopted. The causal attention module calculates the causal correlation strength of node feature representations, generates attention weights, forms a neighborhood set, and performs hierarchical semantic feature extraction through an attention aggregation information strategy. Combined with a gating update module, dynamic weighted updates are performed to form a deep network representation learning structure.

Benefits of technology

Explicitly characterizes the multi-source, heterogeneous, and multi-layered semantic information in the network transaction graph, reduces noise interference, and improves the model's robustness and detection accuracy against complex fraud patterns.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121907574A_ABST
    Figure CN121907574A_ABST
Patent Text Reader

Abstract

The invention provides a hierarchical semantic perception-based network fraud identification method and system. The method comprises the following steps of: constructing a network transaction graph based on network transaction data; carrying out characterization representation on nodes and edges of the network transaction graph; a causal attention module is adopted to calculate causal association strength represented by the node features, and an attention weight is generated based on the causal association strength; carrying out importance measurement and sorting on neighbor nodes through attention weights to form a neighborhood set; performing hierarchical semantic feature extraction on the neighborhood set by adopting an attention aggregation information strategy; the hierarchical semantic features are input into a gating updating module, and dynamic weighted updating is carried out on information from a neighborhood and historical representation of nodes according to gating coefficients; a deep network representation learning structure is formed in the stacking hierarchical semantic extraction and gating updating process, and final feature representation of the nodes is generated; inputting the node feature representation into a classifier to obtain a prediction result; layered modeling is carried out for multi-source, heterogeneous and multi-layer semantic information in a network transaction relation graph, and interference of noise propagation on a detection result is reduced by explicitly describing risk contributions of different semantic levels and different relation types.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of network security and fraud detection technology, specifically relating to a network fraud identification method and system based on hierarchical semantic perception. Background Technology

[0002] With the rapid development of information technology, online fraud has become increasingly frequent, causing huge losses to individuals and businesses. Online fraud takes many forms, including identity theft, fraudulent transactions, and credit card fraud, and typically involves complex network structures and large amounts of transaction data. This data can usually be represented as a graph structure, including social networks and financial transaction networks. In these graph structures, nodes represent entities such as accounts, merchants, devices, IP addresses, and orders, while edges represent relationships such as transactions, logins, and account bindings.

[0003] Graph Neural Networks (GNNs) are powerful tools for processing graph data, effectively capturing the relationships between nodes and learning graph representations in a continuous embedding space. This enables GNNs to demonstrate excellent performance in tasks such as online fraud detection, social network analysis, and recommender systems. Through multi-layer stacking and non-linear activation functions, GNNs can extract complex graph structure features, thereby achieving efficient classification and prediction.

[0004] Therefore, many scholars have begun to study network fraud detection based on graph neural networks. These methods learn the representation vectors of nodes / subgraphs by employing a unified neighborhood aggregation and message passing mechanism, thereby detecting suspicious accounts, transactions, or groups. Compared to traditional methods that rely on manual rules and static feature engineering, deep models based on graph neural networks and various variants can make fuller use of contextual information in relational networks, showing certain advantages in representing and characterizing relational patterns in complex scenarios. The key to these methods lies in two points: first, how to effectively represent multi-layered and multi-source semantic information in network fraud scenarios; and second, how to design reasonable neighborhood information aggregation and feature update strategies so that the model can highlight key risk clues under complex topology and noise interference. Existing research methods mostly follow a homogeneous aggregation paradigm, making it difficult to explicitly distinguish the contributions of different relation types and different semantic levels of information, thus easily causing semantic aliasing and dilution of key risk clues in heterogeneous fraud networks. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes a network fraud identification method based on hierarchical semantic awareness. The method includes: acquiring network transaction data; constructing a network transaction graph based on the data; performing feature representation on the nodes and edges of the network transaction graph; calculating the causal correlation strength of the node feature representations using a causal attention module, and generating attention weights based on the causal correlation strength; measuring and ranking the importance of neighboring nodes using the attention weights to form a neighborhood set; extracting hierarchical semantic features from the neighborhood set using an attention aggregation information strategy; inputting the hierarchical semantic features into a gating update module, dynamically weighting and updating the information from the neighborhood and the node's own historical representation based on the gating coefficient; stacking the hierarchical semantic extraction and gating update processes to form a deep network representation learning structure, generating the final feature representation of the nodes; and inputting the node feature representation into a classifier to obtain prediction results.

[0006] A network fraud identification system based on hierarchical semantic perception, the system includes: a data acquisition module, a data transformation module, a feature representation module, a causal attention module, a node filtering module, an attention aggregation information strategy module, a gating update module, a deep network representation learning module, a prediction module, and an output module;

[0007] The data acquisition module is used to acquire network transaction data;

[0008] The data conversion module is used to convert network transaction data into a network transaction transaction graph.

[0009] The feature representation module is used to perform feature representation on the nodes and edges of the network transaction graph;

[0010] The causal attention module is used to calculate the causal association strength of node feature representations and generate attention weights based on the causal association strength.

[0011] The node filtering module is used to measure and sort the importance of neighboring nodes according to attention weights to form a neighborhood set;

[0012] The attention aggregation information strategy module is used to extract hierarchical semantic features from the neighborhood set;

[0013] The gating update module dynamically weights and updates the information from the neighborhood and the node’s own historical representation based on the gating coefficient.

[0014] The deep network representation learning module is used to form a deep network representation learning structure for the stacked hierarchical semantic extraction and gating update process, and generate the final feature representation of the node;

[0015] The prediction results are used to perform classification prediction on the final feature representation to obtain the prediction results;

[0016] The output module is used to output the prediction results.

[0017] The beneficial effects of this invention are:

[0018] This invention proposes a network fraud detection method based on hierarchical semantic information perception. Unlike existing fraud detection methods that mostly employ unified neighborhood aggregation and homogeneous message passing, this invention performs hierarchical modeling of multi-source, heterogeneous, and multi-layered semantic information in the network transaction relationship graph. By explicitly characterizing the risk contribution of different semantic levels and different relationship types, it reduces the interference of noise propagation on the detection results. First, it uses a causal attention mechanism to mine the causal relationships between nodes in the network transaction graph and identify the importance of different neighboring nodes. Second, based on node importance, it adopts an attention aggregation information strategy to extract and fuse the semantic features of different neighboring nodes hierarchically, strengthening the implicit pattern representation of fraudulent behavior. Finally, it designs a gating update mechanism to dynamically adjust information updates, effectively alleviating the oversmoothing problem in deep networks and improving the robustness of the model to complex fraud patterns. Attached Figure Description

[0019] Figure 1 This is a framework diagram of the network fraud detection method based on hierarchical semantic information perception of the present invention;

[0020] Figure 2 This is an example diagram of the network transaction graph described in this invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] This invention proposes a network fraud identification method based on hierarchical semantic awareness. The method includes: acquiring network transaction data; constructing a network transaction graph based on the data; performing feature representation on the nodes and edges of the network transaction graph; calculating the causal correlation strength of the node feature representations using a causal attention module, and generating attention weights based on the causal correlation strength; measuring and ranking the importance of neighboring nodes using the attention weights to form a neighborhood set; extracting hierarchical semantic features from the neighborhood set using an attention aggregation information strategy; inputting the hierarchical semantic features into a gating update module, dynamically weighting and updating the information from the neighborhood and the node's own historical representation based on the gating coefficient; stacking the hierarchical semantic extraction and gating update processes to form a deep network representation learning structure, generating the final feature representation of the nodes; and inputting the node feature representation into a classifier to obtain prediction results.

[0023] A network fraud detection method based on hierarchical semantic information perception, such as Figure 1 As shown, the specific steps include, but are not limited to, the following:

[0024] S1. Acquire network transaction data and construct a network transaction graph, and perform feature representation on nodes and edges; S2. Calculate the causal correlation strength between nodes using a causal attention module to generate attention weights; S3. Measure and rank the importance of neighboring nodes using attention weights to form a neighborhood set; S4. Use an attention aggregation information strategy to extract hierarchical semantic features from the neighborhood set; S5. Input the hierarchical semantic features into a gating update module, and dynamically update the information from the neighborhood and the node's own historical representation based on the gating coefficient; S6. Stack the above hierarchical semantic extraction and gating update processes to form a deep network representation learning structure, generating the final feature representation of the nodes; S7. Input the node feature representation into a classifier to obtain the prediction result, and iteratively optimize through a loss function. This invention has strong robustness and generalization ability, and can effectively improve the accuracy of network fraud detection.

[0025] In this embodiment, a network transaction graph can be represented as follows: ,in Represents a set of nodes (entities). The total number of nodes. Represents the node feature matrix, A collection of node labels. Represents the set of edges.

[0026] In this embodiment, the generation of attention weights by calculating the causal correlation strength between nodes through the causal attention module includes:

[0027] S21. Input the transaction graph into the causal attention module and obtain the attention score by calculating the causal relationship between nodes;

[0028] Specifically, this includes: assigning appropriate attention weights to each neighbor node, measuring its importance in the causal context, and for the target node... and its neighboring nodes The entire calculation process can be represented as:

[0029]

[0030] in, For a learnable parameter matrix, It is an activation function. Let i be the feature vector of node i. The feature vectors of the neighboring nodes, This represents the concatenation operation, where N is the number of neighboring nodes. The weights are averaged across multiple attention heads to obtain the final attention score for each neighboring node, which is then normalized. The calculation formula is as follows:

[0031]

[0032]

[0033] Where H represents the number of attention heads. Neighboring nodes For nodes Importance score.

[0034] S22. Normalize the multiple attention heads to obtain the final attention weights of the neighboring nodes;

[0035] S23. Use the node importance weights to filter the nodes in the original graph to obtain a set of important neighbor nodes.

[0036] In this embodiment, an attention aggregation information strategy is used to extract hierarchical semantic features from the neighborhood set;

[0037] S41. Map the initial features and self-representation of the important neighbor node set to the query vector, key vector, and value vector;

[0038] Specifically, this includes: targeting nodes The initial feature representation and self-representation of the important neighbor node subset H(i) are mapped to the query vector. Key vector Sum value vector Attention coefficients are calculated using dot product similarity to generate the aggregated message representation. The calculation formula is as follows:

[0039]

[0040]

[0041]

[0042]

[0043] in and This represents the corresponding weight matrix.

[0044] S42. The attention coefficient is obtained by calculating the dot product similarity, and the message representation of the node is generated by using the Softmax normalization operation;

[0045] S43. Parallel computation is performed using a multi-head attention mechanism, and an aggregated result is formed through a splicing operation.

[0046] S44. Enhance message representation using two-layer linear transformation and Gaussian error linear unit activation function.

[0047] Specifically, this includes: utilizing a two-layer linear transformation and the Gaussian Error Linear Unit (GELU) activation function to enhance the model's nonlinear modeling ability and the sufficiency of feature representation. The calculation formula is as follows:

[0048]

[0049]

[0050] in, and The weight matrix is ​​the linear transformation matrix. This represents the activation function. Indicates bias.

[0051] In this embodiment, inputting hierarchical semantic features into the gating update module includes:

[0052] S51, using the reset door and Update Gate Control the retention ratio of historical information and important neighbor aggregation information separately;

[0053] Specifically, this includes: utilizing the reset door and Update Gate The calculation process for control historical information and important neighbor aggregation information is as follows:

[0054]

[0055]

[0056] in, For the Sigmoid activation function, [ The symbol ] represents a vector concatenation operation.

[0057] S52. The historical representation and neighbor information are weighted and fused through a gating mechanism to obtain the updated hidden state.

[0058] The calculation formula for weighted fusion of historical representation and neighbor information using a gating mechanism is as follows:

[0059]

[0060]

[0061] in, Let be the hidden state of node i at step t. This is the update gate for node i at step t. Let be the historical representation of node i at step t-1, and tanh be the activation function. The weight matrix is ​​a learnable matrix. To reset the door, This represents the neighbor information of node i at step t. It is a learnable bias vector.

[0062] In this embodiment, node representations are input into a multilayer perceptron for classification prediction, gradients are calculated using the cross-entropy loss function, and model parameters are iteratively optimized.

[0063] The model's loss function is:

[0064]

[0065]

[0066] In this embodiment, the present invention proposes a network fraud detection method based on hierarchical semantic information perception, such as... Figure 1 As shown, the specific steps include, but are not limited to, the following: First, the data is processed and constructed into a network transaction graph, such as... Figure 2 As shown, nodes and edges are characterized. Next, the causal attention module calculates the strength of causal relationships between nodes to generate attention weights. After obtaining the set of important neighbor nodes, an attention aggregation strategy is used to extract hierarchical semantic features from the neighborhood set. Then, the hierarchical semantic features are input into the gating update module, which updates the information from the neighborhood and the node's own historical representation based on the gating coefficients. Finally, modules S2-S5 are stacked to form a deep network representation learning structure. The node representations are input into a multilayer perceptron for classification and prediction. The gradient is calculated using the cross-entropy loss function, and the model parameters are iteratively optimized.

[0067] A network fraud identification system based on hierarchical semantic perception, the system includes: a data acquisition module, a data transformation module, a feature representation module, a causal attention module, a node filtering module, an attention aggregation information strategy module, a gating update module, a deep network representation learning module, a prediction module, and an output module;

[0068] The data acquisition module is used to acquire network transaction data;

[0069] The data conversion module is used to convert network transaction data into a network transaction transaction graph.

[0070] The feature representation module is used to perform feature representation on the nodes and edges of the network transaction graph;

[0071] The causal attention module is used to calculate the causal association strength of node feature representations and generate attention weights based on the causal association strength.

[0072] The node filtering module is used to measure and sort the importance of neighboring nodes according to attention weights to form a neighborhood set;

[0073] The attention aggregation information strategy module is used to extract hierarchical semantic features from the neighborhood set;

[0074] The gating update module dynamically weights and updates the information from the neighborhood and the node’s own historical representation based on the gating coefficient.

[0075] The deep network representation learning module is used to form a deep network representation learning structure for the stacked hierarchical semantic extraction and gating update process, and generate the final feature representation of the node;

[0076] The prediction results are used to perform classification prediction on the final feature representation to obtain the prediction results;

[0077] The output module is used to output the prediction results.

[0078] The system implementation method of the present invention is the same as the method implementation method.

[0079] It should be noted that those skilled in the art should understand that all or part of the steps in the above method embodiments can be implemented by a computer program, and the program can control the corresponding hardware device to execute them. The program can be stored in a computer-readable storage medium, and when the program runs, it can execute the processes corresponding to the above method embodiments. The computer-readable storage medium includes, for example, a hard disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0080] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A network fraud identification method based on hierarchical semantic awareness, characterized in that, include: Obtain network transaction data, construct a network transaction graph based on the network transaction data, and perform feature representation on the nodes and edges of the network transaction graph; A causal attention module is used to calculate the causal association strength of node feature representations and generate attention weights based on the causal association strength; the importance of neighboring nodes is measured and ranked by the attention weights to form a neighborhood set; and a hierarchical semantic feature extraction is performed on the neighborhood set using an attention aggregation information strategy. The hierarchical semantic features are input into the gating update module, and the information from the neighborhood and the node’s own historical representation are dynamically weighted and updated according to the gating coefficient. The stacked hierarchical semantic extraction and gating update process forms a deep network representation learning structure, generating the final feature representation of the nodes; the node feature representation is then input into the classifier to obtain the prediction result.

2. The network fraud identification method based on hierarchical semantic awareness according to claim 1, characterized in that, The calculation of the causal relationship strength represented by node features includes: inputting the transaction graph into the causal attention module, obtaining attention scores by calculating the causal relationships between nodes; normalizing multiple attention heads to obtain the final attention weights of neighboring nodes; and using the node importance weights to filter the nodes in the original graph to obtain a set of important neighboring nodes.

3. The network fraud identification method based on hierarchical semantic awareness according to claim 2, characterized in that, The causal relationships between computing nodes are as follows: ; in, For a learnable parameter matrix, It is an activation function. Let i be the feature vector of node i. The feature vectors of the neighboring nodes, This indicates a splicing operation, where N is the number of neighboring nodes.

4. The network fraud identification method based on hierarchical semantic awareness according to claim 2, characterized in that, Normalization of multiple attention heads includes: ; ; Where H represents the number of attention heads. Neighboring nodes For nodes Importance score.

5. The network fraud identification method based on hierarchical semantic awareness according to claim 1, characterized in that, The hierarchical semantic feature extraction of the neighborhood set is performed using an attention aggregation information strategy, which includes: mapping the initial features and self-representation of the important neighbor node set to query vector, key vector, and value vector; calculating the attention coefficient through dot product similarity and generating the message representation of the node using Softmax normalization; using a multi-head attention mechanism for parallel computation and forming the aggregation result through concatenation; and strengthening the message representation using two-layer linear transformation and Gaussian error linear unit activation function.

6. The network fraud identification method based on hierarchical semantic awareness according to claim 5, characterized in that, Mapping the initial features and their own representation to the query vector, key vector, and value vector includes: mapping the target node... The initial feature representation and self-representation of the important neighbor node subset H(i) are mapped to the query vector. Key vector Sum value vector Attention coefficients are calculated using dot product similarity to generate the aggregated message representation. The calculation formula is as follows: ; ; ; ; in and This represents the corresponding weight matrix.

7. The network fraud identification method based on hierarchical semantic awareness according to claim 5, characterized in that, The message representation is enhanced using a two-layer linear transformation and a Gaussian error linear unit activation function, including: ; ; in, and The weight matrix is ​​the linear transformation matrix. This represents the activation function. Indicates bias.

8. The network fraud identification method based on hierarchical semantic awareness according to claim 1, characterized in that, The gate update module processes the input data, including by resetting the gate. and Update Gate The retention ratios of historical information and important neighbor aggregation information are controlled separately; the historical representation and neighbor information are weighted and fused through a gating mechanism to obtain the updated hidden state.

9. A network fraud identification method based on hierarchical semantic awareness according to claim 8, characterized in that, The historical representation and neighbor information are weighted and fused using a gating mechanism to form: ; ; in, Let be the hidden state of node i at step t. This is the update gate for node i at step t. Let be the historical representation of node i at step t-1, and tanh be the activation function. The weight matrix is ​​a learnable matrix. To reset the door, This represents the neighbor information of node i at step t. It is a learnable bias vector.

10. A network fraud identification system based on hierarchical semantic awareness, the system being used to execute the network fraud identification method based on hierarchical semantic awareness as described in any one of claims 1 to 9, characterized in that, The system includes: a data acquisition module, a data transformation module, a feature representation module, a causal attention module, a node selection module, an attention aggregation information strategy module, a gating update module, a deep network representation learning module, a prediction module, and an output module; The data acquisition module is used to acquire network transaction data; The data conversion module is used to convert network transaction data into a network transaction transaction graph. The feature representation module is used to perform feature representation on the nodes and edges of the network transaction graph; The causal attention module is used to calculate the causal association strength of node feature representations and generate attention weights based on the causal association strength. The node filtering module is used to measure and sort the importance of neighboring nodes according to attention weights to form a neighborhood set; The attention aggregation information strategy module is used to extract hierarchical semantic features from the neighborhood set; The gating update module dynamically weights and updates the information from the neighborhood and the node’s own historical representation based on the gating coefficient. The deep network representation learning module is used to form a deep network representation learning structure for the stacked hierarchical semantic extraction and gating update process, and generate the final feature representation of the node; The prediction results are used to perform classification prediction on the final feature representation to obtain the prediction results; The output module is used to output the prediction results.