A method, system, device, and medium for fusion characterization of multi-source heterogeneous network threat data.

By combining contrastive learning and graph attention networks, threat feature vectors are generated and node embedding representations are optimized, which solves the problem of low recognition accuracy of multi-source heterogeneous network threat detection models in dynamic environments and achieves efficient recognition and stable detection of complex attack chains.

CN123122883APending Publication Date: 2026-11-17GUANGXI POWER GRID CORP
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611194941.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-07
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing network threat detection models struggle to effectively capture dynamic attack chains and complex attack topologies in multi-source heterogeneous, dynamically evolving network environments with noise interference. Furthermore, they are poorly adaptable to semantic differences and environmental changes in multi-source heterogeneous data, resulting in low identification accuracy.

Method used

A contrastive learning preprocessing mechanism is used to generate threat feature vectors, a node update model is established through a graph attention network, and the node embedding representation is optimized through an adversarial regularization loss function. Finally, a classifier is used to predict threat behavior.

Benefits of technology

It improves the ability to learn semantic associations from multi-source heterogeneous data, enhances the ability to represent complex threat patterns, improves the accuracy and stability of threat detection, and can maintain efficient identification of complex attack chains in dynamic environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123122883A_ABST
    Figure CN123122883A_ABST
Patent Text Reader

Abstract

This invention discloses a method, system, device, and medium for fusing and representing multi-source heterogeneous network threat data, relating to the field of network threat detection technology. The specific steps are as follows: collecting multi-source heterogeneous network threat data to generate meta-event sequences; establishing a contrastive learning preprocessing mechanism to preprocess the meta-event sequences to obtain threat feature vectors; performing sequence segmentation; using a graph attention network to establish a node update model; outputting node embedding representations that fuse temporal and topological information; establishing a joint optimization loss function to robustly optimize the node embedding representations; and inputting the optimized node embedding representations into a classifier for threat behavior prediction and threat type discrimination. This invention achieves unified fusing and representing of multi-source heterogeneous threat data. By constructing graph structure fusion modeling and a differential robust optimization mechanism, it improves the system's accuracy in identifying complex attack chains and enhances the system's adaptability and stability in dynamic environments.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of network threat detection technology, and in particular to a method, system, device and medium for the fusion and characterization of multi-source heterogeneous network threat data. Background Technology

[0002] As cyberattack techniques evolve and network system complexity increases, the data sources required for cybersecurity monitoring are becoming increasingly complex. Cyber ​​threat data may originate from network traffic probes, host audit logs, alarms from various security devices, and external threat intelligence platforms. It exhibits characteristics of being multi-sourced, heterogeneous, and strongly semantically correlated. Various types of cyber threat data show significant differences in format and semantics, and attack behaviors are complex. Existing methods for fusion and characterization of cyber threat data have the following shortcomings in this scenario: Existing threat detection models have limited ability to capture dynamic attack chains. Most of these models employ typical machine learning algorithms and are usually based on manually extracted static features. Therefore, they are weak in representing dynamic attack chains and cannot characterize the temporal features of threat events. Furthermore, for complex attack topologies composed of multiple entity interactions, existing threat detection models only focus on entity structure; they lack the ability to fuse temporal patterns with graph structure relationships, resulting in low accuracy in identifying complex attack scenarios.

[0003] Current feature representation techniques are poorly adaptable to dynamic environments. For example, principal component analysis is designed for single-source data and cannot effectively represent data features when dealing with semantic differences caused by multi-source heterogeneous data. More importantly, existing feature representation techniques lack consideration for natural data drift and changes in the operating environment. Threat detection models are prone to performance degradation in actual deployments due to slight fluctuations in input data. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is: how to obtain threat feature vectors by introducing a contrastive learning preprocessing mechanism in a network operating environment that is multi-source, heterogeneous, dynamically evolving, and subject to noise interference, establish a node update model by using a graph attention network, and finally predict threat behavior using a classifier.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a method for fusing and characterizing multi-source heterogeneous network threat data, comprising, The target data is transformed into a meta-event sequence, which is standardized multi-source threat data; In response to the acquisition of the meta-event sequence, a threat feature vector is obtained, which is obtained by introducing a contrastive learning preprocessing mechanism; The steps of the contrastive learning preprocessing mechanism are as follows: construct positive and negative sample pairs based on the meta-event sequence, input the positive and negative sample pairs into the encoder for training, and output a threat feature vector; An attack behavior graph is constructed based on threat feature vectors, and a node update model is established using a graph attention network. The input is the attack behavior graph, and the output is the updated node embedding representation. An adversarial regularization loss function is established to optimize the updated node embedding representation; The optimized node embedding representation is input into the classifier for threat behavior prediction and threat type identification.

[0007] As a preferred embodiment of the multi-source heterogeneous network threat data fusion and characterization method described in this invention, wherein: in response to the acquisition of the meta-event sequence, a threat feature vector is obtained, and the threat feature vector is obtained by introducing a contrastive learning preprocessing mechanism, including: For a threat event in the meta-event sequence, generate two enhanced samples as positive sample pairs, and the remaining threat events as negative sample pairs; The encoder is trained using positive and negative samples, and a contrastive loss function is introduced to optimize the training process, outputting a threat feature vector.

[0008] This invention introduces a contrastive learning framework to generate positive and negative sample pairs for threat events in a meta-event sequence using different enhancement functions. Furthermore, an encoder is used to automatically learn semantic relationships between multi-source heterogeneous data under unsupervised conditions. This unifies the feature representation space of multi-source heterogeneous data. Secondly, highly discriminative threat feature vectors are generated, enhancing the system's ability to represent complex threat patterns.

[0009] As a preferred embodiment of the multi-source heterogeneous network threat data fusion and characterization method described in this invention, the step of constructing an attack behavior graph based on threat feature vectors, establishing a node update model using a graph attention network, and inputting the attack behavior graph to output the updated node embedding representation includes: The threat feature vector is segmented into behavioral fragment subsequences, and the attack behavior patterns of the behavioral fragment subsequences are mined. Based on the discovered attack behavior patterns, construct an attack behavior map; A node update model is established using a graph attention network, which takes an attack behavior graph as input and outputs an updated node embedding representation.

[0010] This invention employs a mining algorithm to uncover attack behavior patterns from threat feature vectors. It further constructs an attack behavior graph based on these patterns and then uses a graph attention network to establish a node update model, aggregating semantic information of nodes and their corresponding neighbors using the graph attention network's attention mechanism. By fusing the temporal evolution characteristics of threat events with the entity topology, the attack path is reconstructed, enhancing the system's ability to identify complex attack chains such as persistent threats.

[0011] As a preferred embodiment of the multi-source heterogeneous network threat data fusion and characterization method described in this invention, the step of establishing an adversarial regularization loss function to optimize the updated node embedding representation includes: Add perturbations to the updated node embedding representations to construct perturbation samples; Establish an adversarial regularization loss function to minimize the consistency loss between perturbed samples and node embedding representations.

[0012] This invention constructs perturbation samples and establishes an adversarial regularization loss function to minimize the consistency loss between the perturbation samples and the node embedding representation, forcing the threat detection model to learn essential feature representations that are insensitive to abnormal fluctuations. Furthermore, by establishing a joint optimization loss function to robustly optimize the node embedding representation, the threat detection model can improve the stability of the learned representations while maintaining threat identification accuracy.

[0013] As a preferred embodiment of the multi-source heterogeneous network threat data fusion and characterization method described in this invention, the step of inputting the optimized node embedding representation into the classifier for threat behavior prediction and threat type discrimination includes: The optimized node embedding representation is input into the classifier, and the classifier outputs the predicted probability of each threat behavior category; Based on the predicted probability of each threat behavior category, determine whether the threat behavior threshold is exceeded and trigger the linkage response strategy. A cross-entropy loss function is established to optimize the classifier, minimizing the cross-entropy loss between the predicted label and the true label.

[0014] As a preferred embodiment of the multi-source heterogeneous network threat data fusion and characterization method described in this invention, the step of training the input encoder with positive and negative samples, introducing a contrastive loss function to optimize the training process, and outputting a threat feature vector includes: The optimized training process is achieved by calculating the contrastive learning loss value between positive and negative samples; The contrastive learning loss value between positive and negative samples is obtained by calculating the loss value between positive sample pairs and the loss value between positive and negative sample pairs; The calculation of the loss value between positive sample pairs and the calculation of the loss value between positive and negative sample pairs require the introduction of a temperature parameter to adjust the smoothness of the similarity distribution.

[0015] As a preferred embodiment of the multi-source heterogeneous network threat data fusion and characterization method described in this invention, the step of establishing a node update model using a graph attention network, and inputting an attack behavior graph to output an updated node embedding representation, includes: A node update model is established using a graph attention network, expressed as follows: , in, For nodes The set of neighboring nodes, For nodes With nodes Attention weights between them For learnable parameter matrix, For activation function, This represents the index of the network layer in a graph neural network. Represents a node The neighboring nodes, Represents a node The updated node embedding representation in the next layer of the network.

[0016] This invention provides a multi-source heterogeneous network threat data fusion characterization system.

[0017] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a multi-source heterogeneous network threat data fusion characterization system, comprising: a collection module, a preprocessing module, an update module, an optimization module, and a threat prediction module; The acquisition module converts target data into a meta-event sequence, which is standardized multi-source threat data. The preprocessing module responds to the acquisition of the meta-event sequence to obtain a threat feature vector. The threat feature vector is obtained by introducing a contrastive learning preprocessing mechanism. The steps of the contrastive learning preprocessing mechanism are to construct positive and negative sample pairs based on the meta-event sequence, input the positive and negative sample pairs into the encoder for training, and output the threat feature vector. The update module constructs an attack behavior graph based on threat feature vectors, establishes a node update model using a graph attention network, and outputs an updated node embedding representation as input to the attack behavior graph. The optimization module establishes an adversarial regularization loss function to optimize the updated node embedding representation; The threat prediction module embeds the optimized nodes into the input classifier to predict threat behavior and determine threat type.

[0018] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of the multi-source heterogeneous network threat data fusion characterization method.

[0019] The present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the multi-source heterogeneous network threat data fusion characterization method.

[0020] The beneficial effects of this invention are as follows: To enable the threat detection model to learn semantic relationships between multi-source heterogeneous data, this invention performs unified standardization processing on multi-source heterogeneous threat data and introduces a contrastive learning framework to achieve the fusion of multi-source heterogeneous data. Furthermore, it employs sequence pattern mining algorithms and graph neural networks to construct attack behavior graphs, capturing entity relationships between threat events. Secondly, it optimizes the threat detection model's ability to remain stable under noisy perturbation scenarios by jointly optimizing the loss function. Finally, it extracts highly discriminative and stable features from complex network threat data, thereby improving threat detection accuracy. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the overall process of a multi-source heterogeneous network threat data fusion characterization method according to an embodiment of the present invention. Detailed Implementation

[0023] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0024] Example 1, referring to Figure 1 As one embodiment of the present invention, this embodiment provides a method for fusion characterization of multi-source heterogeneous network threat data, including: S1: Transform the target data into a meta-event sequence, wherein the meta-event sequence is standardized multi-source threat data; S2: In response to the acquisition of the meta-event sequence, a threat feature vector is obtained, which is obtained by introducing a contrastive learning preprocessing mechanism; S3: Construct an attack behavior graph based on threat feature vectors, and use a graph attention network to establish a node update model. Input the attack behavior graph and output the updated node embedding representation. S4: Establish an adversarial regularization loss function to optimize the updated node embedding representation; S5: Embed the optimized nodes into the classifier to predict threat behavior and determine threat type.

[0025] Example 2, an embodiment of the present invention, provides a method for fusion and characterization of multi-source heterogeneous network threat data based on the previous embodiment, including: In step S1, the target data is transformed into a meta-event sequence, which is standardized multi-source threat data, including the following steps: Specifically, the target data includes, but is not limited to: network traffic data, host behavior logs, security device alarm information, and attack chain context information.

[0026] Furthermore, we introduce format regularization and field mapping methods to standardize different types of target data into a unified JSON structure and abstract it into a meta-event sequence.

[0027] In step S2, in response to the acquisition of the meta-event sequence, a threat feature vector is obtained. The threat feature vector is obtained by introducing a contrastive learning preprocessing mechanism, including the following steps A1-A2: A1: For a threat event in the meta-event sequence, generate two enhanced samples as positive sample pairs, and the remaining threat events as negative sample pairs.

[0028] Furthermore, for a threat event in the meta-event sequence, two different enhancement functions are used to generate two enhanced samples for this threat event. The enhancement functions can employ random field masking, timestamp perturbation, field value replacement, or log structure shuffling. The resulting enhanced sample expression is as follows: , , in, Indicates the first event in the sequence of meta-events A sample of threat incidents, This indicates the index number of the threat event sample in the meta-event sequence. This represents two different data augmentation functions used to apply different perturbations to the same threat event sample. Indicates the first event in the sequence of meta-events Two augmented samples are generated from a threat event sample using different augmentation functions, and the two samples form a positive sample pair.

[0029] Furthermore, the remaining threat event samples in the meta-event sequence are used as negative samples.

[0030] A2: Train the encoder with positive and negative samples, introduce a contrastive loss function to optimize the training process, and output a threat feature vector.

[0031] Furthermore, the encoder employs a multilayer perceptron network structure. The threat event feature representations corresponding to the positive sample pairs constructed in step A1 are input into the input layer of the multilayer perceptron network. The multilayer perceptron network performs a linear transformation on the input features through at least two fully connected layers, and after each fully connected layer, a non-linear activation function is used to perform layer-by-layer mapping and semantic compression of the threat event features. After multi-layer mapping, the multilayer perceptron network outputs a threat feature vector, which is used to represent the positional relationship of the corresponding threat events in the latent semantic space.

[0032] Furthermore, a contrastive loss function is introduced. By increasing the similarity between different enhanced samples of the same threat event and reducing the similarity difference between samples of different threat events, the encoder can learn threat semantic features with discriminative capabilities and output a threat feature vector. The expression is as follows: , in, Indicates the first event in the sequence of meta-events A sample of threat events and its corresponding positive sample The contrastive learning loss value between them This represents a temperature parameter used to adjust the smoothness of the similarity distribution, preventing the gradient from being too large or too small. This indicates the number of original samples in the current training batch. This represents the total number of samples obtained after two enhancements in a batch. This represents the index number of any sample in the batch. Indicates the first in the batch The embedding vector of each sample, Indicates an indicator function, when When the value is 1, it is used to exclude the current positive sample pair from itself. The cosine similarity is expressed as: , It should be noted that this invention utilizes a contrastive learning preprocessing mechanism and an adaptive enhancement strategy to construct positive and negative sample pairs. This drives the encoder to automatically extract essential semantic features from multi-source heterogeneous original threat events. The contrastive loss function maximizes the similarity between positive samples while minimizing the confusion between negative samples. It automatically learns the consistency and difference features of data semantics, effectively overcoming the limitations of traditional methods that rely on manual annotation and are difficult to integrate heterogeneous semantics.

[0033] In step S3, an attack behavior graph is constructed based on the threat feature vector, and a node update model is established using a graph attention network. The attack behavior graph is input and the updated node embedding representation is output, including the following steps B1-B3: B1: Sequence segmentation of the threat feature vector is performed to obtain behavioral segment subsequences, and attack behavior patterns of the behavioral segment subsequences are mined.

[0034] Furthermore, the threat feature vector is segmented using a sliding window and behavioral fragments are constructed. Let an original time-series event stream be: , in, Represents the original time series event stream, each This represents a timestamped threat event (such as login failure, port scan, abnormal data transmission, etc.). The window size can be set. Divided into behavioral segment subsequences: , Furthermore, a sequence pattern mining algorithm is used to extract high-frequency attack patterns from all segments. This algorithm employs a frequent sequence pattern mining method, statistically analyzing the subsequences of behavioral segments obtained by sliding window segmentation. Each subsequence of behavioral segments is represented as a sequence sample consisting of multiple threat events arranged in chronological order. All sequence samples are traversed, and the frequency of different event sequence patterns in all behavioral segments is counted. When the frequency of a certain event sequence pattern exceeds a preset support threshold, that event sequence pattern is identified as an attack behavior pattern.

[0035] It should be noted that the support threshold is set based on experience, and the attack behavior pattern is used to characterize attack behavior features that recur within multiple time windows.

[0036] B2: Construct an attack behavior map based on the discovered attack behavior patterns.

[0037] Furthermore, in order to represent the entity relationships between different events in the spatial dimension, this invention constructs each threat event as graph-structured data, expressed as: , in, This represents graph-structured data, where V is the set of nodes and E is the set of edges.

[0038] It should be noted that graph-structured data consists of a set of nodes and a set of edges. Each node in the node set corresponds to a threat event. Edges are used to represent the relationships between threat events.

[0039] Furthermore, the edge construction rules include: if two threat events occur sequentially and belong to the same attack context, a connection is established between the corresponding nodes; if two threat events originate from the same source address or target the same asset, a connection is established between the corresponding nodes; if two threat events exhibit adjacent attack behavior patterns in the sequence pattern mining results, a connection is established between the corresponding nodes; and if a causal relationship exists between two threat events and has been manually confirmed, a connection is established between the corresponding nodes.

[0040] B3: A node update model is established using a graph attention network. The input is the attack behavior graph, and the output is the updated node embedding representation.

[0041] Furthermore, a graph attention network is used to establish a node update model, expressed as: , in, For nodes The set of neighboring nodes, For nodes With nodes The attention weights between them represent their relevance in the semantic graph. For learnable parameter matrix, It is the ReLU activation function. This represents the index of the network layer in a graph neural network. Represents a node The neighboring nodes, Represents a node The updated node embedding representation in the next layer of the network.

[0042] It should be noted that this invention constructs an attack behavior graph and introduces a graph attention network to establish a node update model. In the node update model, nodes are used to represent threat events, and edges are used to represent entity associations between various threat events. By using the graph attention mechanism to aggregate the semantic information between neighboring nodes, the temporal context and network topology of various threat events are integrated into a unified node embedding representation, which solves the problem that traditional methods are difficult to simultaneously characterize the temporal evolution and entity topology associations between threat events.

[0043] In step S4, an adversarial regularization loss function is established to optimize the updated node embedding representation, including the following steps C1-C2: C1: Add perturbations to the updated node embedding representation and construct perturbation samples.

[0044] Specifically, the expression for constructing the perturbation sample is: , , in, Represents the nodes in the graph The original embedding representation, This represents the perturbed graph node embedding representation, where v represents the node identifier in the graph structure data. With a mean of zero, For disturbance intensity parameters, This represents a Gaussian distribution, used to describe the random distribution characteristics of the perturbation vector. This represents the square of the disturbance intensity parameter. It is an identity matrix.

[0045] C2: Establish an adversarial regularization loss function to minimize the consistency loss between perturbed samples and node embedding representations.

[0046] Furthermore, the adversarial regularization loss function is expressed as follows: , in, This indicates the loss of consistency due to perturbation. This is a non-linear feature mapping function set before the final classification layer. This represents the Euclidean distance used to measure feature differences. This represents the set of nodes in the constructed graph structure data. Represents a set of nodes Any node in the array.

[0047] Furthermore, by traversing and calculating each node in the graph structure data one by one, and by optimizing this objective, the model will learn a representation pattern that is insensitive to irrelevant perturbations but remains sensitive to real attack characteristics.

[0048] In step S5, the optimized node embedding representation is input into the classifier for threat behavior prediction and threat type discrimination, including the following steps D1-D3: D1: The optimized node embedding representation is input into the classifier, and the classifier outputs the predicted probability of each threat behavior category.

[0049] Furthermore, after graph attention network and robustness optimization, each threat event node... obtaining a set of high-dimensional vector representations: , It should be noted that the classifier adopts a multi-layer perceptron structure, which includes an input layer, a hidden layer and an output layer. The input layer feeds into the multi-layer perceptron. The hidden layer performs nonlinear transformation on the input features, which specifically involves at least one fully connected layer and a nonlinear activation function to obtain intermediate hidden feature representations. This process can be described as high-dimensional mapping and reorganization of the feature space, so that different threat categories are easier to separate in the feature space. Subsequently, linear projection is performed on the intermediate hidden feature representations to obtain the original score value corresponding to each type of threat. The score value reflects the relative confidence that the current sample belongs to each threat category. The output layer performs Softmax normalization processing on the original score value corresponding to each type of threat, and converts it into a predicted probability vector conforming to the nature of probability distribution, where each element represents the predicted probability that the threat event belongs to a certain category.

[0050] Further, the Softmax function is used to output the predicted probability of each threat behavior category, and the expression is as follows: , where represents the predicted probability of the threat behavior category, is the classification weight matrix, is the bias term, is the total number of threat categories.

[0051] D2: According to the predicted probability of each threat behavior category, it is judged whether the probability exceeds the threat behavior threshold, and a linkage response strategy is triggered.

[0052] It should be noted that is the predicted probability of threat behavior, which characterizes the probability that the current network threat event belongs to the threat category, is the threat category index. When the predicted probability of threat behavior of a certain type of threat event exceeds the threat behavior threshold, the system will automatically trigger the corresponding linkage response strategy. The linkage response strategy includes blocking connection, asset isolation or alarm reporting, etc. The threat behavior threshold is set according to experience.

[0053] D3: A cross-entropy loss function is established to optimize the classifier, so as to minimize the cross-entropy loss between the predicted label and the real label.

[0054] Further, output optimization is performed on the multi-layer perceptron by establishing a cross-entropy loss function. The optimization objective is to minimize the cross-entropy loss between the predicted label and the real label, and the expression is as follows: , where is the cross-entropy loss value, One-hot encoding for the real label. The predicted probability output by the model. Total number of threat categories Indexed by threat categories.

[0055] In summary, this invention overcomes the labeling dependency problem by collecting multi-source heterogeneous threat data and constructing positive and negative sample pairs using a contrastive learning mechanism, and employing a driving encoder to learn a unified threat feature representation. Furthermore, it extracts attack behavior patterns and constructs an attack behavior graph through pattern mining algorithms, and uses a graph attention network to semantically aggregate graph nodes, generating node embedding representations that integrate temporal evolution and topological structure. Secondly, it optimizes the robustness of the threat detection model by establishing a joint optimization loss function, and outputs the predicted probability of threat behavior through a multilayer perceptron, achieving efficient detection of network threats in dynamic and complex environments.

[0056] Example 3 is an embodiment of the present invention, which provides a multi-source heterogeneous network threat data fusion characterization system, including: a collection module, a preprocessing module, an update module, an optimization module, and a threat prediction module; The acquisition module converts target data into a meta-event sequence, which is standardized multi-source threat data. The preprocessing module responds to the acquisition of the meta-event sequence to obtain a threat feature vector. The threat feature vector is obtained by introducing a contrastive learning preprocessing mechanism. The steps of the contrastive learning preprocessing mechanism are to construct positive and negative sample pairs based on the meta-event sequence, input the positive and negative sample pairs into the encoder for training, and output the threat feature vector. The update module constructs an attack behavior graph based on threat feature vectors, establishes a node update model using a graph attention network, and outputs an updated node embedding representation as input to the attack behavior graph. The optimization module establishes an adversarial regularization loss function to optimize the updated node embedding representation; The threat prediction module embeds the optimized nodes into the input classifier to predict threat behavior and determine threat type.

[0057] This embodiment also provides an electronic device applicable to a multi-source heterogeneous network threat data fusion characterization method, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the multi-source heterogeneous network threat data fusion characterization method proposed in the above embodiment.

[0058] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements a multi-source heterogeneous network threat data fusion characterization method as proposed in the above embodiment.

[0059] The storage medium proposed in this embodiment belongs to the same inventive concept as the method for fusion characterization of multi-source heterogeneous network threat data proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0060] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0061] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for threat data fusion representation in multi-source heterogeneous networks, characterized in that: include, The target data is transformed into a meta-event sequence, which is standardized multi-source threat data; In response to the acquisition of the meta-event sequence, a threat feature vector is obtained, which is obtained by introducing a contrastive learning preprocessing mechanism; The steps of the contrastive learning preprocessing mechanism are as follows: construct positive and negative sample pairs based on the meta-event sequence, input the positive and negative sample pairs into the encoder for training, and output a threat feature vector; An attack behavior graph is constructed based on threat feature vectors, and a node update model is established using a graph attention network. The input is the attack behavior graph, and the output is the updated node embedding representation. An adversarial regularization loss function is established to optimize the updated node embedding representation; The optimized node embedding representation is input into the classifier for threat behavior prediction and threat type identification.

2. The method for fusing and characterizing multi-source heterogeneous network threat data as described in claim 1, characterized in that: In response to the acquisition of the meta-event sequence, a threat feature vector is obtained. This threat feature vector is obtained by introducing a contrastive learning preprocessing mechanism, including: For a threat event in the meta-event sequence, generate two enhanced samples as positive sample pairs, and the remaining threat events as negative sample pairs; The encoder is trained using positive and negative samples, and a contrastive loss function is introduced to optimize the training process, outputting a threat feature vector.

3. The multi-source heterogeneous network threat data fusion characterization method as described in claim 2, characterized in that: The process of constructing an attack behavior graph based on threat feature vectors, establishing a node update model using a graph attention network, and inputting the attack behavior graph to output updated node embedding representations includes: The threat feature vector is segmented into behavioral fragment subsequences, and the attack behavior patterns of the behavioral fragment subsequences are mined. Based on the discovered attack behavior patterns, construct an attack behavior map; A node update model is established using a graph attention network, which takes an attack behavior graph as input and outputs an updated node embedding representation.

4. The method for fusing and characterizing multi-source heterogeneous network threat data as described in claim 3, characterized in that: The establishment of an adversarial regularization loss function to optimize the updated node embedding representation includes: Add perturbations to the updated node embedding representations to construct perturbation samples; Establish an adversarial regularization loss function to minimize the consistency loss between perturbed samples and node embedding representations.

5. The method for fusing and characterizing multi-source heterogeneous network threat data as described in claim 4, characterized in that: The step of embedding the optimized nodes into the classifier for threat behavior prediction and threat type discrimination includes: The optimized node embedding representation is input into the classifier, and the classifier outputs the predicted probability of each threat behavior category; Based on the predicted probability of each threat behavior category, determine whether the threat behavior threshold is exceeded and trigger the linkage response strategy. A cross-entropy loss function is established to optimize the classifier, minimizing the cross-entropy loss between the predicted label and the true label.

6. The multi-source heterogeneous network threat data fusion characterization method as described in claim 5, characterized in that: The process of training the encoder with positive and negative samples, introducing a contrastive loss function to optimize the training process, and outputting a threat feature vector includes: The optimized training process is achieved by calculating the contrastive learning loss value between positive and negative samples; The contrastive learning loss value between positive and negative samples is obtained by calculating the loss value between positive sample pairs and the loss value between positive and negative sample pairs; The calculation of the loss value between positive sample pairs and the calculation of the loss value between positive and negative sample pairs require the introduction of a temperature parameter to adjust the smoothness of the similarity distribution.

7. The method for fusing and characterizing multi-source heterogeneous network threat data as described in claim 6, characterized in that: The step of establishing a node update model using a graph attention network, inputting an attack behavior graph, and outputting updated node embedding representations includes: A node update model is established using a graph attention network, expressed as follows: , in, For nodes The set of neighboring nodes, For nodes With nodes Attention weights between them For learnable parameter matrix, For activation function, This represents the index of the network layer in a graph neural network. Represents a node The neighboring nodes. Represents a node The updated node embedding representation in the next layer of the network.

8. A multi-source heterogeneous network threat data fusion characterization system, employing the multi-source heterogeneous network threat data fusion characterization method as described in any one of claims 1-7, characterized in that, include: The system includes a data acquisition module, a preprocessing module, an update module, an optimization module, and a threat prediction module. The acquisition module converts target data into a meta-event sequence, which is standardized multi-source threat data. The preprocessing module responds to the acquisition of the meta-event sequence to obtain a threat feature vector. The threat feature vector is obtained by introducing a contrastive learning preprocessing mechanism. The steps of the contrastive learning preprocessing mechanism are to construct positive and negative sample pairs based on the meta-event sequence, input the positive and negative sample pairs into the encoder for training, and output the threat feature vector. The update module constructs an attack behavior graph based on threat feature vectors, establishes a node update model using a graph attention network, and outputs an updated node embedding representation as input to the attack behavior graph. The optimization module establishes an adversarial regularization loss function to optimize the updated node embedding representation; The threat prediction module embeds the optimized nodes into the input classifier to predict threat behavior and determine threat type.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the multi-source heterogeneous network threat data fusion characterization method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the multi-source heterogeneous network threat data fusion characterization method according to any one of claims 1 to 7.