Network space surveying and mapping threat detection method and system based on causal association privacy protection

Abstract

The invention discloses a causal association privacy protection cyberspace surveying and mapping threat detection method and system, and relates to the technical field of information security and cyberspace surveying and mapping. The method specifically comprises the following steps: constructing a causal incidence matrix based on a multi-modal network space surveying and mapping data dynamic fusion method of multi-scale convolution and causal reasoning to realize cross-modal surveying and mapping data feature fusion; according to the dynamic Gaussian noise privacy protection method based on GCN importance evaluation, privacy budget is dynamically allocated by utilizing causal association importance, and fusion feature privacy is protected; according to the self-adaptive pre-training model gradient protection method based on the self-supervised dual disturbance mechanism, the privacy and availability of the model are dynamically protected; according to the network space surveying and mapping threat detection method based on the stacked LSTM model, abnormal modes in multi-scale traffic are learned, and high-precision threat detection is achieved. According to the invention, through causal association modeling and dynamic privacy protection, the threat detection capability in a complex network environment is improved while the data security is guaranteed.

Description

Technical Field

[0001] The present invention relates to the fields of information security and cyberspace mapping, and particularly to a threat detection method and system for cyberspace mapping with causal association privacy protection. Background Art

[0002] Currently, with the rapid development of cyberspace mapping and big data technologies, especially in the field of processing and fusion of multi-modal data, cyberspace mapping, as an important technical means, has been widely used in aspects such as network topology, traffic dynamic analysis, and threat detection. By collecting and analyzing data from different modalities, such as text, images, and numerical data, cyberspace mapping can effectively reveal the structural characteristics, traffic patterns, and potential security threats in the network. With the increase in data volume and the improvement of modal diversity, the accuracy and real-time performance of cyberspace mapping face huge challenges. Especially when processing multi-modal cyberspace mapping data, it is necessary to effectively extract the causal associations of each modality.

[0003] However, a large amount of multi-modal data is involved in the process of cyberspace mapping, including protocol fields, network topology diagrams, traffic dynamics, etc. These data not only contain rich network structure information but also imply the causal dependencies in the network. These causal associations may expose sensitive information of users and devices. For example, the network activities and traffic patterns of users often show strong causal associations, and attackers can infer the user's activity trajectories, access habits, and potential attack behaviors by analyzing the causal associations. Therefore, how to ensure the causal association privacy of multi-modal data while performing cyberspace mapping and threat detection has become a key problem to be solved urgently.

[0004] Existing privacy protection technologies mainly include methods such as differential privacy, homomorphic encryption, and data anonymization. However, for large-scale multi-modal data, traditional privacy protection methods face many challenges. Although the differential privacy method can effectively protect data privacy, its ability to protect the causal associations of multi-modal data is limited, and it often needs to add noise to the data, resulting in a significant decline in the usability and analysis accuracy of the data; although homomorphic encryption can perform calculations on encrypted data, its computational overhead is large, the response time is long, and it is limited to accessing encrypted data, restricting the flexible use of data; although the data anonymization method can, to a certain extent, conceal the user's identity, its privacy protection effect is not good when facing complex multi-modal cyberspace mapping data, and it cannot effectively prevent link attacks. Therefore, how to achieve high-precision network threat detection while ensuring the privacy of multi-modal cyberspace mapping data has become a major challenge in the current technological development. Summary of the Invention

[0005] In response to the above problems, the present invention provides a cyberspace mapping threat detection method and system with causal privacy protection, which aims to ensure the causal privacy of multimodal data while achieving high-precision identification of network security threats.

[0006] The specific scheme of the present invention is as follows: A causal privacy-preserving cyberspace mapping threat detection method comprises the following steps: S1, a dynamic fusion method of multimodal network space mapping data based on multi-scale convolution and causal reasoning, extracts features from multimodal mapping data, captures local details and global temporal dependencies, constructs a dynamic causal association matrix, and realizes cross-modal dynamic feature fusion; S2, a dynamic Gaussian noise privacy protection method based on GCN importance evaluation, combined with the K-means clustering method, reduces the dimension of fusion features, dynamically allocates privacy budgets based on causal correlation importance, and protects the privacy of fusion features; S3, an adaptive pre-trained model gradient protection method based on a self-supervised dual perturbation mechanism, adds perturbation samples and adversarial samples to the pre-trained model, and dynamically protects the model gradient privacy through a momentum mechanism; S4, a cyberspace mapping threat detection method based on a stacked LSTM model, uses a stacked structure to learn traffic features at different time scales and mine abnormal patterns in network traffic.

[0007] Furthermore, the S1 specifically includes: S11, obtaining text modality data, image modality data and numerical modality data, extracting features from text and image data through BERT and Transformer encoders, and converting numerical data into high-dimensional feature representation through an embedding layer; S12, uses multi-scale convolution to extract multi-scale features in numerical data, calculates the similarity matrix between modalities, and performs Softmax normalization on the similarity matrix to obtain the final attention weight; S13, introduce causal reasoning, calculate the causal effect of each modality on the final output, reduce the imbalance of causal effects through causal correction terms, and optimize the information fusion of different modalities; S14, based on the extracted multi-scale features, the final cross-modal fusion result is obtained through weighted summation of dynamic weighting and causal correction terms.

[0008] Feature extraction of multimodal data through BERT and Transformer encoders can give full play to the semantic representation advantages of pre-trained models in text, image, and numerical modalities, ensuring the high-dimensional semantic integrity of basic features; the combination of multi-scale convolution and dynamic attention mechanism can effectively capture the fine-grained associations and global dependencies between different modalities, enabling the model to adaptively focus on key modality interactions and enhance the pertinence of cross-modal information fusion; introducing causal reasoning to optimize modality relationship modeling, through causal effect calculation and correction term adjustment, can avoid redundancy or omission of modality information in traditional fusion methods, and significantly enhance the causal logic rationality of fused features.

[0009] Further, the S2 specifically includes: S21, the edge node receives the multimodal fusion vector and obtains the dimensionality-reduced multimodal fusion vector through the K-means clustering algorithm; S22, using the graph convolutional network to calculate the causal association importance of the dimensionality-reduced fusion vector, and dynamically allocating the privacy budget based on the causal association importance; S23, according to the allocated privacy budget, adding Gaussian noise to the fusion vector through the α-stable distribution mechanism, where the noise intensity is positively correlated with the privacy budget, to obtain the noisy feature vector.

[0010] Further, the obtaining of the dimensionality-reduced multimodal fusion vector through the K-means clustering algorithm specifically includes: Selecting K samples through the K-means clustering algorithm, calculating the Euclidean distance between each vector and the cluster center, and iteratively selecting the cluster center according to the probability distribution of the square of the distance until convergence, to obtain the dimensionality-reduced multimodal fusion vector.

[0011] Using K-means clustering to select representative samples can reduce the dimension of the fused features while retaining the core information, reducing the computational overhead and information loss in the privacy protection process; evaluating the causal association importance based on the graph convolutional network (GCN) can dynamically allocate the privacy budget according to the actual dependence degree between modalities, enabling high-importance features to obtain more strict privacy protection and low-importance features to achieve lightweight processing, and achieving an accurate balance between the privacy protection intensity and data availability; the dynamic addition strategy of Gaussian noise not only meets the security requirements under the differential privacy theoretical framework but also avoids the excessive damage to the feature semantics by the traditional fixed noise mechanism, ensuring that the feature vector after privacy protection can still effectively support subsequent detection tasks.

[0012] Further, the adding of perturbation samples and adversarial samples to the pre-trained model specifically includes: S31, adding perturbation samples to the self-supervised contrastive loss function and the triplet loss function, and combining the two loss functions to form a new loss function for the pre-trained model; S32. Add adversarial samples to the Mixup loss function to adversarially train the model.

[0013] Further, the dynamic protection of the model gradient privacy through the momentum mechanism specifically includes: S33. The cloud server sends the initial model parameters to the edge node; S34. The edge node calculates the local gradient based on the received initial model parameters and local data, and adds Gaussian noise to obtain the updated gradient; S35. Update the noise intensity through the momentum mechanism.

[0014] The combination of self-supervised contrastive learning and the triplet loss function strengthens the model's discriminative ability for sample semantic differences, enabling the pre-trained model to learn strongly discriminative feature representations even in scenarios without labeled data; the Mixup technique generates perturbed samples, enhancing the model's robustness through semantic mixing in the input space and effectively resisting the impact of adversarial attacks on detection accuracy; the dynamic noise adjustment strategy based on the momentum mechanism adaptively adjusts the gradient noise intensity according to the training process, protecting the model gradient privacy while avoiding interference from noise mutations to training stability, and ensuring that the model can still maintain efficient parameter optimization ability and generalization performance during the privacy protection process.

[0015] Further, the specific steps of S4 are as follows: S41. The cloud server performs average aggregation on the gradients uploaded by the edge nodes to obtain the global model parameters, and distributes the updated model parameters to the edge nodes for local model updates; S42. The edge node updates the stacked LSTM model based on the updated global model parameters, and learns features at different time scales through the stacked LSTM model; S43. Determine whether there is traffic anomaly by calculating the reconstruction error of each input.

[0016] The stacked LSTM structure extracts traffic features at different time scales layer by layer through multiple layers of networks. The bottom layer network captures short-period burst anomalies, and the top layer network mines long-term trend attack patterns, realizing multi-granularity dynamic modeling of network traffic; the anomaly detection mechanism based on the reconstruction error can accurately identify traffic behaviors that deviate from the normal mode, avoiding the problem of missed detection of complex attack patterns by traditional single-scale models; the collaborative training framework of the edge node and the cloud server protects local data privacy while using the update of global model parameters to improve the consistency and accuracy of edge detection, meeting the requirements of real-time threat monitoring in a distributed network environment.

[0017] A cyber space mapping threat detection system with causal association privacy protection, comprising: Data fusion module, a multi-modal network space mapping data dynamic fusion method based on multi-scale convolution and causal reasoning, extracts features from multi-modal mapping data, captures local details and global temporal dependencies, constructs a dynamic causal association matrix, and realizes cross-modal dynamic feature fusion; Privacy protection module, a dynamic Gaussian noise privacy protection method based on GCN importance evaluation, combines the K-means clustering method, reduces the dimension of the fusion features, and dynamically allocates privacy budgets using causal association importance to protect the privacy of the fusion features; Gradient protection module, an adaptive pre-training model gradient protection method based on a self-supervised dual perturbation mechanism, adds perturbation samples and adversarial samples to the pre-training model, and dynamically protects the model gradient privacy through a momentum mechanism; Threat detection module, a network space mapping threat detection method based on a stacked LSTM model, uses a stacked structure to learn traffic features at different time scales and mines abnormal patterns in network traffic.

[0018] Through causal association modeling and multi-scale feature learning, the data fusion module breaks through the limitations of modal isolation of traditional fusion methods, provides high-value density fusion features for the system; the privacy protection module combines clustering dimensionality reduction and dynamic privacy budget allocation to achieve hierarchical privacy protection while performing lightweight processing, meeting the sensitive information protection requirements of network space mapping data; the gradient protection module constructs a full-process security protection system for the pre-training model through dual perturbation and momentum noise mechanisms to resist attack risks such as model inversion and gradient leakage; the stacked LSTM structure of the threat detection module realizes in-depth mining of the spatio-temporal features of network traffic. Through this system, the accuracy, robustness, and privacy compliance of network space threat detection have been significantly improved.

[0019] Compared with the prior art, the beneficial effects of the present invention are: through the dynamic fusion mechanism of multi-scale convolution and causal reasoning, it breaks through the limitations of traditional modal independent processing, captures local details and global temporal dependencies of text, image, and numerical modalities, and realizes logical enhanced fusion of cross-modal information through a causal association matrix, providing high-value density feature inputs for threat detection; the dynamic privacy protection method based on GCN and K-means realizes intelligent allocation of privacy budgets through causal association importance evaluation, ensures hierarchical protection of sensitive information while reducing feature dimensions, and balances data availability and security; the self-supervised dual perturbation and momentum gradient protection technology effectively resists model gradient leakage and adversarial attacks, and improves the robustness of the pre-training model; the multi-time scale feature learning ability of the stacked LSTM model accurately mines short-term anomalies and long-term attack patterns in network traffic, significantly improving the accuracy and generalization ability of threat detection. The present invention realizes high-precision identification and real-time response to network security threats while ensuring data privacy compliance. Brief Description of the Drawings

[0020] To more clearly illustrate the technical solutions in the embodiments of the present drawings or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present drawings. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on the structures shown in these drawings.

[0021] Figure 1 It is the flowchart of the method of the present invention; Figure 2 It is a comparison chart of the accuracy between the embodiments of the present invention and the traditional cyber space mapping threat detection method under different total privacy budgets; Figure 3 It is a comparison chart of the accuracy between the embodiments of the present invention and the traditional cyber space mapping threat detection method under different causal association strengths; Figure 4 It is a comparison chart of the accuracy between the embodiments of the present invention and the traditional cyber space mapping threat detection method under different perturbation intensities; Figure 5 It is a comparison chart of the accuracy between the embodiments of the present invention and the traditional cyber space mapping threat detection method under different time windows. Detailed implementation manners

[0022] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the following describes and explains the present invention in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. Based on the embodiments provided by the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0023] To better understand the solutions of the embodiments of the present invention, the following first introduces some related terms and concepts that may be involved in the embodiments of the present invention.

[0024] (1) Graph Convolutional Network (GCN) is a type of neural network model specifically designed for processing graph data. It combines the topological structure of the graph with node features and extracts the relationships and feature information between nodes through graph convolution operations.

[0025] (2) Long Short-Term Memory Artificial Neural Network (LSTM) is a type of recurrent neural network designed to solve the long-term dependence problem existing in general Recurrent Neural Networks (RNNs). All RNNs have a chain form of repeating neural network modules. In a standard RNN, this repeating structural module has a very simple structure, such as a tanh layer.

[0026] (3) K-means clustering is an iterative clustering analysis algorithm that is mainly used to divide a data set into K clusters, where the center point of each cluster is the mean of all the points in the cluster.

[0027] (4) Mixup technology: By fusing samples and labels in the same way, a new training sample is obtained. The general data enhancement method only changes the sample without changing the label, while Mixup changes both the sample and the label.

[0028] (5) Self-supervised contrastive learning is an unsupervised learning method that learns discriminative feature representations by constructing contrast relationships between samples without relying on labeled data.

[0029] (6) Differential Privacy is a cryptographic technique that aims to maximize the accuracy of data queries when querying from a statistical database while minimizing the chance of identifying its records.

[0030] In this embodiment, Figure 1 As shown, the present invention provides a cyberspace mapping threat detection method with causal association privacy protection, which specifically includes the following steps: S1, a dynamic fusion method of multimodal network space mapping data based on multi-scale convolution and causal reasoning, extracts features from multimodal mapping data, captures local details and global temporal dependencies, constructs a dynamic causal association matrix, and realizes cross-modal dynamic feature fusion; S2, a dynamic Gaussian noise privacy protection method based on GCN importance evaluation, combined with the K-means clustering method, reduces the dimension of fusion features, dynamically allocates privacy budgets based on causal correlation importance, and protects the privacy of fusion features; S3, an adaptive pre-trained model gradient protection method based on a self-supervised dual perturbation mechanism, adds perturbation samples and adversarial samples to the pre-trained model, and dynamically protects the model gradient privacy through a momentum mechanism; S4, a cyberspace mapping threat detection method based on a stacked LSTM model, uses a stacked structure to learn traffic features at different time scales and mine abnormal patterns in network traffic.

[0031] Furthermore, the multimodal network space mapping data dynamic fusion method based on multi-scale convolution and causal reasoning specifically encodes the multimodal data through BERT and Transformer encoders, uses multi-scale convolution to obtain attention weights, and constructs multimodal feature causal relationships based on causal relationship mining, and finally dynamically weighted fusion through multi-scale convolution.

[0032] Furthermore, S1 specifically includes the following steps: S11. Obtain text modal data, image modal data, and numerical modal data. Extract features from the text and image data through the BERT and Transformer encoders, and convert the numerical data into a high-dimensional feature representation through the embedding layer.

[0033] Specifically, use two single-modal encoders: a text-image encoder and a numerical encoder to encode the text, image, and numerical values respectively. The formulas are as follows: ; Where, is the embedding of the key ; is the embedding of the value , generated by the BERT encoder; Concat represents the concatenation operation; represents the text information conversion operation performed by the context encoder, and the context encoder is specifically the Transformer encoder; is the generated text feature.

[0034] ; Where I represents the image data; represents the convolution operation on the image data I to capture local details; represents the fusion operation performed by the integration encoder; v is the generated image feature.

[0035] For the input numerical data at each time step t, first convert it into a high-dimensional feature representation through the embedding layer, is the input numerical sequence, and the feature extraction formula is: ; Where, is the embedding that maps the numerical data to the high-dimensional space; is the numerical embedding feature at this time step; T is the time step length.

[0036] S12. Use multi-scale convolution to extract multi-scale features from the numerical data, calculate the similarity matrix between modalities, and perform Softmax normalization on the similarity matrix to obtain the final attention weights.

[0037] Specifically, the multi-scale convolution method is used to extract different-scale features from numerical data, and the formula is as follows: ; Among them, represents feature extraction using convolution kernels of different sizes; k is the scale of the convolutional layer; is the multi-scale feature in the numerical data extracted through the convolutional layer, including local features and global features.

[0038] The dot product attention method is used to adapt to the multi-modal fusion of network space mapping. When calculating the similarity between text (q), numerical (z), and image (v), dynamic weights, adaptive causal correction, and temporal information are added, and the formula is as follows: ; Among them, : The query matrix (Query) at time step t, obtained by linear transformation, where is the modal data of modality m at time step t, is the query weight matrix; : The key matrix (Key) at time step t, obtained by linear transformation, where is the modal data of modality at time step t, is the key weight matrix; is the transposed vector of ; d is the feature dimension, used for scaling to avoid gradient explosion; is the similarity between modality m and modality at time step t. The larger the value, the stronger the correlation between the two modalities; represents the introduction of historical information and considers the temporal evolution of modalities; represents a key parameter for capturing the dynamic relationship between different modalities during the multi-modal data fusion process, which is related to causal association; is the modality adaptive weight, which dynamically adjusts the importance of different modalities; is the causal intervention correction term, which removes non-causal correlations, such as noise and irrelevant features; refers to the control and intervention of the variable m' under the causal inference framework.

[0039] The similarity between the three modalities is calculated using the following formula: ; ; .

[0040] The similarity matrix is obtained by calculating the similarity between modalities, and the Softmax normalization is performed on the similarity matrix to obtain the final attention weights: ; where, is the attention weight of the m pairs of modalities .

[0041] S13, introducing causal reasoning, calculates the causal effect of each modality on the final output, and reduces the causal effect imbalance through the causal correction term to optimize the information fusion of different modalities.

[0042] Specifically, introducing causal reasoning, optimizing the relationship modeling between modalities, and dealing with the variable and complex environment in cyber space mapping. The specific causal effect formula is as follows: ; ; ; ; where, is the causal effect of the text modality attention on the final output O; is the causal effect of the image modality attention on the final output O; is the causal effect of the numerical modality attention on the final output O; is the causal effect of integrating the three modalities; represents the expected value of the text modality attention ; represents the expected value of the image modality attention ; represents the expected value of the numerical modality attention ; represents multiple modality attentions ( , , The combined joint expectation, which is used to measure the comprehensive impact of multimodal collaboration on the final output result O, reflects the overall causal effect of the attention of each modality under multi-factor fusion, rather than a simple summation of independent effects; is the conditional probability, representing the probability of the final output O given the attention of the text modality , the text feature q, and its related probability parameters ; is the conditional probability, representing the probability of the final output O given the attention of the image modality , the image feature v, and its related probability parameters ; is the conditional probability, representing the probability of the final output O given the attention of the numerical modality , the numerical feature z, and its related probability parameters ; is the conditional probability of the multimodality, representing the probability of the final output O given all the modality attentions ( , , ) and all the modality features (q, v, z); represents an intervention on , forcing the value of to be ; represents an intervention on , forcing the value of to be ; represents an intervention on , forcing the value of to be ; is the conditional probability of the final output O given q and after an intervention on ; is the conditional probability of the final output O given v and after an intervention on ; is the conditional probability of the final output O given z and after an intervention on ; After intervening in multiple modalities, given all modal attentions ( , , ), as well as all modal features (q, v, z), the conditional probability of the final output O.

[0043] Design a method for dynamically correcting causal attention to reduce causal effect imbalance, thereby optimizing the information fusion of different modalities. Introduce a causal correction term , and dynamically adjust the relationship between modalities according to the results of counterfactual reasoning: ; Among them, represents the causal influence between the i-th node and the j-th node; represents the feature vector of node i; the transposed feature vector of node j; is the dimension of the k-th feature.

[0044] The causal correction term is dynamically calculated through the following formula: ; represents a hyperparameter in a certain causal correction process, which adjusts the correction intensity.

[0045] S14, based on the extracted multi-scale features, through dynamic weighted sum and weighted sum of causal correction terms, obtains the final cross-modal fusion result. The specific formula is as follows: ; ; Among them, is the dynamic weight of each scale feature; is the final fusion result; is the weight corresponding to each modal feature, indicating the influence of each feature on the final result; represents the fused features or modalities, which is the intermediate result of the fusion of multiple features.

[0046] Furthermore, S2 is specifically to reduce the dimension of the fusion vector according to K-means clustering sampling, dynamically allocate privacy budgets based on causal association importance measurement, use a graph convolutional network (GCN) to calculate the causal importance of the fusion vector, and control the noise intensity according to the α-stable distribution noise, including the following steps: S21. The edge node receives the multi-modal fusion vector and obtains the dimension-reduced multi-modal fusion vector through the K-means clustering algorithm.

[0047] Specifically, the edge node selects N centers through the K-means clustering algorithm, that is, samples N vectors, improves the diversity and representativeness of the sample set, and achieves a better privacy-utility balance in the multi-modal privacy protection scenario.

[0048] Randomly select a vector from the fused vectors as the first clustering center, denoted as c1.

[0049] For each vector , calculate its distance from the currently selected clustering . The formula for calculating the distance is: ; where represents the Euclidean distance between vectors; represents the j-th clustering center.

[0050] Calculate the probability of selecting the next center according to the square of the distance from each point to the nearest center. Specifically, for each vector , the probability that it is selected as the next center is given by the following formula: ; According to the calculated probability distribution, select the next clustering center from with probability, denoted as , and repeat until convergence. Select N clustering centers, and these N centers are the sampled vectors.

[0051] Convergence determination condition: The change in the clustering center between the t-th round and the (t - 1)-th round is less than the set threshold r: ; where is the position of the i-th clustering center in the t-th round.

[0052] S22. Use the graph convolutional network to calculate the causal association importance of the dimension-reduced fusion vector, and dynamically allocate the privacy budget based on the causal association importance. The specific formula is as follows: ; where , respectively represent the preset minimum and maximum privacy budgets, usually ; is to control the privacy budget with Rate of change ( ); represents the causal association importance of node j to node i; represents the privacy budget allocated to the causal association between node i and node j, controlling the noise protection intensity of this association feature.

[0053] In the graph convolutional network GCN, the causal association importance can be dynamically measured through the attention coefficient, and the specific formula is as follows: ; where W is a learnable weight matrix; is the attention parameter vector, is the transposed vector of; , and respectively represent the features of node i, node j, and node k; represents vector concatenation; LeakyReLU is the leaky rectified linear unit function; is the set of all nodes connected to node i.

[0054] S23, according to the allocated privacy budget, adds Gaussian noise to the fused vector through the α-stable distribution mechanism. The noise intensity is positively correlated with the privacy budget to obtain the noise-added feature vector. The specific process is as follows: The fused feature vector after dimensionality reduction , add noise to it, and the process of adding noise is as follows: ; where, is the noise-added feature vector; is the intensity of the noise, controlling the amplitude of the noise; is the noise vector from the α-stable distribution.

[0055] The noise vector each component of (for ) comes from the α-stable distribution, and the characteristic function of the α-stable distribution is: ; where, controls the tail of the distribution; is the Gaussian distribution when; It is a heavy-tailed distribution with infinite variance; u is the value of a certain component of the noise vector and is one of the variables used to calculate the expectation; i is an index indicating the component in the noise vector .

[0056] Under the dynamic privacy budget allocation mechanism, the noise intensity is directly related to the causal association importance of the vector; and the privacy budget will be dynamically adjusted according to the real-time changes of modalities such as text, images, and numerical values.

[0057] Specifically, it is calculated according to the following formula: ; where represents the total privacy budget allocated to the causal association between node k and other nodes.

[0058] Furthermore, S3 is specifically to improve the model robustness through the pre-training of the model with double perturbations, and to protect the privacy of model parameters while improving the model performance by adding noise to the dynamic gradient, including the following steps: S31, adding perturbed samples to the self-supervised contrast loss function and the triplet loss function, and combining the two loss functions to form a new loss function to pre-train the model.

[0059] Specifically, based on the data after sampling and adding noise, self-supervised double perturbations are used to enhance the model robustness by generating adversarial samples and perturbing the model itself, enhance the accuracy of traffic anomaly recognition, and reduce the negative impact of noise on pre-training.

[0060] The first stage (pre-training stage): Self-supervised learning: ; where m is the sample vector after adding noise, that is, the feature vector after adding noise in S23; is the feature representation obtained by the sample vector m after adding noise through the encoder; is the positive sample related to m; is the sample not similar to the sample m, and the negative sample is taken here; sim is the similarity function; is the self-supervised contrast loss function.

[0061] Triplet loss function: ; Among them, is the first hyperparameter, representing the minimum interval between positive and negative samples; is the triplet loss function; is the negative sample.

[0062] Combine the triplet loss with the original contrastive loss to form a new loss function: ; Among them, is the second hyperparameter, used to control the weight of the triplet loss in the overall loss.

[0063] S32. Add adversarial samples to the Mixup loss function to adversarially train the model.

[0064] Specifically, the second stage (adversarial training stage): Use the Mixup technique to perturb the model. The formula is: ; Among them, is the input mixed sample; is a model sample under attack, serving as the target sample for mixing, obtained by attacking the model; is the prediction of the model for the input sample ; is the weight randomly sampled from the range [0, 1].

[0065] Input mixing: Perform weighted averaging on the input samples and to obtain the mixed sample: ; Among them, is the third hyperparameter sampled from the uniform distribution U(0, 1).

[0066] is 's true label; is the adversarial sample.

[0067] S33. The cloud server sends the initial model parameters to the edge node; the edge node calculates the local gradient based on the received initial model parameters and local data, and adds Gaussian noise to obtain the updated gradient; update the noise intensity through the momentum mechanism.

[0068] After receiving the initial model parameters, the edge node calculates the gradient according to the local model. When calculating, the edge node adds noise to the gradient to ensure the privacy of hidden data.

[0069] Specifically, the gradient is calculated by the following formula: ; where are the local model parameters; are the parameters sent by the cloud server; is the gradient operator; is the loss function.

[0070] The noise intensity is added to the gradient, and the updated gradient is: ; where represents the added Gaussian noise.

[0071] The update of the noise uses the momentum mechanism to avoid too drastic changes in the noise. The specific formula is: ; where is the noise intensity in the t-th round; is the change amount of the noise, which is calculated based on the change of the loss function; ρ is the attenuation rate, which is used to adjust the smoothness of the noise intensity update process.

[0072] Furthermore, S4 is specifically that according to the uploaded gradient, the cloud server performs average aggregation, the model parameters are sent to the edge node for local model update, and the edge node processes the data by stacking multiple LSTM layers, learns features at different time scales at different levels, and performs threat detection according to the error. Specifically, it includes the following steps: S41, the cloud server performs average aggregation according to the gradients uploaded by the edge nodes, obtains the global model parameters, and sends the updated model parameters to the edge nodes for local model update. The formula is as follows: ; where M is the number of edge nodes; are the global model parameters; are the model parameters of the i-th client; is the local data volume of the i-th client; is the total data volume.

[0073] S42. The edge node updates the stacked LSTM model based on the updated global model parameters, and learns features at different time scales through the stacked LSTM model.

[0074] Specifically, according to the model parameters sent by the cloud server, the stacked LSTM threat detection model is updated. Through the multi-level learning mechanism, the stacked LSTM model can accurately capture the long-term and short-term dependencies in network traffic, extract effective features from the noisy data, and improve the accuracy and robustness of threat detection.

[0075] The stacked LSTM architecture includes: Forget gate: ; Among them, is the sigmoid activation function; is the weight matrix of the forget gate; is the hidden state at the previous moment, containing historical information of past traffic patterns; is the input numerical data at each time step t; represents the vector formed by concatenating the hidden state at the previous moment and the input at the current moment. By fusing the historical traffic pattern and the current real-time data, the LSTM can capture the temporal dependencies in the traffic, thereby identifying abnormal patterns; is the bias term of the forget gate; is the output of the forget gate.

[0076] Input gate: ; Among them, is the weight matrix of the input gate; is the bias term of the input gate; is the output of the input gate.

[0077] Candidate memory cell: ; Among them, is the weight matrix of the candidate memory cell; is the candidate cell state; is the hyperbolic tangent activation function; is the bias term of the candidate memory cell.

[0078] Cell state update: ; Among them, represents the cell state at the current moment.

[0079] Output gate: ; Among them, is the weight matrix of the output gate; is the bias term of the output gate.

[0080] Hidden state update: .

[0081] SkLSTM processes data by stacking multiple LSTM layers, and uses the hidden state output by the first LSTM layer as the input of the second LSTM layer: .

[0082] Through such a stacking structure, SkLSTM can learn features at different time scales at different levels.

[0083] S43, determines whether there is abnormal traffic by calculating the reconstruction error of each input.

[0084] By calculating the traffic data of each input of the reconstruction error to determine whether there is abnormal traffic, the specific formula is as follows: ; Among them, is the reconstructed input, that is, the reconstructed output generated by the stacked LSTM model after processing the original input data.

[0085] If the reconstruction error exceeds the set threshold , it is determined as a threat, and the specific formula is as follows: .

[0086] The present invention also provides a cyberspace mapping threat detection system for causal association privacy protection, specifically including the following modules: Data fusion module, based on a multi-modal network space mapping data dynamic fusion method of multi-scale convolution and causal reasoning, extracts features from multi-modal mapping data, captures local details and global temporal dependencies, constructs a dynamic causal association matrix, and realizes cross-modal dynamic feature fusion.

[0087] Specifically, the data fusion module uses multi-scale convolution to calculate the similarity of the three modalities, performs Softmax normalization on these similarity matrices to obtain the final attention weights, and uses a structural causal model to quantify the causal impact of different modalities on the final mapping results in order to capture the deep relationship between the modalities. Dynamic attention causal correction is used to optimize the information fusion of different modalities.

[0088] The privacy protection module uses a dynamic Gaussian noise privacy protection method based on GCN importance evaluation and K-means clustering method to reduce the dimension of fusion features, dynamically allocate privacy budgets based on causal correlation importance, and protect the privacy of fusion features.

[0089] Specifically, the privacy protection module reduces the dimension of the fusion vector through K-means clustering sampling, combines α-stable distribution noise with dynamic privacy budget allocation, and protects the causal correlation privacy of the fusion features.

[0090] The gradient protection module is an adaptive pre-trained model gradient protection method based on a self-supervised dual perturbation mechanism. It adds perturbation samples and adversarial samples to the pre-trained model and dynamically protects the model gradient privacy through a momentum mechanism.

[0091] Specifically, the gradient protection module adds dynamic attenuation noise to the edge node gradients during the federated learning process, and combines gradient correlation clipping with weighted aggregation to optimize the global model update.

[0092] The threat detection module, a cyberspace mapping threat detection method based on the stacked LSTM model, adopts a stacked structure to learn traffic characteristics at different time scales and mine abnormal patterns in network traffic.

[0093] Specifically, the threat detection module is based on the network space mapping anomaly detection module of the stacked LSTM model. It processes data by stacking multiple LSTM layers, learns features of different time scales at different levels, and performs threat detection based on errors. It mines abnormal patterns in network traffic to achieve highly robust threat detection.

[0094] In order to ensure the privacy of multimodal cyberspace mapping, it is necessary to prove that the cyberspace mapping threat detection method for causal association privacy protection in this embodiment satisfies differential privacy protection.

[0095] After multi-modal data is fused and dimensionally reduced, a special type of noise (α-stable distribution noise) is added to these features. The intensity of this noise is initially set to a baseline value, and each feature is allocated a different privacy budget according to its sensitivity. For example, key features with strong causal associations will receive a higher privacy budget, and thus stronger noise will be added to mask sensitive information. Since this method of adding noise conforms to the mathematical definition of differential privacy, this approach can ensure the privacy security of the fused features, preventing attackers from inferring real sensitive information from the data.

[0096] During the model training process, the gradients of model updates also contain sensitive information. Therefore, another type of noise (Gaussian noise) needs to be added to the gradients, and the noise intensity is controlled by the privacy budget. The higher the budget, the stronger the added noise, and the better the sensitive information in the gradients is masked. This approach also meets the requirements of differential privacy, which can not only prevent attackers from reverse engineering the original data through the gradients but also ensure that the model can normally learn network threat patterns.

[0097] The above two privacy protection measures (adding noise to features and adding noise to gradients) are implemented independently. According to the composition principle of differential privacy, when multiple privacy protection methods are used in parallel, the overall privacy protection strength can be measured by combining their respective privacy budgets. Therefore, the overall solution of the present invention forms the final total privacy budget by adding the privacy budgets of feature protection and gradient protection, mathematically proving that the entire threat detection method meets the differential privacy standard and achieving a high level of privacy protection for multi-modal network data.

[0098] In this embodiment, based on the real network threat dataset RT-IoT, different parameters are used: total privacy budget, causal association strength, perturbation strength, and time window to evaluate the threat detection accuracy of the present invention for the privacy protection of cyberspace mapping data. The results of the comparative experiments are as Figures 2 - 5 shown. From Figures 2 - 5 it can be seen that the cyberspace mapping threat detection effects of this embodiment under different parameters are all better than traditional methods.

[0099] It should be noted that the present invention is not limited to the above embodiments. The above embodiments are only examples, and embodiments with the same composition and the same effect as the technical idea within the technical scope of the present invention are all included in the technical scope of the present invention. In addition, within the scope of not departing from the gist of the present invention, various modifications that can be thought of by those skilled in the art to the embodiments and other ways constructed by combining some constituent elements of the embodiments are also included in the scope of the present invention.

Claims

1. A method for detecting threats in cyberspace mapping for causal association privacy protection, characterized in that, The following steps are involved: S1, a dynamic fusion method of multimodal network space mapping data based on multi-scale convolution and causal reasoning, extracts features from multimodal mapping data, captures local details and global temporal dependencies, constructs a dynamic causal association matrix, and realizes cross-modal dynamic feature fusion; S2, a dynamic Gaussian noise privacy protection method based on GCN importance evaluation, combined with the K-means clustering method, reduces the dimension of fusion features, dynamically allocates privacy budgets based on causal correlation importance, and protects the privacy of fusion features; S3, an adaptive pre-trained model gradient protection method based on a self-supervised dual perturbation mechanism, adds perturbation samples and adversarial samples to the pre-trained model, and dynamically protects the model gradient privacy through a momentum mechanism; S4, a cyberspace mapping threat detection method based on a stacked LSTM model, uses a stacked structure to learn traffic features at different time scales and mine abnormal patterns in network traffic.

2. The method for detecting threats in cyber space mapping for causal association privacy protection according to claim 1, wherein, The S1 specifically includes: S11, obtaining text modality data, image modality data and numerical modality data, extracting features from text and image data through BERT and Transformer encoders, and converting numerical data into high-dimensional feature representation through an embedding layer; S12, uses multi-scale convolution to extract multi-scale features in numerical data, calculates the similarity matrix between modalities, and performs Softmax normalization on the similarity matrix to obtain the final attention weight; S13, introduce causal reasoning, calculate the causal effect of each modality on the final output, reduce the imbalance of causal effects through causal correction terms, and optimize the information fusion of different modalities; S14, based on the extracted multi-scale features, the final cross-modal fusion result is obtained through weighted summation of dynamic weighting and causal correction terms.

3. The method for detecting threats in cyberspace mapping for causal association privacy protection according to claim 1, wherein The S2 specifically includes: S21, the edge node receives the multimodal fusion vector and obtains the multimodal fusion vector after dimensionality reduction through the K-means clustering algorithm; S22, uses graph convolutional networks to calculate the causal importance of the fusion vector after dimension reduction, and dynamically allocates the privacy budget based on the causal importance; S23, according to the allocated privacy budget, Gaussian noise is added to the fusion vector through the α-stable distribution mechanism. The noise intensity is positively correlated with the privacy budget, and the noisy feature vector is obtained.

4. The method for detecting threats in cyberspace mapping for causal association privacy protection according to claim 3, wherein, The multimodal fusion vector after dimensionality reduction is obtained by using the K-means clustering algorithm, specifically including: K samples are selected through the K-means clustering algorithm, the Euclidean distance between each vector and the cluster center is calculated, and the cluster center is iteratively selected according to the probability distribution of the square of the distance until convergence, and the multimodal fusion vector after dimensionality reduction is obtained.

5. The method for detecting threats in cyberspace mapping for causal association privacy protection according to claim 1, wherein Adding perturbation samples and adversarial samples to the pre-trained model specifically includes: S31, adding perturbation samples to the self-supervised contrast loss function and the triplet loss function, and combining the two loss functions to form a new loss function to pre-train the model; S32, add adversarial samples to the Mixup loss function to adversarially train the model.

6. The method for detecting threats in cyber space mapping with causal association privacy protection according to claim 1, wherein The dynamic protection of model gradient privacy through the momentum mechanism specifically includes: S33, the cloud server sends initial model parameters to the edge node; S34. The edge node calculates the local gradient based on the received initial model parameters and local data, and adds Gaussian noise to obtain the updated gradient. S35. Update the noise intensity through the momentum mechanism.

7. The method for detecting threats in cyber space mapping for causal association privacy protection according to claim 1, characterized in that, The specific steps of S4 are as follows: S41. The cloud server performs average aggregation on the gradients uploaded by the edge nodes to obtain the global model parameters, and distributes the updated model parameters to the edge nodes for local model update. S42. The edge node updates the stacked LSTM model based on the updated global model parameters, and learns features at different time scales through the stacked LSTM model. S43. Determine whether there is traffic anomaly by calculating the reconstruction error of each input.

8. A cyber space mapping threat detection system for causal association privacy protection, characterized in that, It includes: A data fusion module, which uses a multi-modal network spatial mapping data dynamic fusion method based on multi-scale convolution and causal inference to extract features from multi-modal mapping data, capture local details and global temporal dependencies, construct a dynamic causal association matrix, and achieve cross-modal dynamic feature fusion. A privacy protection module, which uses a dynamic Gaussian noise privacy protection method based on GCN importance evaluation, combines the K-means clustering method to reduce the dimension of the fused features, and dynamically allocates privacy budgets using causal association importance to protect the privacy of the fused features. A gradient protection module, which uses an adaptive pre-training model gradient protection method based on a self-supervised dual perturbation mechanism to add perturbation samples and adversarial samples to the pre-training model, and dynamically protects the model gradient privacy through the momentum mechanism. A threat detection module, which uses a network spatial mapping threat detection method based on a stacked LSTM model, adopts a stacked structure to learn traffic features at different time scales, and mines abnormal patterns in network traffic.

Images