Cyberspace mapping threat detection method and system with causal association privacy protection

Abstract

The present invention discloses a cyberspace mapping threat detection method and system with causal privacy protection, which relates to the field of information security and cyberspace mapping technology. Specifically, it includes: a dynamic fusion method of multimodal cyberspace mapping data based on multi-scale convolution and causal reasoning, which constructs a causal matrix to realize cross-modal mapping data feature fusion; a dynamic Gaussian noise privacy protection method based on GCN importance evaluation, which dynamically allocates privacy budgets based on causal importance to protect the privacy of fusion features; an adaptive pre-training model gradient protection method based on a self-supervised dual perturbation mechanism, which dynamically protects model privacy and availability; a cyberspace mapping threat detection method based on a stacked LSTM model, which learns abnormal patterns in multi-scale traffic to achieve high-precision threat detection. The present invention improves threat detection capabilities in complex network environments while ensuring data security through causal modeling and dynamic privacy protection.

Description

Technical Field

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

[0002] With the rapid development of cyberspace mapping and big data technologies, particularly in the processing and fusion of multimodal data, cyberspace mapping, as an important technical tool, has been widely applied in network topology, traffic dynamics analysis, and threat detection. By collecting and analyzing data from different modalities, such as text, images, and numerical data, cyberspace mapping can effectively reveal structural characteristics, traffic patterns, and potential security threats within a network. With the increase in data volume and modal diversity, the accuracy and real-time performance of cyberspace mapping face significant challenges. This is particularly true when processing multimodal cyberspace mapping data, which requires effectively extracting causal relationships between modalities.

[0003] However, the process of cyberspace mapping involves a large amount of multimodal data, including protocol fields, network topology diagrams, and traffic dynamics. This data not only contains rich network structural information but also contains causal dependencies within the network. These causal dependencies can expose sensitive information about users and devices. For example, a user's network activity and traffic patterns often exhibit strong causal dependencies, allowing attackers to infer user activity trajectories, access habits, and potential attack behaviors by analyzing these causal dependencies. Therefore, ensuring the privacy of causal dependencies in multimodal data while performing cyberspace mapping and threat detection has become a key issue that needs to be addressed.

[0004] Existing privacy protection technologies primarily include differential privacy, homomorphic encryption, and data anonymization. However, traditional privacy protection methods face numerous challenges when dealing with large-scale multimodal data. While differential privacy methods can effectively protect data privacy, their ability to protect the causal relationships of multimodal data is limited, and they often require the addition of noise to the data, significantly reducing the data's practicality and analytical accuracy. While homomorphic encryption can perform computations on encrypted data, it suffers from high computational overhead, long response times, and is limited to accessing encrypted data, limiting its flexible use. While data anonymization methods can conceal user identities to a certain extent, they offer limited privacy protection when dealing with complex multimodal cyberspace mapping data, failing to effectively prevent linking attacks. Therefore, ensuring the privacy of multimodal cyberspace mapping data while achieving high-precision network threat detection has become a major challenge in 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 correlation privacy protection, aiming to ensure the causal correlation privacy of multimodal data while achieving high-precision identification of network security threats.

[0006] The specific scheme of the present invention is as follows:

[0007] A causal association privacy-preserving cyberspace mapping threat detection method comprises the following steps:

[0008] S1, a dynamic fusion method for multimodal network spatial 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;

[0009] 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;

[0010] S3, an adaptive pre-trained model gradient protection method based on a self-supervised dual perturbation mechanism, adds perturbed samples and adversarial samples to the pre-trained model and dynamically protects the model gradient privacy through a momentum mechanism;

[0011] S4, a cyberspace mapping threat detection method based on a stacked LSTM model, uses a stacked structure to learn traffic characteristics at different time scales and mine abnormal patterns in network traffic.

[0012] Furthermore, the S1 specifically includes:

[0013] S11, obtain text modal data, image modal data, and numerical modal data, perform feature extraction on the text and image data through BERT and Transformer encoders, and convert the numerical data into high-dimensional feature representation through the embedding layer;

[0014] S12 uses multi-scale convolution to extract multi-scale features from numerical data, calculates the similarity matrix between modalities, and performs Softmax normalization on the similarity matrix to obtain the final attention weight;

[0015] S13, introduces causal reasoning, calculates the causal effect of each modality on the final output, reduces the imbalance of causal effects through causal correction terms, and optimizes the information fusion of different modalities;

[0016] S14, based on the extracted multi-scale features, the final cross-modal fusion result is obtained through dynamic weighting and weighted summation of causal correction terms.

[0017] By extracting features from multimodal data using BERT and Transformer encoders, we can fully leverage 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 fine-grained associations and global dependencies between different modalities, enabling the model to adaptively focus on key modal interactions and improve the targeted nature of cross-modal information fusion. Causal reasoning is introduced to optimize modal relationship modeling. By calculating causal effects and adjusting correction terms, the redundancy or omission of modal information in traditional fusion methods can be avoided, significantly enhancing the causal logical rationality of the fusion features.

[0018] Furthermore, the S2 specifically includes:

[0019] S21, the edge node receives the multimodal fusion vector and obtains the multimodal fusion vector after dimensionality reduction through the K-means clustering algorithm;

[0020] S22 uses a graph convolutional network to calculate the causal importance of the fusion vector after dimensionality reduction and dynamically allocates the privacy budget based on the causal importance.

[0021] 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.

[0022] Furthermore, the multimodal fusion vector after dimensionality reduction is obtained by using the K-means clustering algorithm, specifically including:

[0023] 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.

[0024] K-means clustering is used to select representative samples, retaining core information while reducing the dimension of fused features, thereby reducing computational overhead and information loss in the privacy protection process; the importance of causal associations is evaluated based on the graph convolutional network (GCN), which can dynamically allocate privacy budgets according to the actual degree of dependence between modalities, so that high-importance features can obtain stricter privacy protection and low-importance features can be processed in a lightweight manner, achieving a precise balance between privacy protection strength and data availability; the dynamic addition strategy of Gaussian noise not only meets the security requirements under the differential privacy theory framework, but also avoids excessive damage to feature semantics by traditional fixed noise mechanisms, ensuring that the feature vectors after privacy protection can still effectively support subsequent detection tasks.

[0025] Furthermore, the adding of perturbation samples and adversarial samples to the pre-trained model specifically includes:

[0026] S31, add perturbation samples to the self-supervised contrast loss function and the triplet loss function, and combine the two loss functions to form a new loss function to pre-train the model;

[0027] S32, add adversarial samples to the Mixup loss function to adversarially train the model.

[0028] Furthermore, the dynamic protection of model gradient privacy through the momentum mechanism specifically includes:

[0029] S33, the cloud server sends the initial model parameters to the edge node;

[0030] 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;

[0031] S35, updates the noise intensity via the momentum mechanism.

[0032] The combination of self-supervised contrastive learning and the triplet loss function strengthens the model's ability to discriminate semantic differences between samples, enabling the pre-trained model to learn highly discriminative feature representations even in unlabeled data scenarios. Mixup technology generates perturbed samples, enhancing the robustness of the model through semantic mixing in the input space, 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, while protecting the privacy of the model gradients and avoiding the interference of noise mutations on training stability, ensuring that the model can still maintain efficient parameter optimization capabilities and generalization performance during the privacy protection process.

[0033] Furthermore, the S4 specifically includes:

[0034] S41: The cloud server averages and aggregates the gradients uploaded by the edge nodes to obtain global model parameters, and sends the updated model parameters to the edge nodes for local model update.

[0035] S42, the edge node updates the stacked LSTM model based on the updated global model parameters, and learns features of different time scales through the stacked LSTM model;

[0036] S43, determining whether the traffic is abnormal by calculating the reconstruction error of each input.

[0037] The stacked LSTM structure extracts traffic features of different time scales layer by layer through a multi-layer network. The bottom-layer network captures short-term sudden anomalies, and the high-layer network mines long-term trend attack patterns, realizing multi-granular dynamic modeling of network traffic; the anomaly detection mechanism based on reconstruction error can accurately identify traffic behaviors that deviate from normal patterns, avoiding the problem of traditional single-scale models missing complex attack patterns; the collaborative training framework of edge nodes and cloud servers uses global model parameter updates to improve the consistency and accuracy of edge detection while protecting local data privacy, adapting to the real-time threat monitoring needs in distributed network environments.

[0038] A cyberspace mapping threat detection system with causal association privacy protection, comprising:

[0039] The data fusion module uses a dynamic fusion method for multimodal cyberspace mapping data based on multi-scale convolution and causal reasoning to extract features from multimodal mapping data, capture local details and global temporal dependencies, construct a dynamic causal association matrix, and achieve cross-modal dynamic feature fusion.

[0040] The privacy protection module uses a dynamic Gaussian noise privacy protection method based on GCN importance evaluation, combined with the K-means clustering method to reduce the dimension of fused features, dynamically allocate privacy budgets based on causal correlation importance, and protect the privacy of fused features.

[0041] 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.

[0042] The threat detection module is a cyberspace mapping threat detection method based on the stacked LSTM model. It uses a stacked structure to learn traffic characteristics at different time scales and mine abnormal patterns in network traffic.

[0043] The data fusion module overcomes the modal isolation limitations of traditional fusion methods through causal association modeling and multi-scale feature learning, providing the system with high-value fusion features. The privacy protection module combines clustering dimensionality reduction with dynamic privacy budget allocation to achieve hierarchical privacy protection while simultaneously performing lightweight processing, meeting the sensitive information protection needs of cyberspace mapping data. The gradient protection module uses dual perturbation and momentum noise mechanisms to build a full-process security protection system for pre-trained models, defending against attacks such as model inversion and gradient leakage. The threat detection module's stacked LSTM structure enables in-depth mining of the spatiotemporal characteristics of network traffic. This system significantly improves the accuracy, robustness, and privacy compliance of cyberspace threat detection.

[0044] Compared with the existing technology, 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, while capturing the local details and global temporal dependencies of text, image, and numerical modalities, it realizes the logical enhancement fusion of cross-modal information through the causal association matrix, and provides high-value density feature input for threat detection; based on the dynamic privacy protection method of GCN and K-means, it realizes the intelligent allocation of privacy budget through causal association importance evaluation, while reducing the feature dimension, it ensures the hierarchical protection of sensitive information, and balances data availability and security; self-supervised dual perturbation and momentum gradient protection technology, effectively resists model gradient leakage and adversarial attacks, and improves the robustness of pre-trained models; 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, and significantly improves the accuracy and generalization ability of threat detection. While ensuring data privacy compliance, the present invention achieves high-precision identification and real-time response to network security threats. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present drawings or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present drawings. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying any creative work.

[0046] Figure 1 is a flow chart of the method of the present invention;

[0047] Figure 2 This figure compares the accuracy of the embodiment of the present invention and the traditional cyberspace mapping threat detection method under different total privacy budgets;

[0048] Figure 3 A comparison chart of the accuracy of the embodiment of the present invention and the traditional cyberspace mapping threat detection method under different causal association strengths;

[0049] Figure 4 This is a comparison chart of the accuracy of the embodiment of the present invention and the traditional cyberspace mapping threat detection method under different disturbance intensities;

[0050] Figure 5 The figure compares the accuracy of the embodiment of the present invention and the traditional cyberspace mapping threat detection method under different time windows. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is described and illustrated below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended 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 work are within the scope of protection of the present invention.

[0052] In order to better understand the solutions of the embodiments of the present invention, some relevant terms and concepts that may be involved in the embodiments of the present invention are first introduced below.

[0053] (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 the node features and extracts the relationship and feature information between nodes through graph convolution operations.

[0054] (2) Long Short-Term Memory (LSTM) artificial neural networks are a type of time-recurrent neural network designed to address the long-term dependency issues inherent in general RNNs (recurrent neural networks). All RNNs have a chain-like structure of repeated neural network modules. In standard RNNs, this repeated module consists of a very simple structure, such as a tanh layer.

[0055] (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.

[0056] (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.

[0057] (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.

[0058] (6) Differential Privacy is a method in cryptography that aims to maximize the accuracy of data queries when querying from a statistical database while minimizing the chance of identifying its records.

[0059] 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:

[0060] S1, a dynamic fusion method for multimodal network spatial 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;

[0061] 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;

[0062] S3, an adaptive pre-trained model gradient protection method based on a self-supervised dual perturbation mechanism, adds perturbed samples and adversarial samples to the pre-trained model and dynamically protects the model gradient privacy through a momentum mechanism;

[0063] S4, a cyberspace mapping threat detection method based on a stacked LSTM model, uses a stacked structure to learn traffic characteristics at different time scales and mine abnormal patterns in network traffic.

[0064] Furthermore, the dynamic fusion method of multimodal cyberspace mapping data based on multi-scale convolution and causal reasoning specifically encodes 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.

[0065] Furthermore, S1 specifically includes the following steps:

[0066] S11, obtain text modal data, image modal data and numerical modal data, extract features from text and image data through BERT and Transformer encoders, and convert numerical data into high-dimensional feature representation through the embedding layer.

[0067] Specifically, two single-modal encoders are used: a text-image encoder and a numerical encoder to encode text, images, and numerical values ​​respectively. The formulas are as follows:

[0068] ;

[0069] in, It is a key Embedding

[0070] It is a pair The embedding of is generated by the BERT encoder;

[0071] Concat represents the concatenation operation;

[0072] represents a text information conversion operation performed by a context encoder, wherein the context encoder is specifically a Transformer encoder;

[0073] is the generated text feature.

[0074] ;

[0075] Wherein, I represents image data;

[0076] Indicates a convolution operation on the image data I to capture local details;

[0077] represents the fusion operation performed by the integrated encoder;

[0078] v is the generated image features.

[0079] For each time step t, the input numerical data , first convert it into a high-dimensional feature representation through the embedding layer, For the input numerical sequence, the feature extraction formula is:

[0080] ;

[0081] in, It is an embedding that maps numerical data into a high-dimensional space;

[0082] Embed the numerical features for this time step;

[0083] T is the time step.

[0084] S12 uses multi-scale convolution to extract multi-scale features from numerical data, calculates the similarity matrix between modalities, and performs Softmax normalization on the similarity matrix to obtain the final attention weight.

[0085] Specifically, the multi-scale convolution method is used to extract different scale features in numerical data. The formula is as follows:

[0086] ;

[0087] in, Indicates the use of convolution kernels of different sizes for feature extraction;

[0088] k is the scale of the convolutional layer;

[0089] It is the multi-scale features in the numerical data extracted by the convolutional layer, including local features and global features.

[0090] The dot product attention method is used to adapt to the multimodal fusion of cyberspace mapping. When calculating the similarity between text (q), numerical value (z), and image (v), dynamic weights, adaptive causal correction, and temporal information are added. The formula is as follows:

[0091] ;

[0092] in, : The query matrix (Query) at time step t is obtained by linear transformation, where is the modal data of mode m at time step t, is the query weight matrix;

[0093] : The key matrix (Key) at time step t, obtained by linear transformation, where is the mode at time step t The modal data of is the bond weight matrix;

[0094] for The transposed vector of

[0095] d is the feature dimension, used for scaling to avoid gradient explosion;

[0096] is the mode m pair mode The similarity at time step t, the larger the value, the stronger the correlation between the two modes;

[0097] Indicates the introduction of historical information and the consideration of the time evolution of the mode;

[0098] It represents a key parameter that captures the dynamic relationship between different modalities in the process of multimodal data fusion and is related to causal association;

[0099] It is the modal adaptive weight, which dynamically adjusts the importance of different modes;

[0100] It is a causal intervention correction term that removes non-causal correlations, such as noise and irrelevant features;

[0101] Refers to the control and intervention of the variable m' within the framework of causal reasoning.

[0102] Specifically, the following formula is used to calculate the similarity between the three modalities:

[0103] ;

[0104] ;

[0105] .

[0106] The similarity matrix is ​​obtained by calculating the similarity between modalities, and the similarity matrix is ​​normalized by Softmax to obtain the final attention weight:

[0107] ;

[0108] in, is m pairs of modes The attention weight of .

[0109] S13 introduces causal reasoning to calculate the causal effect of each modality on the final output, reduces the imbalance of causal effects through causal correction terms, and optimizes the information fusion of different modalities.

[0110] Specifically, causal reasoning is introduced to optimize the relationship modeling between modalities to cope with the changing and complex environment in cyberspace mapping. The specific causal effect formula is as follows:

[0111] ;

[0112] ;

[0113] ;

[0114] ;

[0115] in, Text-modal attention The causal effect on the final output O;

[0116] Image modality attention The causal effect on the final output O;

[0117] Numerical mode attention The causal effect on the final output O;

[0118] To integrate the causal effects of the three modalities;

[0119] Representing text modal attention expected value;

[0120] Representing image modality attention expected value;

[0121] Represents numerical mode attention expected value;

[0122] Represents multi-modal attention ( 、 、 ) is combined together to measure the combined effect of multimodal collaboration on the final output O, reflecting the overall causal effect of each modal attention under the fusion of multiple factors, rather than the simple summation of independent effects;

[0123] is the conditional probability, indicating the attention given the text modality , text features q and their related probability parameters In the case of , the probability of the final output is O;

[0124] is the conditional probability, indicating the attention of the given image modality , image features v and their related probability parameters In the case of , the probability of the final output is O;

[0125] is the conditional probability, indicating the attention given the numerical mode , numerical features z and their related probability parameters In the case of , the probability of the final output is O;

[0126] is the conditional probability of multimodality, indicating that given all modal attentions ( 、 、 ) and all modal features (q, v, z), the probability of the final output O;

[0127] Express Intervene, force The value of ;

[0128] Express Intervene, force The value of ;

[0129] Express Intervene, force The value of ;

[0130] For After intervention, given q and In the case of , the final output is the conditional probability of O;

[0131] For After intervention, given v and In the case of , the final output is the conditional probability of O;

[0132] For After intervention, given z and In the case of , the final output is the conditional probability of O;

[0133] After intervening on multiple modalities, given all modal attention ( 、 、 ) and all modal features (q, v, z), the final output is the conditional probability of O.

[0134] Design a dynamic attention causal correction method to reduce the imbalance of causal effects and thus optimize the information fusion of different modalities. Introduce the causal correction term , dynamically adjust the relationship between modalities according to the results of counterfactual reasoning:

[0135] ;

[0136] in, Represents the causal influence between the i-th node and the j-th node;

[0137] Represents the feature vector of node i;

[0138] The transposed eigenvector of node j;

[0139] is the dimension of the k-th feature.

[0140] Causal Modifier Dynamically calculated using the following formula:

[0141] ;

[0142] It represents a hyperparameter in a causal correction process and adjusts the correction strength.

[0143] S14, based on the extracted multi-scale features, the final cross-modal fusion result is obtained through dynamic weighting and weighted summation of causal correction terms. The specific formula is as follows:

[0144] ;

[0145] ;

[0146] in, is the dynamic weight of each scale feature;

[0147] is the final fusion result;

[0148] is the weight corresponding to each modal feature, indicating the impact of each feature on the final result;

[0149] Represents the fused feature or modality, which is the intermediate result of fusing multiple features.

[0150] Furthermore, S2 specifically reduces the dimensionality of the fusion vector based on K-means clustering sampling, dynamically allocates the privacy budget based on the causal importance metric, uses a graph convolutional network (GCN) to calculate the causal importance of the fusion vector, and controls the noise intensity based on α-stable distributed noise. It includes the following steps:

[0151] S21, the edge node receives the multimodal fusion vector and obtains the multimodal fusion vector after dimensionality reduction through the K-means clustering algorithm.

[0152] Specifically, edge nodes select N centers through the K-means clustering algorithm, that is, sample N vectors, to improve the diversity and representativeness of the sample set, and achieve a better privacy-utility balance in multimodal privacy protection scenarios.

[0153] From the fused vector A vector is randomly selected as the first cluster center, denoted as c1.

[0154] For each vector , calculate its difference with the currently selected cluster The distance is calculated as follows:

[0155] ;

[0156] in, Represents the Euclidean distance between vectors;

[0157] represents the jth cluster center.

[0158] The probability of selecting the next center is calculated based on the square of the distance from each point to the nearest center. Specifically, for each vector , the probability of being selected as the next center It is given by the following formula:

[0159] ;

[0160] According to the calculated probability distribution, from The next cluster center is selected according to the probability, recorded as , repeat until convergence, and select N cluster centers. These N centers are the sampling vectors.

[0161] Convergence judgment condition: the change of cluster center between round t and round t-1 is less than the set threshold r:

[0162] ;

[0163] in, is the position of the i-th cluster center in the t-th round.

[0164] S22 uses a graph convolutional network to calculate the causal importance of the fused vector after dimensionality reduction, and dynamically allocates the privacy budget based on the causal importance. The specific formula is as follows:

[0165] ;

[0166] in, 、 Represent the preset minimum and maximum privacy budgets, respectively, usually ;

[0167] To control the privacy budget The rate of change ( );

[0168] Indicates the causal importance of node j to node i;

[0169] It represents the privacy budget allocated to the causal relationship between node i and node j, and controls the noise protection strength of the relationship feature.

[0170] In the graph convolutional network (GCN), the importance of causal relationships can be dynamically measured through the attention coefficient. The specific formula is as follows:

[0171] ;

[0172] Where W is the learnable weight matrix;

[0173] is the attention parameter vector, yes The transposed vector of

[0174] 、 and Represent the characteristics of node i, node j and node k respectively;

[0175] Represents vector concatenation;

[0176] LeakyReLU is a leaky linear rectification function;

[0177] is the set of all nodes connected to node i.

[0178] 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. The specific process is as follows:

[0179] Fusion feature vector after dimensionality reduction , add noise to it, the noise adding process is as follows:

[0180] ;

[0181] in, is the eigenvector after adding noise;

[0182] is the intensity of the noise, which controls the amplitude of the noise;

[0183] is a noise vector from an α-stable distribution.

[0184] Noise Vector Each component (for ) comes from the α-stable distribution, and the characteristic function of the α-stable distribution is:

[0185] ;

[0186] in, Control the tails of the distribution;

[0187] When is Gaussian distribution;

[0188] When is a heavy-tailed distribution with infinite variance;

[0189] u is the noise vector The value of a component of , used to calculate one of the expected variables;

[0190] i is an index indicating the noise vector The amount in.

[0191] Under the dynamic privacy budget allocation mechanism, the noise intensity It is directly related to the causal importance of the vector; and the privacy budget is dynamically adjusted according to the real-time changes of modalities such as text, images, and numbers.

[0192] Specifically, Calculation is performed according to the following formula:

[0193] ;

[0194] in, represents the total privacy budget allocated to the causal relationships between node k and other nodes.

[0195] Furthermore, S3 specifically improves model robustness by adding double perturbation to model pre-training. It also protects model parameter privacy while improving model performance through dynamic gradient noise. The steps include:

[0196] S31, add perturbation samples to the self-supervised contrast loss function and the triplet loss function, and combine the two loss functions to form a new loss function to pre-train the model.

[0197] Specifically, based on the sampled and noisy data, self-supervised dual perturbation is used to enhance the robustness of the model by generating adversarial samples and perturb the model itself, thereby improving the accuracy of traffic anomaly recognition and reducing the negative impact of noise on pre-training.

[0198] The first stage (pre-training stage):

[0199] Self-supervised learning:

[0200] ;

[0201] Wherein, m is the sample vector after noise addition, i.e., the feature vector after noise addition in S23;

[0202] is the feature representation of the sample vector m after adding noise through the encoder;

[0203] is a positive sample related to m;

[0204] is a sample that is not similar to sample m, and a negative sample is taken here;

[0205] sim is the similarity function;

[0206] is the self-supervised contrastive loss function.

[0207] Triplet loss function:

[0208] ;

[0209] in, is the first hyperparameter, which represents the minimum interval between positive and negative samples;

[0210] is the ternary loss function;

[0211] is a negative sample.

[0212] Combine the triplet loss with the original contrast loss to form a new loss function:

[0213] ;

[0214] in, is the second hyperparameter, which is used to control the weight of triplet loss in the overall loss.

[0215] S32, add adversarial samples to the Mixup loss function to adversarially train the model.

[0216] Specifically, the second stage (adversarial training stage):

[0217] Use Mixup technology to perform model perturbation, the formula is:

[0218] ;

[0219] in, is the mixed sample of input;

[0220] is an attacked model sample, which is obtained by attacking the model as a mixed target sample;

[0221] Input samples to the model predictions;

[0222] are weights randomly sampled from the range [0,1].

[0223] Input mixing: for input samples and Perform weighted averaging to obtain mixed samples:

[0224] ;

[0225] in, is the third hyperparameter sampled from the uniform distribution U(0,1).

[0226] for The true label of

[0227] It is an adversarial example.

[0228] In step 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. The noise intensity is updated through the momentum mechanism.

[0229] After receiving the initial model parameters, the edge node calculates the gradient based on the local model. During the calculation, the edge node adds noise to the gradient to ensure the privacy of the hidden data.

[0230] The gradient is calculated using the following formula:

[0231] ;

[0232] in, are local model parameters;

[0233] Parameters delivered to the cloud server;

[0234] is the gradient operator;

[0235] is the loss function.

[0236] The noise intensity Added to the gradient, the updated gradient is:

[0237] ;

[0238] in, represents the added Gaussian noise.

[0239] The noise update uses the momentum mechanism to avoid excessive noise changes. The specific formula is:

[0240] ;

[0241] in, is the noise intensity of the tth round;

[0242] is the change in noise, calculated based on the change in the loss function;

[0243] ρ is the decay rate, which is used to adjust the smoothness of the noise intensity update process.

[0244] Furthermore, S4 specifically involves averaging and aggregating the uploaded gradients on the cloud server, sending model parameters to the edge nodes for local model updates. The edge nodes process the data by stacking multiple LSTM layers, learning features of different time scales at different levels, and performing threat detection based on the error. This specifically includes the following steps:

[0245] In step S41, the cloud server averages and aggregates the gradients uploaded by the edge nodes to obtain the global model parameters, and sends the updated model parameters to the edge nodes for local model update. The formula is as follows:

[0246] ;

[0247] Where M is the number of edge nodes;

[0248] is the global model parameter;

[0249] is the model parameter of the i-th client;

[0250] is the local data volume of the i-th client;

[0251] is the total data volume.

[0252] In S42, the edge node updates the stacked LSTM model based on the updated global model parameters and learns features of different time scales through the stacked LSTM model.

[0253] Specifically, the stacked LSTM threat detection model is updated based on the model parameters sent by the cloud server. Through a multi-level learning mechanism, the stacked LSTM model can accurately capture long-term and short-term dependencies in network traffic, extract effective features from noisy data, and improve the accuracy and robustness of threat detection.

[0254] The stacked LSTM architecture consists of:

[0255] Forget Gate:

[0256] ;

[0257] in, is the sigmoid activation function;

[0258] is the weight matrix of the forget gate;

[0259] is the hidden state at the previous moment, which contains historical information about past traffic patterns;

[0260] is the input numerical data at each time step t;

[0261] The vector formed by concatenating the hidden state at the previous moment with the input at the current moment. By fusing historical traffic patterns with current real-time data, the LSTM can capture the temporal dependencies in traffic and identify abnormal patterns.

[0262] is the forget gate bias term;

[0263] is the output of the forget gate.

[0264] Input Gate:

[0265] ;

[0266] in, is the weight matrix of the input gate;

[0267] is the input gate bias term;

[0268] is the output of the input gate.

[0269] Candidate memory cells:

[0270] ;

[0271] in, is the weight matrix of the candidate memory unit;

[0272] is a candidate cell state;

[0273] is the hyperbolic tangent activation function;

[0274] is the candidate memory cell bias term.

[0275] Unit status update:

[0276] ;

[0277] in, Indicates the cell state at the current moment.

[0278] Output Gate:

[0279] ;

[0280] in, is the weight matrix of the output gate;

[0281] is the output gate bias term.

[0282] Hide status update:

[0283] .

[0284] SkLSTM processes data by stacking multiple LSTM layers, and the hidden state output by the first LSTM layer As input to the second LSTM layer:

[0285] .

[0286] Through such a stacked structure, SkLSTM is able to learn features of different time scales at different levels.

[0287] S43, determining whether the traffic is abnormal by calculating the reconstruction error of each input.

[0288] By calculating the flow data for each input The reconstruction error To determine whether the traffic is abnormal, the specific formula is as follows:

[0289] ;

[0290] in, is the reconstructed input, that is, the reconstructed output generated by the stacked LSTM model after processing the original input data.

[0291] If the reconstruction error Exceeding the set threshold , it is determined to be a threat. The specific formula is as follows:

[0292] .

[0293] The present invention also provides a cyberspace mapping threat detection system for causal association privacy protection, which specifically includes the following modules:

[0294] The data fusion module is a dynamic fusion method of multimodal cyberspace mapping data based on multi-scale convolution and causal reasoning. It 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.

[0295] 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. In order to capture the deep relationship between the modalities, a structural causal model is used to quantify the causal impact of different modalities on the final mapping results, and dynamic attention causal correction is used to optimize the information fusion of different modalities.

[0296] The privacy protection module uses a dynamic Gaussian noise privacy protection method based on GCN importance evaluation, combined with the K-means clustering method to reduce the dimension of fused features, dynamically allocate privacy budgets based on causal correlation importance, and protect the privacy of fused features.

[0297] 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.

[0298] 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 the momentum mechanism.

[0299] 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.

[0300] The threat detection module is a cyberspace mapping threat detection method based on the stacked LSTM model. It uses a stacked structure to learn traffic characteristics at different time scales and mine abnormal patterns in network traffic.

[0301] Specifically, the threat detection module is based on the cyberspace mapping anomaly detection module of the stacked LSTM model. This module processes data by stacking multiple LSTM layers, learning features of different time scales at different levels and performing threat detection based on errors. This module mines anomalous patterns in network traffic to achieve highly robust threat detection.

[0302] 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 meets the requirements of differential privacy protection.

[0303] After multimodal data is fused and reduced in dimension, a special type of noise (α-stable distributed noise) is added to these features. The intensity of this noise is initially set at a baseline value, and each feature is assigned a different privacy budget based on its sensitivity. For example, key features with strong causal relationships receive a higher privacy budget, resulting in the addition of stronger noise to mask sensitive information. Because this noise addition method adheres to the mathematical definition of differential privacy, it ensures the privacy of the fused features, preventing attackers from inferring true sensitive information from the data.

[0304] During model training, the gradients of model updates also contain sensitive information. To address this, we add a different type of noise (Gaussian noise) to the gradients, controlling the noise intensity using a privacy budget. A higher budget increases the added noise intensity, effectively masking sensitive information in the gradients. This approach also meets the requirements of differential privacy, preventing attackers from inferring the original data through the gradients while ensuring the model can properly learn cyber threat patterns.

[0305] The two privacy-preserving measures described above (feature-based noising and gradient-based noising) are implemented independently. According to the combinatorial principle of differential privacy, when multiple privacy-preserving methods are used in parallel, the overall privacy protection strength can be measured by combining their respective privacy budgets. Therefore, the overall solution of this invention combines the privacy budgets of feature-based and gradient-based noising to form a final total privacy budget. This mathematically proves that the entire threat detection method meets the differential privacy criteria, achieving a high level of privacy protection for multimodal network data.

[0306] In this embodiment, based on the real network threat dataset RT-IoT, different parameters are used: total privacy budget, causal correlation strength, disturbance strength, and time window to evaluate the threat detection accuracy of the present invention for privacy protection of cyberspace mapping data. The results of the comparative experiment are as follows: Figure 2-Figure 5 As shown, from Figure 2-Figure 5 It can be seen that the cyberspace mapping threat detection effect of this embodiment is better than the traditional method under different parameters.

[0307] It should be noted that the present invention is not limited to the above-mentioned embodiments. The above-mentioned embodiments are merely examples, and any embodiments having substantially the same structure and effect as the technical concept within the scope of the technical solution of the present invention are all included in the technical scope of the present invention. In addition, without departing from the scope of the present invention, other embodiments that can be conceived by those skilled in the art and that combine some of the constituent elements in the embodiments are also included in the scope of the present invention.

Claims

1. A cyberspace mapping threat detection method with causal privacy protection, characterized by: The following steps are involved: S1, a dynamic fusion method for multimodal cyberspace 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; specifically, it includes: S11, obtain text modal data, image modal data, and numerical modal data, perform feature extraction on the text and image data through BERT and Transformer encoders, and convert the numerical data into high-dimensional feature representation through the embedding layer; S12 uses multi-scale convolution to extract multi-scale features from numerical data, calculates the similarity matrix between modalities, and performs Softmax normalization on the similarity matrix to obtain the final attention weight; S13, introduces causal reasoning, calculates the causal effect of each modality on the final output, reduces the imbalance of causal effects through causal correction terms, and optimizes the information fusion of different modalities; S14, based on the extracted multi-scale features, the final cross-modal fusion result is obtained through dynamic weighting and weighted summation of causal correction terms; S2, a dynamic Gaussian noise privacy protection method based on GCN importance evaluation, combined with the K-means clustering method, reduces the dimension of fused features, dynamically allocates privacy budgets based on causal correlation importance, and protects the privacy of fused features; specifically, it 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 a graph convolutional network to calculate the causal importance of the fusion vector after dimensionality 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; S3, an adaptive pre-trained model gradient protection method based on a self-supervised dual perturbation mechanism, adds perturbed 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 characteristics at different time scales and mine abnormal patterns in network traffic.

2. The causal privacy-preserving cyberspace mapping threat detection method according to claim 1, wherein: The multimodal fusion vector after dimensionality reduction obtained by the K-means clustering algorithm specifically includes: 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.

3. The causal privacy-preserving cyberspace mapping threat detection method according to claim 1, wherein: Adding perturbation samples and adversarial samples to the pre-trained model specifically includes: S31, add perturbation samples to the self-supervised contrast loss function and the triplet loss function, and combine 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.

4. The causal privacy-preserving cyberspace mapping threat detection method according to claim 1, wherein: The dynamic protection of 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, updates the noise intensity via the momentum mechanism.

5. The causal privacy-preserving cyberspace mapping threat detection method according to claim 1, wherein: The S4 specifically includes: S41: The cloud server averages and aggregates the gradients uploaded by the edge nodes to obtain global model parameters, and sends 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 of different time scales through the stacked LSTM model; S43, determining whether the traffic is abnormal by calculating the reconstruction error of each input.

6. A cyberspace mapping threat detection system with causal privacy protection, characterized by: include: The data fusion module uses a dynamic fusion method for multimodal cyberspace mapping data based on multi-scale convolution and causal reasoning to extract features from multimodal mapping data, capture local details and global temporal dependencies, construct a dynamic causal association matrix, and achieve cross-modal dynamic feature fusion. Specifically, it includes: S11, obtain text modal data, image modal data, and numerical modal data, perform feature extraction on the text and image data through BERT and Transformer encoders, and convert the numerical data into high-dimensional feature representation through the embedding layer; S12 uses multi-scale convolution to extract multi-scale features from numerical data, calculates the similarity matrix between modalities, and performs Softmax normalization on the similarity matrix to obtain the final attention weight; S13, introduces causal reasoning, calculates the causal effect of each modality on the final output, reduces the imbalance of causal effects through causal correction terms, and optimizes the information fusion of different modalities; S14, based on the extracted multi-scale features, the final cross-modal fusion result is obtained through dynamic weighting and weighted summation of causal correction terms; The privacy protection module uses a dynamic Gaussian noise privacy protection method based on GCN importance evaluation, combined with the 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. Specifically, it 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 a graph convolutional network to calculate the causal importance of the fusion vector after dimensionality 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; 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. The threat detection module is a cyberspace mapping threat detection method based on the stacked LSTM model. It uses a stacked structure to learn traffic characteristics at different time scales and mine abnormal patterns in network traffic.

Images