An explainable attack intention determination method for an industrial internet edge gateway

Abstract

The present application relates to an explainable attack intention determination method for an industrial internet edge gateway, belonging to the technical field of artificial intelligence and industrial internet security, and comprising the following steps: establishing an instance representation packet-level data set with a time window level attack label for the edge gateway; generating an IF-THEN type fuzzy rule antecedent by FCM clustering; constructing a joint optimization objective function combining packet-level attack labels based on a TSK fuzzy system and a fusion mechanism of fuzzy confidence and gated attention aggregation, to train fuzzy rule consequent parameters and gated attention weights; and constructing an explainable knowledge base and attack intention determination model after joint optimization, to output attack intention labels and their explanation basis; the method can realize industrial internet attack intention recognition in the absence of accurate labeling of a single communication event, improve the explainability of the determination process, the model generalization ability and the security operation response efficiency, and is suitable for security detection of industrial internet edge gateways and industrial control networks.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and industrial internet security technology, specifically to a method for determining interpretable attack intent for industrial internet edge gateways. Background Technology

[0002] With the development of Industrial Internet, edge computing, and intelligent manufacturing technologies, devices such as PLCs, sensors, actuators, CNC equipment, industrial robots, and monitoring systems in industrial settings are increasingly connected to enterprise networks, cloud platforms, or Industrial Internet platforms through Industrial Internet edge gateways. Industrial Internet edge gateways typically handle protocol conversion, data acquisition, edge computing, device access, and security isolation, serving as a crucial hub connecting field devices, control networks, and upper-layer application systems. While edge gateways improve the efficiency of industrial data aggregation, remote operation and maintenance, and production collaboration, they also expose industrial control networks to more complex cyberattack environments.

[0003] The environment in which industrial internet edge gateways operate is characterized by diverse protocol types, strong device heterogeneity, high business continuity requirements, strict real-time requirements, and a wide impact from security incidents. Attacks targeting the industrial internet edge can manifest as abnormal scanning, unauthorized access, malicious control command injection, abnormal industrial protocol messages, device status tampering, denial-of-service attacks, and multi-stage combined attacks. These attacks are often mixed in with normal industrial communication traffic and control processes. A single communication event may not be sufficient to directly determine the attack intent. Only by combining multi-source information such as network traffic, control commands, device status, and security alarms over a time window can the attack behavior and its potential intent be more accurately identified.

[0004] Existing industrial internet security detection methods mainly include rule-based feature matching methods, statistical anomaly detection methods, and intrusion detection methods based on machine learning or deep learning. Rule-based methods rely on expert experience and pre-set feature libraries, and while they have some detection capability against known attacks, they are less adaptable to unknown attacks, variant attacks, and cross-protocol combined attacks. Deep learning-based methods can learn complex features from large amounts of industrial communication data, giving them an advantage in detection performance. However, their training typically relies on sufficient and accurately labeled data, while security alarms in industrial settings are mostly labeled at the time period, device, or event level, making it difficult to provide accurate instance-level attack labels for each communication message, control command, or device status record.

[0005] Furthermore, the data collected by industrial internet edge gateways typically exhibits characteristics such as high dimensionality, strong noise, multi-source heterogeneity, and temporal correlation. Directly utilizing these high-dimensional industrial communication features for modeling makes them susceptible to redundant features, noise disturbances, and abnormal fluctuations, leading to unstable localization of critical attack instances. While existing attention mechanisms can assign weights to different instances, their weight allocation process is primarily data-driven, lacking confidence constraints that are combined with industrial security knowledge or rule-based reasoning results. In complex industrial scenarios, this may result in issues such as misidentification of critical instances, unstable attention weights, and difficulty in interpreting the judgment criteria.

[0006] In industrial security operations and maintenance scenarios, attack intent determination results not only need to indicate the existence or type of attack, but also need to explain the key communication instances that triggered the determination, the relevant rules and their confidence levels, so that security operations and maintenance personnel can quickly locate the source of the anomaly, understand the attack path, and take appropriate measures. However, existing detection methods based on black-box models often fail to provide clear and traceable determination criteria, making it difficult to meet the comprehensive requirements of industrial control networks for security, reliability, real-time performance, and interpretability. Therefore, there is an urgent need for an interpretable attack intent determination method for industrial internet edge gateways to solve the above-mentioned technical problems. Summary of the Invention

[0007] The purpose of this invention is to provide an interpretable attack intent determination method for industrial internet edge gateways, overcoming the problem in the prior art that it is difficult to achieve interpretable attack intent determination in the case of a lack of accurate labeling of individual communication events in the security detection scenario of industrial internet edge gateways.

[0008] To address the aforementioned technical problems, the present invention adopts the following technical solution: a method for determining interpretable attack intent for industrial internet edge gateways, comprising the following steps:

[0009] Establish a packet-level dataset of instance representations with time window-level attack labels for edge gateways, and obtain low-dimensional instance feature representations after preprocessing;

[0010] The fuzzy C-means algorithm is used to cluster the feature representations of low-dimensional instances, and the antecedents of IF-THEN type fuzzy rules in the interpretable knowledge base are constructed based on the clustering results.

[0011] A TSK fuzzy system is built. Based on the aforementioned antecedent, the rule activation results and instance fuzzy confidence of the low-dimensional instance feature representation are calculated. Packet-level attack labels are used as supervision signals, and instance fuzzy confidence is introduced as a semantic prior into a gating attention aggregation mechanism to construct a coupled joint optimization objective function.

[0012] The joint optimization objective function is optimized to identify the fuzzy rule consequent parameters and update the gradient of the gating attention weights, forming an interpretable knowledge base and attack intent determination model.

[0013] The features of each instance in the packet-level dataset are used as input to the knowledge base and attack intent determination model. The output is the attack intent label of the corresponding time window. Based on the final instance, the weights and rule activation strengths are aggregated, and the key instances, key rules, rule activation values ​​and instance confidence are backtracked and output to form the basis for decision interpretability.

[0014] Preferably, feature extraction is performed on the network traffic, control commands, device status, and security alarm data collected by the edge gateway within a preset time window, and each data item in each packet is converted into instance features to construct a packet-level dataset based on instance feature representation; wherein, each packet has a time window-level attack label, and each instance in the packet does not have an instance-level label.

[0015] Preferably, the preprocessing includes: performing linear transformation, nonlinear activation, and normalization on each instance feature in the instance representation packet-level dataset to project the high-dimensional industrial communication features into a low-dimensional instance feature representation.

[0016] Preferably, the fuzzy C-means algorithm takes low-dimensional instance feature representation as input to obtain K cluster centers and membership matrices; and updates the cluster centers and membership matrices through alternating optimization until a preset convergence condition is met to obtain optimized membership matrices and cluster centers; the optimized cluster centers serve as the centers of the antecedents of IF-THEN type fuzzy rules.

[0017] Preferably, the TSK fuzzy system uses a Gaussian membership function to calculate the degree of matching of low-dimensional instance feature representations with respect to each fuzzy rule, and obtains the rule activation result based on the degree of matching.

[0018] Preferably, the k-th IF-THEN type fuzzy rule is expressed as:

[0019] ;

[0020] in, This represents the feature representation of a low-dimensional instance. This represents the dimension of the feature representation of a low-dimensional instance. Indicates the first Rule # The fuzzy set corresponding to each feature and They represent the first The consequent weight parameters and bias parameters of the rule. This represents the linear output of the k-th fuzzy rule.

[0021] Preferably, the gated attention aggregation mechanism includes a two-branch gating structure, wherein the first branch uses the tanh activation function to extract the relevance representation of instance features, and the second branch uses the sigmoid activation function to form gating weights to suppress invalid or noisy instance features; the outputs of the first branch and the second branch are fused element-wise and mapped to attention scores, and then normalized by Softmax to obtain class-related attention weights. Finally, the instance fuzzy confidence is fused with the class-related attention weights to obtain the final instance aggregation weights.

[0022] Preferably, the weights are aggregated based on the final instance. We perform a weighted summation of instance features within the same package to obtain a category-related time-window aggregated representation:

[0023] ;

[0024] in, This represents the intermediate feature representation obtained after the first nonlinear transformation; based on the time window-level aggregation representation, the attack intent label is output, and instances with higher attention weights are taken as key communication instances, and fuzzy rules with higher rule activation strength are taken as key fuzzy rules. At the same time, the rule activation result, instance fuzzy confidence, and final instance aggregation weight are output as provided in the above technical solution.

[0025] Beneficial effects: This invention initializes fuzzy rule centers based on low-dimensional instance feature representation using the fuzzy C-means algorithm and constructs IF-THEN type fuzzy rules using the TSK fuzzy system, enabling the model to express fuzzy prototypes of different industrial communication behaviors or attack patterns in rule form. Specifically, the TSK fuzzy system calculates the rule activation strength and instance fuzzy confidence of each instance, and introduces the instance fuzzy confidence as a semantic prior into a gated attention aggregation mechanism to calibrate the attention weights of each instance within a packet. Compared to aggregation methods that rely solely on data-driven attention weights, this invention can simultaneously combine fuzzy rule evidence and gated attention weights to identify key communication instances, reducing the interference of noisy instances, irrelevant instances, and abnormal fluctuations on the attack intent determination results, and improving the reliability of key instance location.

[0026] Furthermore, while outputting attack intent labels, this invention also outputs explanatory information such as key communication instances, key fuzzy rules, rule activation strength, instance fuzzy confidence, category-related attention weights, and final instance aggregation weights. This enables industrial security operations personnel to understand the source of the model's judgment results and quickly locate abnormal communication behaviors and potential attack paths. Because the judgment process of this invention is based on TSK fuzzy rules and gated attention aggregation mechanisms, it has better readability, traceability, and interpretability than traditional black-box detection models, making it suitable for security detection and operation and maintenance response scenarios in industrial internet edge gateways and industrial control networks. Experimental results show that the method of this invention performs well in terms of test performance and interpretability. Attached Figure Description

[0027] Figure 1 This is an overall flowchart of the method of the present invention;

[0028] Figure 2 This is an experimental flowchart of the method of the present invention;

[0029] Figure 3 This is an overall architecture diagram of the method of the present invention during the training phase;

[0030] Figure 4 This is a diagram illustrating the attack intent determination and interpretation output process of the method of the present invention during the inference phase. Detailed Implementation

[0031] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.

[0032] An example is a method for determining interpretable attack intent for industrial internet edge gateways, such as... Figure 1 and Figure 4 As shown, it includes the following steps:

[0033] S1: Use a large model to extract features from network traffic, control commands, device status and security alarm data collected by the industrial internet edge gateway, convert each data item in each preset time window into instance features, and construct a packet-level dataset based on instance feature representation;

[0034] In this embodiment, after performing protocol parsing, time alignment, outlier handling, missing value completion, and normalization on the network traffic, control commands, device status, and security alarm data, feature information is extracted using a pre-trained language model with industrial communication semantic parsing capabilities. The time-window level attack labels include 0 representing a normal state and 1 representing an attack state, or at least one attack intent category label from the following: normal, scanning and probing, abnormal control, data theft, denial of service, and abnormal device linkage, resulting in a packet-level dataset. ;in ;in, Indicates the number of packages, each package Corresponding to a preset time window, and by indivual Dimensional instance feature composition; Indicates the first The first of the packages Individual instance characteristics; for each package Labeling time window level attacks The instances within the package do not have instance tags.

[0035] S2: Perform low-dimensional feature projection on the high-dimensional industrial communication features of all instances in the package-level dataset to suppress redundant information and noise interference, and obtain a low-dimensional instance feature representation suitable for fuzzy reasoning and attention aggregation;

[0036] Among them, for the first The first of the packages Individual Instance Features The feature adaptation process is represented as follows:

[0037] ;

[0038] ;

[0039] in, Indicates the first The first of the packages High-dimensional instance features This represents the intermediate feature representation obtained after the first nonlinear transformation. This represents the low-dimensional instance feature representation after layer normalization. , and For learnable weight matrix, , and The bias term is ReLU, which is a non-linear activation function. For the Sigmoid function, LN represents layer normalization. and These represent the learnable scaling and translation parameters, respectively. This represents element-wise product.

[0040] S3: The fuzzy C-means algorithm is used to cluster the low-dimensional instance feature representations to obtain the membership matrix and cluster centers. Based on the clustering results, the antecedents of IF-THEN type fuzzy rules in the interpretable knowledge base are constructed.

[0041] In this embodiment, the objective function of the fuzzy C-means algorithm is expressed as:

[0042] ;

[0043] The constraints are:

[0044] ;

[0045] in: Indicates the total number of instances. Indicates the number of clusters. Indicates the first The instance is for the first Cluster centers membership degree Indicates the fuzzy weight index; The larger the value, the more ambiguous the clustering results. When the value approaches 1, FCM degenerates into the K-Means algorithm for hard clustering.

[0046] The membership degree and cluster center are updated by alternating optimization until the preset convergence condition is met, resulting in the optimized membership degree matrix and cluster center. The cluster center is used to characterize the fuzzy prototype of different industrial communication behavior patterns or attack patterns, and serves as the center of the antecedent of IF-THEN type fuzzy rules.

[0047] S4: Build a TSK fuzzy system, use IF-THEN type fuzzy rule antecedents to calculate the rule activation results, instance fuzzy confidence and fuzzy system output corresponding to each instance feature, and combine time window level attack labels to construct a joint optimization objective function that integrates fuzzy confidence and gating attention aggregation mechanism.

[0048] In this embodiment, the TSK fuzzy system uses a Gaussian membership function to calculate the degree of matching between the low-dimensional instance feature representation and each fuzzy rule, and obtains the rule activation result based on the degree of matching; the k-th IF-THEN type fuzzy rule is represented as:

[0049] ;

[0050] in, This represents the feature representation of a low-dimensional instance. This represents the dimension of the feature representation of a low-dimensional instance. Indicates the first Rule # The fuzzy set corresponding to each feature and They represent the first The consequent weight parameters and bias parameters of the rule. This represents the linear output of the \(k\)th fuzzy rule.

[0051] ( ) indicates the first The first fuzzy rule A fuzzy set of features is defined by a membership function, i.e. , This represents the k-th ensemble cluster center obtained above. The first of the preceding cluster centers 1 eigenvalue, , ; It is the first Linear output under fuzzy rules It is the feature transformation matrix, which determines how the input features affect the rule output. It is the baseline adjustment item, which determines the rule input when the input is 0.

[0052] For low-dimensional instance feature representation ; Calculate its rule activation strength relative to each fuzzy rule. And calculate the fuzzy confidence of the instance based on the importance weights of the learnable rules:

[0053] ;

[0054] ;

[0055] in, Indicates the first The learnable importance parameters of a fuzzy rule. This represents the normalized importance weight of the rules. Indicates the first The first of the packages Fuzzy confidence of each instance; instance fuzzy confidence is used to characterize the degree of trustworthy matching between the corresponding instance and the effective industrial communication pattern or attack pattern, and serves as a semantic prior for the gating attention aggregation mechanism.

[0056] The joint optimization objective function includes a classification loss based on time-window-level attack labels, a rule sparsity regularization term to constrain the importance weights of fuzzy rules, and an attention smoothing regularization term to stabilize the distribution of instance attention weights. Through end-to-end joint optimization, it simultaneously updates the low-dimensional feature projection parameters, fuzzy rule antecedent parameters, fuzzy rule consequent parameters, rule importance weight parameters, and the weight parameters of the gated attention aggregation mechanism. The joint optimization objective function is expressed as:

[0057] ;

[0058] in, Indicates classification loss; Represents sparse regularization terms for rules. This represents an attention smoothing regularization term; and This represents the regularization coefficient.

[0059] S5: The joint optimization method is used to optimize the joint optimization objective function, obtain the optimized fuzzy rule consequent parameters and the weight parameters of the gating attention aggregation mechanism, and construct an interpretable knowledge base and an attack intent determination model.

[0060] In this embodiment, the backpropagation algorithm and the chain rule of gradient descent are used to jointly update the low-dimensional feature projection parameters, fuzzy rule antecedent parameters, fuzzy rule consequent parameters, rule importance weight parameters, and the weight parameters of the gated attention aggregation mechanism, so that the fuzzy rule output, instance fuzzy confidence and attention aggregation weight are collaboratively optimized under the supervision of time window-level attack labels.

[0061] S6: Input the output results of the interpretable knowledge base on the features of each instance in the package into the optimized gating attention aggregation mechanism to obtain the attack intent label of the corresponding time window, and output the key communication instance, key fuzzy rule, rule activation result and instance fuzzy confidence as the basis for interpretation.

[0062] The gated attention aggregation mechanism includes a two-branch gating structure. The first branch uses the tanh activation function to extract the relevance representation of instance features, while the second branch uses the sigmoid activation function to form gating weights to suppress invalid or noisy instance features. The outputs of the first and second branches are fused element-wise and mapped to an attention score. In this embodiment, for the first... The first of the packages The first instance, for the first The gating attention score for attack intent is represented as:

[0063] ;

[0064] in, Indicates the first The first of the packages The first instance is for the first The original attention score for the type of attack intent; Indicates the first The first of the packages Low-dimensional feature representation of an instance; and Indicates that for the first Learnable weight matrices representing attack intent are used to extract relevance representations of instance features and generate gating weights, respectively. Indicates that for the first Learnable attention projection parameters for attack intents are used to map the fused features into scalar attention scores; Represents the hyperbolic tangent activation function; This represents the Sigmoid activation function. Represents element-wise product; Indicates the package or time window number. Indicates the instance number within the package. This indicates the category number of the attack intent.

[0065] The original attention scores are Softmax normalized to obtain the category-related attention weights. Then, fuzzy confidence of the instances. Category-related attention weights The final instance aggregate weights are obtained by merging the aggregate weights.

[0066] ;

[0067] ;

[0068] in, Indicates the first The first of the packages The first instance is for the first The final instance aggregation weight of the attack intent. It is a numerically stable term.

[0069] Based on the final instance aggregation weights, the instance features within the same bag are weighted and summed to obtain a category-related time window-level aggregated representation:

[0070] ;

[0071] The attack intent label is output based on the time window-level aggregation representation. Instances with high attention weights are designated as key communication instances, and fuzzy rules with high rule activation strength are designated as key fuzzy rules. At the same time, the rule activation results, instance fuzzy confidence, and final instance aggregation weight are output as the basis for interpreting the attack intent determination results.

[0072] This application also provides a specific embodiment that comprehensively analyzes the effectiveness and superiority of the method provided above; the following settings are made:

[0073] To verify the effectiveness of this invention in determining interpretable attack intent under packet-level labeling and instance-level unlabeling conditions, this embodiment uses a publicly available multi-instance learning benchmark dataset to experimentally verify the multi-instance classification capability and interpretable rule modeling capability of the method. The publicly available dataset is used to simulate a weakly supervised data organization format in an industrial internet edge gateway scenario, where "one preset time window corresponds to one packet, the packet contains multiple unlabeled instances, and only has packet-level labels," and does not limit the invention to being applied only to the publicly available dataset. The dataset information used in this embodiment is shown in Table 1:

[0074] Table 1 Dataset Information

[0075] Dataset Sample size Instance count Feature number in the instance Number of categories ElephantFoxTigerMusk1Musk2 10010010092102 2-132-132-134766598 230230230166166 22222

[0076] In the experiment, a five-step grid search was used to determine the model parameters and conduct the experiment. (Refer to...) Figure 2 In practical industrial internet edge gateway applications, network traffic, control commands, device status, and security alarm data are preprocessed and feature extracted to form a packet-level dataset containing multiple instances. In the experimental verification of this embodiment, each sample in the public multi-instance learning dataset is regarded as a packet, and each instance in the sample is regarded as an instance within the packet, used to simulate the above-mentioned packet-level weakly labeled data organization. Subsequently, a model optimization process with grid optimization as the core is entered. In each iteration, the attack intent judgment model is trained and the classification performance under the current parameter combination is evaluated. If the current parameter combination does not reach the optimal evaluation result, the optimization continues within the preset parameter range. If the current optimal evaluation result is reached, the optimal parameter combination is updated, and it is determined whether the termination condition is met. If the termination condition is not met, the optimization range is adjusted and the search continues until the termination condition is met. Finally, the determined optimal parameters are used to complete model training and testing.

[0077] The specific experimental steps are as follows: Figure 3 As shown, it includes:

[0078] (1) Data preprocessing. For actual industrial internet edge gateway data, network traffic, control commands, device status and security alarm data within a preset time window are acquired, and protocol parsing, time alignment, outlier handling, missing value completion and normalization are performed; for the public multi-instance learning dataset used in this embodiment, its packet-level samples, instance features and packet-level labels are directly read, and training packet-level data and testing packet-level data are formed according to the data organization method of this invention.

[0079] (2) Obtain training and testing package-level data. The preprocessed and feature-extracted package-level datasets are divided into training and testing package-level data according to a preset ratio. Each package corresponds to a preset time window and has a time window-level attack label, while the instances in the package do not have instance-level labels. The training package-level data is used for the construction, parameter learning, and grid optimization of the interpretable attack intent determination model, while the testing package-level data is used to verify the attack intent determination performance and interpretation results of the trained model.

[0080] (3) Model training and parameter adjustment range. The interpretable attack intent determination model is trained based on the training packet-level data, and the optimal parameter combination of the model is determined by a grid search method. Specifically, within the preset parameter range, the following parameter combinations are selected: low-dimensional feature projection dimension, number of fuzzy rules, fuzzy weight index, learning rate, batch size, regularization coefficient, gated attention hidden dimension, and FCM clustering termination condition. For each parameter combination, low-dimensional feature projection, FCM clustering, TSK fuzzy rule construction, instance fuzzy confidence calculation, gated attention aggregation, and joint optimization training are performed sequentially, and the model evaluation index is calculated based on the training results or validation results. The optimal parameter combination is determined by comparing the evaluation index under different parameter combinations. In the above dataset, the relevant parameters of the algorithm during optimization are shown in Table 2:

[0081] Table 2 Relevant Parameters

[0082] parameter value parameter value Low-dimensional feature projection dimension {8, 16, 24} Learning rate parameter {0.05, 0.1, 0.15} Sparse regularization coefficients {0.001, 0.01, 0.1} Attention smoothing regularization coefficient {0.001, 0.01, 0.1} Batch size {16,32,64} Number of cluster centers / Number of rules {2,4，6,…,16,18,20} FCM Termination Iteration Number 100 FCM iteration termination threshold 1e-4

[0083] (4) Retrain the model by using the optimal hyperparameters found by grid optimization.

[0084] (5) Model testing: Input the test package-level data into the trained model, observe the test results and record them.

[0085] (6) Evaluation indicators: Determine the evaluation indicators for this experiment and calculate the values ​​of multiple evaluation indicators corresponding to the test results output by the model. The evaluation indicator used in this experiment is accuracy.

[0086] This embodiment compares the method of the present invention with related multi-instance learning methods; the comparison algorithms include: ABMIL, GAMIL, DSMIL, MAMIL, and AttriMIL. As shown in Table 3, the classification evaluation index results of each algorithm after five grid search validations on five public multi-instance learning datasets are recorded. The classification evaluation index is accuracy, and the results in Table 3 are the mean and standard deviation of the accuracy; for ease of observation, all results are retained to four decimal places.

[0087] Table 3. Results of Classification Evaluation Indicators

[0088] Comparison Algorithm Dataset Elephant Tiger Fox Musk1 Musk2 ABMIL Accuracy ± Standard Deviation 0.8484±0.0143 0.6064±0.0601 0.8105±0.0312 0.8844±0.0221 0.8221±0.0179 GAMIL Accuracy ± Standard Deviation 0.8686±0.0221 0.8434±0.0184 0.6355±0.0132 0.9000±0.0503 0.8638±0.0423 DSMIL Accuracy ± Standard Deviation 0.8904±0.0186 0.8518±0.0151 0.6328±0.0082 0.9333±0.0737 0.8147±0.0611 MAMIL Accuracy ± Standard Deviation 0.8904±0.0186 0.8493±0.0161 0.6328±0.0082 0.8278±0.0671 0.8147±0.0611 AttriMIL Accuracy ± Standard Deviation 0.8937±0.0461 0.8500±0.0605 0.6562±0.0814 0.9019±0.0848 0.8404±0.0449 This invention Accuracy ± Standard Deviation 0.9250±0.0318 0.8750±0.0342 0.6937±0.0537 0.9571±0.0571 0.8897±0.0433

[0089] Based on comprehensive comparison metrics, the method of this invention exhibits better test performance among all compared methods, demonstrating its effectiveness. Furthermore, considering the interpretability of the method, the use of fuzzy clustering in solving for the rule antecedent centers results in more stable antecedent centers. This not only improves the standard performance of the fuzzy system but also enhances stable interpretability, thereby improving the model's interpretability. Additionally, by considering the importance selection for rules, the fuzzy rules are simplified while maintaining good test performance, further improving interpretability. The average number of rules in the experiment is shown in Table 4.

[0090] Table 4 Average number of rules in this invention

[0091] Datasets Elephant Fox Tiger Musk1 Musk2 This invention 8 10 8 6 5

[0092] The interpretability of the construction method of this invention is higher than that of all the comparison algorithms in most cases. This is because this invention adds fuzzy rule selection items and obtains stable rule antecedent centers by fuzzy ensemble clustering.

[0093] In summary, the method of this invention not only achieves good performance in determining attack intent but also provides interpretable evidence based on fuzzy rules, rule activation results, instance fuzzy confidence, and gating attention aggregation weights. Furthermore, the determination performance and interpretability of this method are related to the number of fuzzy rules. Fewer fuzzy rules generally result in a simpler rule structure, making it easier for security operations personnel to understand. Appropriately increasing the number of fuzzy rules allows the model to depict more granular industrial communication behavior patterns and attack patterns, but also increases the interpretability complexity. Therefore, in practical industrial internet edge gateway security operations scenarios, the number of fuzzy rules can be flexibly adjusted according to requirements such as attack intent determination accuracy, interpretability granularity, and operational response efficiency to achieve a balance between determination performance and interpretability.

[0094] The embodiments of the present invention have been described in detail above with reference to the examples. However, the present invention is not limited to the above embodiments. For those skilled in the art, after learning the contents described in the present invention, several equivalent changes and substitutions can be made without departing from the principle of the present invention. These equivalent changes and substitutions should also be considered to fall within the protection scope of the present invention.

Claims

1. A method for determining interpretable attack intent for industrial internet edge gateways, characterized in that: Includes the following steps: Establish a packet-level dataset of instance representations with time window-level attack labels for edge gateways, and obtain low-dimensional instance feature representations after preprocessing; The fuzzy C-means algorithm is used to cluster the feature representations of low-dimensional instances, and the antecedents of IF-THEN type fuzzy rules in the interpretable knowledge base are constructed based on the clustering results. A TSK fuzzy system is built. Based on the aforementioned antecedent, the rule activation results and instance fuzzy confidence of the low-dimensional instance feature representation are calculated. Packet-level attack labels are used as supervision signals, and instance fuzzy confidence is introduced as a semantic prior into a gating attention aggregation mechanism to construct a coupled joint optimization objective function. The joint optimization objective function is optimized to identify the fuzzy rule consequent parameters and update the gradient of the gating attention weights, forming an interpretable knowledge base and attack intent determination model. The features of each instance in the packet-level dataset are used as input to the knowledge base and attack intent determination model. The output is the attack intent label of the corresponding time window. Based on the final instance, the weights and rule activation strengths are aggregated, and the key instances, key rules, rule activation values ​​and instance confidence are backtracked and output to form the basis for decision interpretability.

2. The method for determining interpretable attack intent for industrial internet edge gateways according to claim 1, characterized in that: Feature extraction is performed on network traffic, control commands, device status, and security alarm data collected by the edge gateway within a preset time window. Each data item in each packet is converted into instance features to construct a packet-level dataset based on instance feature representation. Each packet has a time window-level attack label, while each instance in the packet does not have an instance-level label.

3. The method for determining interpretable attack intent for industrial internet edge gateways according to claim 1, characterized in that: The preprocessing includes: performing linear transformation, nonlinear activation, and normalization on the instance features in the instance representation packet-level dataset, projecting the high-dimensional industrial communication features into low-dimensional instance feature representations.

4. The method for determining interpretable attack intent for industrial internet edge gateways according to claim 3, characterized in that: The fuzzy C-means algorithm takes low-dimensional instance feature representation as input to obtain K cluster centers and membership matrices; and updates the cluster centers and membership matrices through alternating optimization until a preset convergence condition is met to obtain optimized membership matrices and cluster centers; the optimized cluster centers serve as the centers of the antecedents of IF-THEN type fuzzy rules.

5. The method for determining interpretable attack intent for industrial internet edge gateways according to claim 1, characterized in that: The TSK fuzzy system uses a Gaussian membership function to calculate the degree of matching of low-dimensional instance feature representations with respect to each fuzzy rule, and obtains the rule activation result based on the degree of matching.

6. The method for determining interpretable attack intent for industrial internet edge gateways according to claim 5, characterized in that: The k-th IF-THEN type fuzzy rule is represented as follows: ； in, This represents the feature representation of a low-dimensional instance. This represents the dimension of the feature representation of a low-dimensional instance. Indicates the first Rule # The fuzzy set corresponding to each feature and They represent the first The consequent weight parameters and bias parameters of the rule. This represents the linear output of the k-th fuzzy rule.

7. The method for determining interpretable attack intent for industrial internet edge gateways according to claim 3, characterized in that: The gated attention aggregation mechanism includes a two-branch gating structure. The first branch uses the tanh activation function to extract the relevance representation of instance features, and the second branch uses the sigmoid activation function to form gating weights to suppress invalid or noisy instance features. The outputs of the first and second branches are fused element-wise and mapped to attention scores. After being normalized by Softmax, category-related attention weights are obtained. Then, the instance fuzzy confidence is fused with the category-related attention weights to obtain the final instance aggregation weights.

8. The method for determining interpretable attack intent for industrial internet edge gateways according to claim 1, characterized in that: Aggregate weights based on the final instance We perform a weighted summation of instance features within the same package to obtain a category-related time-window aggregated representation: ； in, This represents the intermediate feature representation obtained after the first nonlinear transformation; based on the time window-level aggregation representation, an attack intent label is output, and instances with higher attention weights are taken as key communication instances, and fuzzy rules with higher rule activation strength are taken as key fuzzy rules. At the same time, the rule activation result, instance fuzzy confidence, and final instance aggregation weight are output as the basis for interpreting the attack intent determination result.

Images