Nuclear power instrument control system network attack detection method and system based on adversarial migration

Abstract

The application discloses a kind of nuclear power instrument control system network attack detection method and system based on confrontation migration, other industrial control field disclosed network attack data set is selected to construct source domain data set, and target domain data set is constructed by nuclear power instrument control system semi-physical simulation;Source domain feature extractor and source domain classifier are constructed, and it is trained using source domain training data set, and the model saved after training;Adaptive model of confrontation domain is constructed, and confrontation training is completed using target domain data set, unknown attack classifier is also constructed, and it is cascaded into nuclear power instrument control system network attack detection model with the saved target domain feature extractor;The detection performance of nuclear power instrument control system network attack detection model is tested and verified using target domain test set.The application uses the idea of confrontation domain adaptation, realizes the migration reuse of source domain attack knowledge in target domain, solves the problem that nuclear power instrument control system network attack data is few and characteristic is deficient, and shows superior detection performance.

Description

Technical Field

[0001] This invention belongs to the field of network attack detection technology for nuclear power plant instrumentation and control systems, specifically relating to a method and system for detecting network attacks on nuclear power plant instrumentation and control systems based on adversarial migration. Background Technology

[0002] Instrumentation and control systems (ICS) are the "central nervous system" of nuclear power plants, crucial for their safe, reliable, and economical operation. Against the backdrop of energy independence and national security, nuclear power ICS are transitioning from early analog systems to digital systems. This will significantly enhance the competitiveness of nuclear power plants in terms of performance, data analysis capabilities, edge computing power, and cost. However, this also exposes them to increasingly serious cybersecurity challenges.

[0003] Nuclear power plant digital instrumentation and control systems typically employ a complex layered architecture. This multi-layered architecture increases the probability of cyberattacks on the system and also makes attack methods increasingly diverse. Statistics show that since 1990, there have been more than 20 cyberattacks targeting nuclear facilities worldwide. This demonstrates that with the development of digitalization and network information technology in nuclear power plant instrumentation and control systems, cyberattack methods are constantly evolving, posing significant challenges to network security protection. Currently, the core equipment most frequently attacked in nuclear power plant instrumentation and control systems is located at the field control layer. This field control layer generally uses an internal local area network (LAN), which is physically isolated from external communication systems. This means that most current cyberattacks begin with phishing emails or mobile devices. In response to the rapidly evolving and diverse network intrusion methods, scholars both domestically and internationally have conducted research on various methods for detecting cyberattacks on nuclear power plant instrumentation and control systems. Among these methods, deep learning-based cyberattack detection methods stand out, as they can learn from large amounts of traffic data to extract numerous effective features. Based on this characteristic of deep learning, combining it with transfer learning can leverage comprehensive cyberattack knowledge from other fields to help address the problem of scarce data and limited features in the nuclear power plant instrumentation and control system field. Therefore, researching cyberattack detection methods for nuclear power plant instrumentation and control systems based on deep transfer learning is of great significance for their safe and stable operation. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method and system for detecting network attacks on nuclear power plant instrumentation and control systems based on adversarial migration, in order to address the technical problem of scarce network attack data and lack of features in the field of nuclear power plant instrumentation and control systems.

[0005] The present invention adopts the following technical solution:

[0006] A method for detecting network attacks on nuclear power plant instrumentation and control systems based on adversarial migration includes the following steps:

[0007] S1. Select the publicly available network attack datasets UNSW-NB15 and CICIDS2017 in the IT field, and the publicly available network attack dataset Electra in the industrial control field to construct a source domain sample database.

[0008] S2. Collect network traffic from the hardware-in-the-loop simulation platform of the nuclear power plant instrumentation and control system under various network attack scenarios, and construct a target domain sample database;

[0009] S3. Preprocess and standardize the source domain sample database obtained in step S1 and the target domain sample database obtained in step S2, and divide the source domain / target domain training dataset and test dataset.

[0010] S4. Construct a global feature extractor based on a self-attention mechanism and a local feature extractor based on a convolutional neural network. Merge the output features of the global and local feature extractors to form a source domain feature extractor. Then, construct a source domain classifier through multiple fully connected layers. Finally, cascade the source domain feature extractor and the source domain classifier into an OSSNformer model for the detection of source domain network attacks.

[0011] S5. Train the OSSNformer model constructed in step S4 using the preprocessed and standardized source domain training data from step S3, and save the trained source domain feature extractor and source domain classifier.

[0012] S6. Construct a target domain feature extractor based on a state space model and a domain discriminator composed of multiple fully connected layers, and use the domain discriminator, the source domain feature extractor and the source domain classifier obtained in step S5 to construct an adversarial transfer architecture.

[0013] S7. Use the preprocessed and standardized target domain training data from step S3 to perform adversarial training on the adversarial domain adaptation architecture constructed in step S6, and save the trained target domain feature extractor.

[0014] S8. Construct a class clusterer to reduce the intra-class distribution distance of attack data. Use K-means clustering to obtain the class center vector of each attack type and the class radius threshold vector of each attack type to determine whether the current data is an unknown attack. Use the class clusterer, class center vector, class radius threshold vector and source domain classifier obtained in step S5 to construct an unknown attack classifier. Concatenate the unknown attack classifier with the target domain feature extractor obtained in step S7 to form a network attack detection model for nuclear power instrumentation and control system.

[0015] S9. Input the preprocessed and standardized target domain test dataset from step S3 into the nuclear power plant instrumentation and control system network attack detection model obtained in step S8, classify various network attack data including unknown attacks, and realize the detection of network attacks on the nuclear power plant instrumentation and control system.

[0016] Preferably, the attack types in the source domain sample database and the target domain sample database include denial-of-service attacks, distributed denial-of-service attacks, man-in-the-middle attacks, backdoor attacks, replay attacks, and data tampering; the samples in the source domain sample database and the target domain sample database are time series composed of network flows within a fixed time period, and each series corresponds to a network attack type label.

[0017] Preferably, in step S4, the OSSNformer model includes an input layer, a feature extraction layer, a classifier, and an output layer; the input layer includes a location encoding and spatial feature normalization module; the feature extraction layer is an encoder-decoder architecture, the encoder includes a global feature extractor and a local feature extractor, both the global feature extractor and the decoder in the encoder are built based on several self-attention layers, and the local feature extractor is built based on full-size convolution; the encoder output encoding is composed of global features and local features concatenated, and the decoder takes the output encoding as input and decodes and outputs the final spatiotemporal features.

[0018] More preferably, the input layer receives the source domain network traffic time series with multidimensional characteristics after normalization processing;

[0019] Feature extraction layer: An encoder-decoder architecture is adopted. In the decoder, the self-attention module extracts global spatiotemporal features, and the full-size block composed of convolutional layers extracts local spatiotemporal features. The global spatiotemporal features and the local spatiotemporal features extracted by the full-size block are concatenated to form the output of the decoder. The decoder decodes the source domain network traffic time series through output encoding and finally outputs the extracted spatiotemporal features.

[0020] The classifier consists of multiple fully connected layers. The spatiotemporal features output by the feature extraction layer are used as input, and the output value is a probability vector of the same length as the number of attack types.

[0021] Output layer: The probability vector output by the classifier is taken as input and Softmax activation is performed. The label at the position corresponding to the maximum probability is output as the attack type detected.

[0022] More preferably, the operation process of the feature extraction layer is as follows:

[0023] x in =Snorm(PE(x))

[0024]

[0025]

[0026] out Encoder =Cat(out) EncoderAtten ,out OSCNN )

[0027]

[0028] Where, x in The time series is processed by the input layer position encoding and spatial normalization module; The feature dimension corresponding to the time series; out EncoderAtten and out OSCNN These are the outputs of the global feature extraction layer and the local feature extraction layer, respectively. b represents the weights of the full-size convolutional kernel in the local feature extractor. conv,1D The bias of the convolutional layer in the local feature extractor; x Decoderout The output after decoding by the decoder, out Encoder These are the global and local features output by the encoder.

[0029] Preferably, in step S6, the adversarial domain adaptation architecture includes a source domain feature extractor, a target domain feature extractor, a domain discriminator, and a source domain classifier; the source domain feature extractor and the source domain classifier are the feature extraction layer and classifier in the saved OSSNformer model, wherein the source domain feature extractor takes the source domain time series as input and outputs effective source domain spatiotemporal features; the target domain feature extractor is a state space transition model, takes the target domain time series as input and outputs the extracted target domain spatiotemporal features; the domain discriminator is used to determine whether the feature comes from the source domain or the target domain; the source domain classifier takes the output of the target domain feature extractor as input and outputs the classification result.

[0030] More preferably, the objective function is as follows:

[0031]

[0032] in, and Let these represent the expectations in the source domain and the target domain, respectively. Let y be the cross-entropy loss function corresponding to the classifier. t For the labels corresponding to the data, θ G For generator parameters, θ D For the domain discriminator parameters, D s For the source domain, D(F) s (x) represents the output of the source domain spatiotemporal features after passing through the domain discriminator, D t Let G(x) be the target domain, and D(G(x)) be the output of the spatiotemporal features of the target domain after passing through the domain discriminator.

[0033] Preferably, in step S8, the unknown attack classifier includes a class clusterer, a source domain classifier, a class center vector, and a class radius threshold vector; the class clusterer is constructed based on metric learning, taking the output features of the target domain feature extractor as input, and outputting spatiotemporal features with close intra-class distribution; the source domain classifier is the classifier in the saved OSSNformer model.

[0034] More preferably, the classification process of the unknown attack classifier is as follows:

[0035] x c =Center(out) t )

[0036]

[0037]

[0038] Among them, out t The output of the target domain feature extractor; x c The target domain features are those aggregated by similar aggregators. Let x be the distance of the i-th sample from the class center of label j; B is a function comparing the distance with a class radius threshold. out To output the classification results of network traffic, x i For the i-th sample, For the j-th class center, V cmax The class radius threshold for each class.

[0039] Secondly, embodiments of the present invention provide a network attack detection system for nuclear power plant instrumentation and control systems based on adversarial migration, comprising:

[0040] The first data module selects the publicly available network attack datasets UNSW-NB15 and CICIDS2017 in the IT field, as well as the publicly available network attack dataset Electra in the industrial control field, to construct a source domain sample database.

[0041] The second data module collects network traffic from the hardware-in-the-loop simulation platform of the nuclear power instrumentation and control system under various network attack scenarios, and constructs a target domain sample database.

[0042] The data preprocessing module preprocesses and standardizes the source domain sample database and the target domain sample database, and divides the source domain / target domain training dataset and test dataset.

[0043] The model building module constructs a global feature extractor based on a self-attention mechanism and a local feature extractor based on a convolutional neural network. The output features of the global and local feature extractors are merged to form a source domain feature extractor. Then, a source domain classifier is constructed through multiple fully connected layers. Finally, the source domain feature extractor and the source domain classifier are cascaded into an OSSNformer model for the detection of source domain network attacks.

[0044] The model training module trains the OSSNformer model using preprocessed and standardized source domain training data, and saves the trained source domain feature extractor and source domain classifier.

[0045] The adversarial transfer module constructs a target domain feature extractor based on a state-space model and a domain discriminator composed of multiple fully connected layers. It then uses the domain discriminator, source domain feature extractor, and source domain classifier to construct an adversarial transfer architecture. The adversarial domain adaptation architecture is trained using preprocessed and standardized target domain training data, and the trained target domain feature extractor is saved.

[0046] The clustering module constructs a clusterer of similar types to reduce the intra-class distribution distance of attack data. It uses the K-means clustering method to obtain the class center vector and the class radius threshold vector of each attack type to determine whether the current data is an unknown attack. It uses the clusterer of similar types, the class center vector, the class radius threshold vector and the source domain classifier to construct an unknown attack classifier. The unknown attack classifier is then cascaded with the target domain feature extractor to form a network attack detection model for nuclear power instrumentation and control systems.

[0047] The detection module inputs the preprocessed and standardized target domain test dataset into the nuclear power plant instrumentation and control system network attack detection model, classifies various network attack data including unknown attacks, and realizes the detection of network attacks on the nuclear power plant instrumentation and control system.

[0048] Thirdly, a computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the above-described method for detecting network attacks on nuclear power plant instrumentation and control systems based on adversarial migration.

[0049] Fourthly, embodiments of the present invention provide a computer-readable storage medium including a computer program, which, when executed by a processor, implements the steps of the above-described method for detecting network attacks on nuclear power plant instrumentation and control systems based on adversarial migration.

[0050] Compared with the prior art, the present invention has at least the following beneficial effects:

[0051] A method for detecting cyberattacks on nuclear power plant instrumentation and control systems based on adversarial transfer is proposed. This method constructs a source domain dataset using publicly available cyberattack datasets from other industrial control fields, and a target domain dataset using hardware-in-the-loop simulation. A source domain feature extractor and a source domain classifier are built and trained using the source domain training dataset, and the trained model is saved. An adversarial domain adaptation model is then constructed, trained using the target domain dataset, and an unknown attack classifier is built. This classifier is then cascaded with the saved target domain feature extractor to form a nuclear power control system cyberattack detection network (NCS-CADN). The detection performance of NCS-CADN is tested and verified using a target domain test set. This method employs the idea of ​​adversarial domain adaptation, realizing the transfer and reuse of attack knowledge from the source domain to the target domain. It solves the problem of scarce and feature-deficient cyberattack data for nuclear power plant instrumentation and control systems, demonstrating superior detection performance. This method is the first to propose constructing a nuclear power plant instrumentation and control system cyberattack detection model using adversarial domain adaptation, which can accurately detect various forms of cyberattacks targeting nuclear power plant instrumentation and control systems.

[0052] Furthermore, considering that nuclear power plant instrumentation and control systems may face unknown attacks, an unknown attack classifier is constructed based on the known attack classifier to achieve the detection of unknown attacks while ensuring the detection accuracy of known attacks.

[0053] Furthermore, preprocessing and normalizing the sample data eliminates the dimensional differences between various features in the data, which helps improve the speed of the cross-entropy loss function in finding the optimal solution using gradient descent.

[0054] Furthermore, a source domain feature extractor is constructed from both global and local perspectives to ensure the comprehensiveness of the extracted spatiotemporal features. Simultaneously, a spatial feature annotation module is used to refine the features in the spatial dimension of the data, improving feature quality.

[0055] Furthermore, setting the source domain network attack detection model appropriately and selecting the optimal hyperparameters can help improve detection accuracy. Moreover, the source domain classifier can integrate and refine the extracted features to achieve accurate detection of network attacks.

[0056] Furthermore, a source domain classifier is introduced during the adversarial migration process. By utilizing the label information of the source domain data, the features extracted by the target domain feature extractor become more distinguishable between classes, effectively improving the detection accuracy of network attacks on nuclear power plant instrumentation and control systems.

[0057] It is understandable that the beneficial effects of the second aspect mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0058] In summary, this invention can accurately detect network attacks on nuclear power plant instrumentation and control systems, including unknown attacks.

[0059] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0060] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0061] Figure 2 Architecture diagram for network attack experiment of hardware-in-the-loop simulation platform for nuclear power digital instrumentation and control system;

[0062] Figure 3 Information graph of the source domain dataset;

[0063] Figure 4 Infographic of the target domain dataset;

[0064] Figure 5 This is the architecture diagram of OSSNformer;

[0065] Figure 6 The diagram shows the training process of OSSNformer, where (a) represents the accuracy and (b) represents the loss curve.

[0066] Figure 7 To combat migration architecture diagram;

[0067] Figure 8 Architecture diagram of an unknown attack classifier;

[0068] Figure 9 The confusion matrix results for NCS-CADN in the target domain test set are shown in the figure.

[0069] Figure 10 This is a graph showing the detection accuracy of NCS-CADN for each type of attack in the target domain test set.

[0070] Figure 11 A schematic diagram of a computer device provided in an embodiment of the present invention;

[0071] Figure 12 This is a block diagram of a chip provided according to an embodiment of the present invention. Detailed Implementation

[0072] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] In the description of this invention, it should be understood that the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0074] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0075] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Additionally, the character " / " in this invention generally indicates that the preceding and following objects have an "or" relationship.

[0076] It should be understood that although terms such as first, second, third, etc., may be used in the embodiments of the present invention to describe the preset range, these preset ranges should not be limited to these terms. These terms are only used to distinguish the preset ranges from one another. For example, without departing from the scope of the embodiments of the present invention, the first preset range may also be referred to as the second preset range, and similarly, the second preset range may also be referred to as the first preset range.

[0077] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0078] The accompanying drawings illustrate various structural schematic diagrams according to embodiments disclosed in this invention. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0079] This invention provides a network attack detection method for nuclear power plant instrumentation and control systems based on adversarial migration. It constructs a source domain feature extractor to extract effective features from the source domain dataset, and then builds a target domain feature extractor to perform adversarial domain adaptation training. Leveraging the abundant attack knowledge in publicly available datasets from traditional IT and other industrial control fields, it helps address the problem of scarce network attack data and feature deficiencies in the nuclear power plant instrumentation and control system domain. Simultaneously, it constructs an unknown attack classifier, demonstrating advanced detection performance for unknown attacks occurring in nuclear power plant instrumentation and control systems.

[0080] Please see Figure 1 This invention discloses a network attack detection method for nuclear power plant instrumentation and control systems based on adversarial migration, comprising the following steps:

[0081] S1. Select the publicly available network attack datasets UNSW-NB15 and CICIDS2017 from the traditional IT field and the publicly available network attack dataset Electra from the industrial control security field to construct the source domain sample database;

[0082] The three publicly available datasets were obtained in real-world cyberattack environments and include various common cyberattack types in nuclear power plant instrumentation and control systems, such as DoS attacks, man-in-the-middle attacks, replay attacks, and erroneous data injection. These datasets provide a rich reserve of attack knowledge for the detection of cyberattacks in nuclear power plant instrumentation and control systems and lay a solid foundation for the subsequent process of transferring cyberattack knowledge to countermeasures.

[0083] S2. Collect network traffic from the hardware-in-the-loop simulation platform of nuclear power instrumentation and control system under various network attack scenarios to construct a target domain sample database;

[0084] Please see Figure 2 Based on the developed hardware-in-the-loop simulation platform for nuclear power digital instrumentation and control systems, various network attack experiments were conducted to obtain data that reflects the actual operating status of nuclear power plant instrumentation and control systems. Six network attack scenarios were set up, including DoS attack, DDoS attack, man-in-the-middle attack, backdoor attack, replay attack, and data tampering. The network traffic of the platform under each attack scenario was captured using the Wireshark network traffic capture tool to construct a target domain sample database.

[0085] S3. Preprocess and standardize the source domain / target domain sample database in steps S1 and S2. Preprocessing involves integrating and processing the data and making it into a format supported by the corresponding training model. Then, divide the source domain / target domain training dataset and test dataset into an 8:2 ratio.

[0086] The normalization method is as follows:

[0087]

[0088] in, It is the value corresponding to the j-th feature at the i-th position in the sequence, max(x) j ) is the maximum value of the j-th feature in the entire sequence, min(x) j ) is the minimum value of the j-th feature in the entire sequence.

[0089] After normalization, the influence of the differences in the dimensions of the parameters is eliminated, which can speed up the gradient descent to find the optimal solution during network training and is helpful for training; the structure of the data does not change after normalization.

[0090] Please see Figure 3 and Figure 4 The figure shows the detailed information of the source and target domain datasets after proportional division. The source domain time series length is set to 1000, and the target domain time series length is set to 32. Compared with the source domain data, the target domain data volume is significantly reduced, which matches the phenomenon of scarce and lacking features in actual nuclear power plant instrumentation and control system network attack data.

[0091] S4. Construct a global feature extractor based on a self-attention mechanism and a local feature extractor based on a convolutional neural network, and merge the output features of the two to form a source domain feature extractor; then construct a source domain classifier through multiple fully connected layers, and finally cascade the source domain feature extractor and the source domain classifier into an OSSNformer model for the detection of source domain network attacks.

[0092] Please see Figure 5 The OSSNformer model consists of an input layer, a feature extraction layer, a classifier, and an output layer. All initial network weights and biases in the OSSNformer are randomly generated, follow a normal distribution, and are continuously optimized during network training.

[0093] The OSSNformer model is as follows:

[0094] Input layer: Receives the source network traffic time series with multidimensional characteristics after normalization processing;

[0095] Feature extraction layer: Composed of encoder-decoder architecture. In the decoder, the self-attention module extracts global spatiotemporal features, and the full-size convolutional layer extracts local spatiotemporal features. The two are concatenated to form the output of the decoder. The decoder decodes the source domain network traffic time series through output encoding and finally outputs the extracted spatiotemporal features.

[0096] The classifier consists of multiple fully connected layers. It takes the spatiotemporal features output by the feature extraction layer as input and outputs a probability vector of the same length as the number of attack types.

[0097] Output layer: The probability vector output by the classifier is taken as input and activated by Softmax. The label at the position of the maximum probability is output as the type of attack detected.

[0098] The feature extraction layer follows an encoder-decoder architecture. The encoder consists of a global feature extractor and local feature extractors. Both the global and local feature extractors are built upon several self-attention layers, while the local feature extractors are built upon full-size convolutional layers (OSCNN). The encoder outputs encoded data. Encoder The spatiotemporal feature (x) is composed of global and local features. The decoder takes the output code as input and decodes to output the final spatiotemporal feature (x). f The settings are as follows:

[0099] Self-attention layer: The number of layers in the global feature extractor is set to 2, the number of layers in the decoder is set to 1, and the number of attention heads is set to 16;

[0100] Local feature extraction layer: The kernel size is set to all prime numbers in the range of 1 to 89, the number of input channels is set to 76, and the number of output channels is set to 76;

[0101] Enter x in The output x is given by the position encoding and spatial feature normalization module, and the final output x is determined by the decoder. out The operation process of the feature extraction layer is represented as follows:

[0102] x in =Snorm(PE(x))

[0103]

[0104]

[0105] out Encoder =Cat(out) EncoderAtten ,out OSCNN )

[0106]

[0107] Where, x in The time series is processed by the input layer position encoding and spatial normalization module; The feature dimension corresponding to the time series; out EncoderAtten and out OSCNN These are the outputs of the global feature extraction layer and the local feature extraction layer, respectively. b represents the weights of the full-size convolutional kernel in the local feature extractor. conv,1DThe bias of the convolutional layer in the local feature extractor; x Decoderout This is the output after decoding by the decoder.

[0108] S5. Use the preprocessed and standardized source domain training data from step S3 to train the OSSNformer constructed in step S4, and save the trained source domain feature extractor and source domain classifier.

[0109] Input the source domain training data into OSSNformer for training. Since deep neural networks need to continuously extract features from a large number of samples and optimize the weight parameters in the network to achieve the ideal prediction effect, it is recommended that the amount of training sample data reach tens of thousands for each attack; the number of training rounds of the network model is set to 100 rounds, and the learning rate decay value is set to 0.5 every 50 iterations during training.

[0110] Cross-validation was used to validate the OSSNformer, specifically by referring to the trends in accuracy and error during its training. The closer the final accuracy is to 100% and the closer the error is to 0, the better the performance. Figure 6 As shown in the figure. The cross-entropy loss function was chosen during training, which performs well for multi-class classification tasks. The trained OSSNformer achieved 99.87% accuracy on the source domain test set, providing a guarantee for the transfer of knowledge from subsequent source domain attacks.

[0111] S6. Construct a target domain feature extractor based on a state space model and a domain discriminator composed of multiple fully connected layers. Use it together with the source domain feature extractor and source domain classifier saved in step S5 to build an adversarial domain adaptation architecture to realize the transfer of source domain knowledge to the target domain for reuse.

[0112] Please see Figure 7 The adversarial domain adaptation architecture includes a source domain feature extractor, a target domain feature extractor, a domain discriminator, and a source domain classifier.

[0113] Source Domain Feature Extractor (F s ):

[0114] The feature extraction layer in the saved OSSNformer model takes the source domain time series as input and outputs effective source domain spatiotemporal features.

[0115] Source Domain Classifier (C s ):

[0116] To classify the classifier in the saved OSSNformer model, the output of the target domain feature extractor is used as input, and the classification result is output.

[0117] Target domain feature extractor (G):

[0118] It is constructed by a state-space transition model, taking the target domain time series as input and outputting the extracted target domain spatiotemporal features;

[0119] Domain discriminator (D):

[0120] Used to determine whether a feature comes from the source domain or the target domain.

[0121] In the process of adversarial domain adaptation, the alignment of real and fake data distributions in traditional generative adversarial networks (GANs) transforms into the alignment of source domain features and target domain features. The adversarial relationship between the generator and discriminator evolves into an adversarial relationship between the target domain feature extractor and the feature domain discriminator. This adversarial approach effectively reduces the distributional differences between the source and target domains, enabling the transfer and reuse of source domain knowledge in the target domain. The entire process can be specifically described as an optimization process of the objective function, as follows:

[0122]

[0123]

[0124] Where k is the number of label categories in the target domain data; C s (G(x t )) j The source domain classifier will classify the target domain data x t The probability of predicting it as class j; and These represent the expectations in the source and target domains, respectively.

[0125] S7. Use the preprocessed and standardized target domain training data from step S3 to perform adversarial training on the adversarial domain adaptation architecture constructed in step S6, and save the trained target domain feature extractor.

[0126] Source domain data is input into the source domain feature extractor, and target domain training data is input into the target domain feature extractor for adversarial training. During the adversarial process, the domain discriminator strives to improve its ability to distinguish between source and target domain features, while the target domain feature extractor continuously improves its feature extraction capabilities, making its extracted features as indistinguishable as possible from those of the domain discriminator. Under these conditions, the two continuously compete and update the network parameters until the loss approaches equilibrium. The adversarial training rounds are set to 1500, and the learning rate decays to 0.5 every 300 iterations during training.

[0127] S8. Construct a homogeneous aggregator to reduce the intra-class distribution distance of attack data. Use K-means clustering to obtain the class center vector for each attack type and the class radius threshold vector for each attack type to determine whether the current data is an unknown attack. Construct an unknown attack classifier using the homogeneous aggregator, class center vector, class radius threshold vector, and the source domain classifier saved in step S5. Concatenate this classifier with the target domain feature extractor saved in step S7 to form the NCS-CADN (Nuclear Power Equipment Control System Network Attack Detection Model).

[0128] Please see Figure 8 Unknown attack classifiers include a class aggregator (Center) and a source domain classifier (Center). s ), class center vector (V c ) and class radius threshold vector (V cmax ).

[0129] Homogeneous clusterers: These are built based on metric learning, taking the output features of the target domain feature extractor as input, and outputting spatiotemporal features that are closely distributed within the same class.

[0130] Source domain classifier: The classifier in the saved OSSNformer model takes the output of the target domain feature extractor as input and outputs the classification result.

[0131] The classifier input is the output of the target domain feature extractor. t Determine and output the classification result of the network traffic x. out .

[0132] The classification process of the unknown attack classifier is represented as follows:

[0133] x c =Center(out) t )

[0134]

[0135]

[0136] Among them, out t The output of the target domain feature extractor; x c The target domain features are those aggregated by similar aggregators. is the distance of the i-th sample from the class center of label j; B is a function that compares the distance with the class radius threshold, and finally outputs the comparison result.

[0137] The NCS-CADN correspondence is formed by cascading the target domain feature extractor and the unknown attack classifier. Figure 1 The third part of the model enables the detection of network attacks on nuclear power plant instrumentation and control systems.

[0138] S9. Input the standardized target domain test dataset from the preprocessed set in step S3 into the NCS-CADN constructed in step S8 to classify various network attack data, including unknown attacks, and realize the detection of network attacks on nuclear power instrumentation and control systems.

[0139] In another embodiment of the present invention, a nuclear power plant instrumentation and control system network attack detection system based on adversarial migration is provided. This system can be used to implement the above-mentioned nuclear power plant instrumentation and control system network attack detection method based on adversarial migration. Specifically, the nuclear power plant instrumentation and control system network attack detection system based on adversarial migration includes a first data module, a second data module, a data preprocessing module, a model building module, a model training module, an adversarial migration module, a clustering module, and a detection module.

[0140] The first data module selects the publicly available network attack datasets UNSW-NB15 and CICIDS2017 in the IT field, and the publicly available network attack dataset Electra in the industrial control field to construct a source domain sample database.

[0141] The second data module collects network traffic from the hardware-in-the-loop simulation platform of the nuclear power instrumentation and control system under various network attack scenarios, and constructs a target domain sample database.

[0142] The data preprocessing module preprocesses and standardizes the source domain sample database and the target domain sample database, and divides the source domain / target domain training dataset and test dataset.

[0143] The model building module constructs a global feature extractor based on a self-attention mechanism and a local feature extractor based on a convolutional neural network. The output features of the global and local feature extractors are merged to form a source domain feature extractor. Then, a source domain classifier is constructed through multiple fully connected layers. Finally, the source domain feature extractor and the source domain classifier are cascaded into an OSSNformer model for the detection of source domain network attacks.

[0144] The model training module trains the OSSNformer model using preprocessed and standardized source domain training data, and saves the trained source domain feature extractor and source domain classifier.

[0145] The adversarial transfer module constructs a target domain feature extractor based on a state-space model and a domain discriminator composed of multiple fully connected layers. It then uses the domain discriminator, the source domain feature extractor, and the source domain classifier to construct an adversarial transfer architecture.

[0146] The adversarial domain adaptation architecture is trained using preprocessed and standardized target domain training data, and the trained target domain feature extractor is saved.

[0147] The clustering module constructs a clusterer of similar types to reduce the intra-class distribution distance of attack data. It uses the K-means clustering method to obtain the class center vector and the class radius threshold vector of each attack type to determine whether the current data is an unknown attack. It uses the clusterer of similar types, the class center vector, the class radius threshold vector and the source domain classifier to construct an unknown attack classifier. The unknown attack classifier is then cascaded with the target domain feature extractor to form a network attack detection model for nuclear power instrumentation and control systems.

[0148] The detection module inputs the preprocessed and standardized target domain test dataset into the nuclear power plant instrumentation and control system network attack detection model, classifies various network attack data including unknown attacks, and realizes the detection of network attacks on the nuclear power plant instrumentation and control system.

[0149] In another embodiment of the present invention, a terminal device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to implement a corresponding method flow or corresponding function. The processor described in this embodiment can be used in the operation of a network attack detection method for nuclear power plant instrumentation and control systems based on adversarial migration, including:

[0150] We selected publicly available network attack datasets from the IT field, UNSW-NB15 and CICIDS2017, and from the industrial control field, Electra, to construct a source domain sample database. We collected network traffic from a hardware-in-the-loop simulation platform of a nuclear power plant instrumentation and control system under various network attack scenarios to construct a target domain sample database. We preprocessed and standardized both the source and target domain sample databases, and divided them into source / target domain training and test datasets. We constructed a global feature extractor based on a self-attention mechanism and a local feature extractor based on a convolutional neural network. The output features of the global and local feature extractors were merged to form a source domain feature extractor. Then, a source domain classifier was constructed using multiple fully connected layers. Finally, the source domain feature extractor and the source domain classifier were cascaded into an OSSNformer model for detecting source domain network attacks. We trained the constructed OSSNformer model using the preprocessed and standardized source domain training data, and then tested the trained source domain feature extractor and source domain classifier. The process involves: saving the data; constructing a target domain feature extractor based on a state-space model and a domain discriminator composed of multiple fully connected layers; using the domain discriminator, source domain feature extractor, and source domain classifier to construct an adversarial transfer architecture; training the constructed adversarial domain adaptation architecture using preprocessed and standardized target domain training data, and saving the trained target domain feature extractor; constructing a class clusterer to reduce the intra-class distribution distance of attack data; obtaining the class center vector for each attack type and the class radius threshold vector for each attack type using the K-means clustering method to determine whether the current data is an unknown attack; constructing an unknown attack classifier using the class clusterer, class center vector, class radius threshold vector, and source domain classifier; cascading the unknown attack classifier with the target domain feature extractor to form a nuclear power plant instrumentation and control system network attack detection model; inputting the preprocessed and standardized target domain test dataset into the nuclear power plant instrumentation and control system network attack detection model to classify various network attack data, including unknown attacks, and to achieve the detection of nuclear power plant instrumentation and control system network attacks.

[0151] In another embodiment of the present invention, a storage medium is also provided, specifically a computer-readable storage medium (memory). This computer-readable storage medium is a memory device in a terminal device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and extended storage media supported by the terminal device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device.

[0152] One or more instructions stored in a computer-readable storage medium can be loaded and executed by a processor to implement the corresponding steps of the network attack detection method for nuclear power plant instrumentation and control systems based on adversarial migration in the above embodiments; one or more instructions in the computer-readable storage medium are loaded and executed by the processor to perform the following steps:

[0153] We selected publicly available network attack datasets from the IT field, UNSW-NB15 and CICIDS2017, and from the industrial control field, Electra, to construct a source domain sample database. We collected network traffic from a hardware-in-the-loop simulation platform of a nuclear power plant instrumentation and control system under various network attack scenarios to construct a target domain sample database. We preprocessed and standardized both the source and target domain sample databases, and divided them into source / target domain training and test datasets. We constructed a global feature extractor based on a self-attention mechanism and a local feature extractor based on a convolutional neural network. The output features of the global and local feature extractors were merged to form a source domain feature extractor. Then, a source domain classifier was constructed using multiple fully connected layers. Finally, the source domain feature extractor and the source domain classifier were cascaded into an OSSNformer model for detecting source domain network attacks. We trained the constructed OSSNformer model using the preprocessed and standardized source domain training data, and then tested the trained source domain feature extractor and source domain classifier. The process involves: saving the data; constructing a target domain feature extractor based on a state-space model and a domain discriminator composed of multiple fully connected layers; using the domain discriminator, source domain feature extractor, and source domain classifier to construct an adversarial transfer architecture; training the constructed adversarial domain adaptation architecture using preprocessed and standardized target domain training data, and saving the trained target domain feature extractor; constructing a class clusterer to reduce the intra-class distribution distance of attack data; obtaining the class center vector for each attack type and the class radius threshold vector for each attack type using the K-means clustering method to determine whether the current data is an unknown attack; constructing an unknown attack classifier using the class clusterer, class center vector, class radius threshold vector, and source domain classifier; cascading the unknown attack classifier with the target domain feature extractor to form a nuclear power plant instrumentation and control system network attack detection model; inputting the preprocessed and standardized target domain test dataset into the nuclear power plant instrumentation and control system network attack detection model to classify various network attack data, including unknown attacks, and to achieve the detection of nuclear power plant instrumentation and control system network attacks.

[0154] Please see Figure 11 The terminal device is a computer device. In this embodiment, the computer device 60 includes a processor 61, a memory 62, and a computer program 63 stored in the memory 62 and executable on the processor 61. When executed by the processor 61, the computer program 63 implements the fluid composition calculation method in the reservoir stimulation wellbore of this embodiment. To avoid repetition, these details are not elaborated here. Alternatively, when executed by the processor 61, the computer program 63 implements the functions of each model / unit in the fluid composition calculation system in the reservoir stimulation wellbore of this embodiment. To avoid repetition, these details are not elaborated here.

[0155] Computer device 60 can be a desktop computer, laptop, handheld computer, cloud server, or other computing device. Computer device 60 may include, but is not limited to, a processor 61 and a memory 62. Those skilled in the art will understand that... Figure 11 This is merely an example of computer device 60 and does not constitute a limitation on computer device 60. It may include more or fewer components than shown, or combine certain components, or different components. For example, computer device may also include input / output devices, network access devices, buses, etc.

[0156] The processor 61 may be a central processing unit (CPU), or other general-purpose processors, CPUs, graphics processing units (GPUs), digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, quantum computing-based data processing logic units, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0157] The memory 62 can be an internal storage unit of the computer device 60, such as a hard disk or RAM of the computer device 60. The memory 62 can also be an external storage device of the computer device 60, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. equipped on the computer device 60.

[0158] Furthermore, the memory 62 may include both internal storage units of the computer device 60 and external storage devices. The memory 62 is used to store computer programs and other programs and data required by the computer device. The memory 62 can also be used to temporarily store data that has been output or will be output.

[0159] Any references to memory, databases, or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0160] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0161] Please see Figure 12 The terminal device is a chip. In this embodiment, the chip 600 includes a processor 622, which may be one or more, and a memory 632 for storing computer programs executable by the processor 622. The computer program stored in the memory 632 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processor 622 may be configured to execute the computer program to perform the generalizable monocular absolute depth map estimation method described above.

[0162] Additionally, chip 600 may also include a power supply component 626 and a communication component 650. The power supply component 626 can be configured to perform power management of chip 600, and the communication component 650 can be configured to enable communication of chip 600, such as wired or wireless communication. Furthermore, chip 600 may also include an input / output interface 658. Chip 600 can operate on an operating system stored in memory 632.

[0163] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0164] Please see Figure 9 and Figure 10 , Figure 9 The performance of NCS-CADN was tested and verified using a confusion matrix. Each row of the confusion matrix represents the true class, each column represents the class predicted by the model, and each value in the matrix indicates the number of data points in the current result. The confusion matrix shows that NCS-CADN achieved high accuracy on all seven types of attack data. Figure 10 The results show the accuracy of NCS-CADN for each type of attack on the target domain test dataset, which includes unknown attacks. It can be seen that the accuracy for unknown attacks is 87.4%, and the overall accuracy reaches 95.81%, demonstrating excellent attack detection performance.

[0165] In summary, this invention provides a method and system for detecting network attacks on nuclear power plant instrumentation and control systems based on adversarial transfer. It constructs a source domain sample database using publicly available network attack datasets from traditional IT and industrial control fields, and a target domain sample database by collecting network traffic from a hardware-in-the-loop simulation platform of the nuclear power plant instrumentation and control system under various network attack scenarios. The source / target domain data are preprocessed and normalized, and the training and test sets are divided in an 8:2 ratio. A source domain feature extractor is constructed to extract effective features from the source domain dataset, and a target domain feature extractor is then built and trained adversarially. This leverages the rich attack knowledge from publicly available datasets in traditional IT and other industrial control fields to address the problem of scarce network attack data and feature deficiencies in the nuclear power plant instrumentation and control system domain. An unknown attack classifier is constructed based on the source domain classifier, demonstrating good detection performance for unknown attacks occurring in the nuclear power plant instrumentation and control system. The unknown attack classifier is cascaded with the saved target domain feature extractor to form the final detection model, achieving accurate detection of network attacks on the nuclear power plant instrumentation and control system, including unknown attacks.

[0166] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0167] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0168] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0169] In the embodiments provided by this invention, it should be understood that the disclosed devices / terminals and methods can be implemented in other ways. For example, the device / terminal embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0170] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0171] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0172] If the integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0173] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0174] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0175] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0176] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A method for detecting network attacks on nuclear power plant instrumentation and control systems based on adversarial migration, characterized in that, Includes the following steps: S1. Select the publicly available network attack datasets UNSW-NB15 and CICIDS2017 in the IT field, and the publicly available network attack dataset Electra in the industrial control field to construct a source domain sample database. S2. Collect network traffic from the hardware-in-the-loop simulation platform of the nuclear power plant instrumentation and control system under various network attack scenarios, and construct a target domain sample database; S3. Preprocess and standardize the source domain sample database obtained in step S1 and the target domain sample database obtained in step S2, and divide the source domain / target domain training dataset and test dataset. S4. Construct a global feature extractor based on a self-attention mechanism and a local feature extractor based on a convolutional neural network. Merge the output features of the global and local feature extractors to form a source domain feature extractor. Then, construct a source domain classifier through multiple fully connected layers. Finally, cascade the source domain feature extractor and the source domain classifier into an OSSNformer model for the detection of source domain network attacks. The OSSNformer model includes an input layer, a feature extraction layer, a classifier, and an output layer; the input layer includes position encoding and spatial feature normalization modules. The feature extraction layer adopts an encoder-decoder architecture. The encoder includes a global feature extractor and a local feature extractor. Both the global feature extractor and the decoder in the encoder are built based on several self-attention layers, while the local feature extractor is built based on full-size convolution. The encoder output encoding is composed of concatenation of global and local features, and the decoder takes the output encoding as input to decode and output the final spatiotemporal features. S5. Train the OSSNformer model constructed in step S4 using the preprocessed and standardized source domain training data from step S3, and save the trained source domain feature extractor and source domain classifier. S6. Construct a target domain feature extractor based on a state space model and a domain discriminator composed of multiple fully connected layers, and use the domain discriminator, the source domain feature extractor and the source domain classifier obtained in step S5 to construct an adversarial transfer architecture. S7. Use the preprocessed and standardized target domain training data from step S3 to perform adversarial training on the adversarial domain adaptation architecture constructed in step S6, and save the trained target domain feature extractor. S8. Construct a class clusterer to reduce the intra-class distribution distance of attack data. Use K-means clustering to obtain the class center vector of each attack type and the class radius threshold vector of each attack type to determine whether the current data is an unknown attack. Use the class clusterer, class center vector, class radius threshold vector and source domain classifier obtained in step S5 to construct an unknown attack classifier. Concatenate the unknown attack classifier with the target domain feature extractor obtained in step S7 to form a network attack detection model for nuclear power instrumentation and control system. S9. Input the preprocessed and standardized target domain test dataset from step S3 into the nuclear power plant instrumentation and control system network attack detection model obtained in step S8, classify various network attack data including unknown attacks, and realize the detection of network attacks on the nuclear power plant instrumentation and control system.

2. The method for detecting network attacks on nuclear power plant instrumentation and control systems based on adversarial migration as described in claim 1, characterized in that, The attack types in the source domain sample database and the target domain sample database include denial-of-service attacks, distributed denial-of-service attacks, man-in-the-middle attacks, backdoor attacks, replay attacks, and data tampering. The samples in the source domain sample database and the target domain sample database are time series composed of network flows within a fixed time period, and each series corresponds to a network attack type label.

3. The method for detecting network attacks on nuclear power plant instrumentation and control systems based on adversarial migration as described in claim 1, characterized in that, In step S4, the input layer receives the source network traffic time series with multidimensional characteristics after normalization. Feature extraction layer: An encoder-decoder architecture is adopted. In the decoder, the self-attention module extracts global spatiotemporal features, and the full-size block composed of convolutional layers extracts local spatiotemporal features. The global spatiotemporal features and the local spatiotemporal features extracted by the full-size block are concatenated to form the output of the decoder. The decoder decodes the source domain network traffic time series through output encoding and finally outputs the extracted spatiotemporal features. The classifier consists of multiple fully connected layers. The spatiotemporal features output by the feature extraction layer are used as input, and the output value is a probability vector of the same length as the number of attack types. Output layer: The probability vector output by the classifier is taken as input and Softmax activation is performed. The label at the position corresponding to the maximum probability is output as the attack type detected.

4. The method for detecting network attacks on nuclear power plant instrumentation and control systems based on adversarial migration as described in claim 1, characterized in that, In step S4, the operation process of the feature extraction layer is as follows: in, The time series is processed by the input layer position encoding and spatial normalization module; The feature dimension corresponding to the time series; and These are the outputs of the global feature extraction layer and the local feature extraction layer, respectively. The weights of the convolutional kernels with full size in the local feature extractor; The bias of the convolutional layer in the local feature extractor; This is the output after decoding by the decoder. These are the global and local features output by the encoder.

5. The method for detecting network attacks on nuclear power plant instrumentation and control systems based on adversarial migration as described in claim 1, characterized in that, In step S6, the adversarial domain adaptation architecture includes a source domain feature extractor, a target domain feature extractor, a domain discriminator, and a source domain classifier. The source domain feature extractor and the source domain classifier are the feature extraction layer and classifier in the saved OSSNformer model. The source domain feature extractor takes the source domain time series as input and outputs effective source domain spatiotemporal features. The target domain feature extractor is a state-space transition model that takes the target domain time series as input and outputs the extracted spatiotemporal features of the target domain. The domain discriminator is used to determine whether a feature comes from the source domain or the target domain; the source domain classifier takes the output of the target domain feature extractor as input and outputs the classification result.

6. The method for detecting network attacks on nuclear power plant instrumentation and control systems based on adversarial migration as described in claim 5, characterized in that, The objective function is as follows: in, and Let these represent the expectations in the source domain and the target domain, respectively. Let cross-entropy be the loss function corresponding to the classifier. For the data, corresponding labels For generator parameters, For domain discriminator parameters, For the source domain, The output of the source domain spatiotemporal features after passing through the domain discriminator. For the target domain, The output of the target domain spatiotemporal features after passing through the domain discriminator.

7. The method for detecting network attacks on nuclear power plant instrumentation and control systems based on adversarial migration as described in claim 1, characterized in that, In step S8, the unknown attack classifier includes a class clusterer, a source domain classifier, a class center vector, and a class radius threshold vector. The class clusterer is constructed based on metric learning, taking the output features of the target domain feature extractor as input, and outputting spatiotemporal features with close intra-class distribution. The source domain classifier is the classifier in the saved OSSNformer model.

8. The method for detecting network attacks on nuclear power plant instrumentation and control systems based on adversarial migration as described in claim 7, characterized in that, The classification process of the unknown attack classifier is as follows: in, This is the output of the target domain feature extractor; The target domain features are those aggregated by similar aggregators. For the first i Each sample is away from the label j Distance from the center of the class; B A function to compare the distance with a class radius threshold. To output the classification results of network traffic, For the first i One sample, For the first j Individual class center The class radius threshold for each class.

9. A network attack detection system for nuclear power plant instrumentation and control systems based on adversarial migration, characterized in that, include: The first data module selects the publicly available network attack datasets UNSW-NB15 and CICIDS2017 in the IT field, as well as the publicly available network attack dataset Electra in the industrial control field, to construct a source domain sample database. The second data module collects network traffic from the hardware-in-the-loop simulation platform of the nuclear power instrumentation and control system under various network attack scenarios, and constructs a target domain sample database. The data preprocessing module preprocesses and standardizes the source domain sample database and the target domain sample database, and divides the source domain / target domain training dataset and test dataset. The model building module constructs a global feature extractor based on a self-attention mechanism and a local feature extractor based on a convolutional neural network. The output features of the global and local feature extractors are merged to form a source domain feature extractor. Then, a source domain classifier is built through multiple fully connected layers. Finally, the source domain feature extractor and the source domain classifier are cascaded into an OSSNformer model for detecting source domain network attacks. The OSSNformer model includes an input layer, a feature extraction layer, a classifier, and an output layer; the input layer includes position encoding and spatial feature normalization modules. The feature extraction layer adopts an encoder-decoder architecture. The encoder includes a global feature extractor and a local feature extractor. Both the global feature extractor and the decoder in the encoder are built based on several self-attention layers, while the local feature extractor is built based on full-size convolution. The encoder output encoding is composed of concatenation of global and local features, and the decoder takes the output encoding as input to decode and output the final spatiotemporal features. The model training module trains the OSSNformer model using preprocessed and standardized source domain training data, and saves the trained source domain feature extractor and source domain classifier. The adversarial transfer module constructs a target domain feature extractor based on a state-space model and a domain discriminator composed of multiple fully connected layers. It then uses the domain discriminator, the source domain feature extractor, and the source domain classifier to construct an adversarial transfer architecture. The adversarial domain adaptation architecture is trained using preprocessed and standardized target domain training data, and the trained target domain feature extractor is saved. The clustering module constructs a clusterer of similar types to reduce the intra-class distribution distance of attack data. It uses the K-means clustering method to obtain the class center vector and the class radius threshold vector of each attack type to determine whether the current data is an unknown attack. It uses the clusterer of similar types, the class center vector, the class radius threshold vector and the source domain classifier to construct an unknown attack classifier. The unknown attack classifier is then cascaded with the target domain feature extractor to form a network attack detection model for nuclear power instrumentation and control systems. The detection module inputs the preprocessed and standardized target domain test dataset into the nuclear power plant instrumentation and control system network attack detection model, classifies various network attack data including unknown attacks, and realizes the detection of network attacks on the nuclear power plant instrumentation and control system.

Images