A Method for Constructing an Industrial Control System Intrusion Detection System Based on Adversarial Domain Adaptation

Abstract

This invention belongs to the field of industrial control system intrusion detection technology, specifically a method for constructing an industrial control system intrusion detection system based on adversarial domain adaptation. It includes the following steps: S100 – Capture traffic data from two different industrial control communication networks and label the types of the traffic data, using the traffic data from one industrial control network as the source domain data and the other as the target domain data; S200 – Perform numerical preprocessing on the labeled traffic data; S300 – Construct a homogeneous feature space for the preprocessed traffic data; S400 – Construct a multi-view feature extractor to extract feature vectors from the source and target domain data; S500 – Introduce an attention mechanism model into the feature vectors; S600 – Quantize the loss function of the label predictor; S700 – Design multiple domain discriminators for the captured traffic data labels; S800 – Quantize the loss function of the entire adversarial domain adaptation network to complete model training.

Description

Technical Field

[0001] This invention belongs to the field of industrial control system intrusion detection technology, specifically a method for constructing an industrial control system intrusion detection system based on adversarial domain adaptation. Background Technology

[0002] In industrial control systems, intrusion detection technology can detect abnormal attack behaviors by monitoring and analyzing system communication behavior in real time, and perform operations such as interception, alarm, and system recovery before the attack causes harm. The advantages of machine learning and deep network models in feature extraction and fitting have improved the detection efficiency of intrusion detection systems. These algorithms allow models to learn on labeled traffic training datasets, and the trained models are used to predict the labels of future samples, i.e., to identify normal and abnormal traffic. Early on, Beaver et al. evaluated the feasibility and detection capabilities of some supervised machine learning methods in natural gas pipeline intrusion detection systems (Beaver JM, Borges-Hink RC, Buckner M A. An Evaluation of Machine Learning Methods to Detect Malicious SCADA Communications, Proc. of IEEE ICMLA, 2014). Subsequently, Hou et al. constructed an intrusion detection system using a deeper Gated Recurrent Unit (GRU) model, achieving a lower false positive rate and better real-time performance (Hou J, Liu F, Zhuang X. A New Intrusion Detection Model Based on GRU and SalientFeatureApproach. 2019). However, the training of the aforementioned supervised learning algorithms requires a large amount of labeled data. This data often resides in different network layers and has different data distributions and label types. Re-collecting, labeling, retraining, and deploying this data is costly. To address this issue, transfer learning algorithms have been introduced into the construction of industrial control intrusion detection systems, primarily including Fine-tuning and adversarial domain adaptation. Fine-tuning refers to pre-training the model in labeled source domain data and then fine-tuning the pre-trained model in an unlabeled target domain.Singla et al. used a fine-tune method to treat attack traffic data and some normal traffic data in network traffic packets as target domain data, and the remaining data as source domain data (Singla A, Bertino E, Verma D. Overcoming the Lack of Labeled Data: Training Intrusion Detection Models Using Transfer Learning, Proc. of IEEE SMARTCOMP, 2019). However, the effectiveness of fine-tune is closely related to the amount of source domain data and the similarity between the source and target domain data, which greatly limits its application scenarios. Domain adaptation is a branch of transfer learning, aiming to map the source and target domains to the same feature space through a feature extractor. Adversarial domain adaptation builds on this by inputting the extracted features into a label predictor and a domain discriminator. Through adversarial training, the label predictor can correctly predict the category of the source domain data, while the domain discriminator cannot recognize the domain information. Zhang et al. trained a robust intrusion detection system for detecting malicious attacks in the smart grid using an adversarial domain adaptation method (Zhang Y, Yan J. Domain-Adversarial Transfer Learning for Robust Intrusion Detection in the Smart Grid, Proc. IEEE SmartGridComm, 2019). Similarly, Singla et al. built a high-precision intrusion detection system using adversarial domain adaptation even with very small labeled samples in the target domain dataset (Singla A, Bertino E, Verma D. Preparing network intrusion detection deep learning models with minimal data using adversarial domain adaptation, Proc. ACM ASIACCS, 2020). However, their setup is overly idealistic, ignoring the multimodal structure of cross-domain data distribution, and their methods have very limited effectiveness with traffic data. Summary of the Invention

[0003] To address the aforementioned problems, this invention provides a method for constructing an industrial control intrusion detection system based on adversarial domain adaptation.

[0004] This invention adopts the following technical solution: a method for constructing an industrial control intrusion detection system based on adversarial domain adaptation, comprising the following steps: S100 – capturing traffic data in two different industrial control communication networks and labeling the type of traffic data, using the traffic data in one industrial control network as the source domain data and the other as the target domain data; S200 – performing numerical preprocessing on the labeled traffic data; S300 – constructing an isomorphic feature space for the preprocessed traffic data; S400 – constructing a multi-view feature extractor to extract feature vectors from the source domain data and the target domain data; S500 – introducing an attention mechanism model into the feature vectors; S600 – quantizing the loss function of the label predictor; S700 – designing multiple domain discriminators for the captured traffic data labels; S800 – quantizing the loss function of the entire adversarial domain adaptation network to complete model training.

[0005] The specific process of step S100 is as follows: Select two different industrial control communication networks, capture their SCADA network layer traffic data, analyze the attribute values, data characteristics, and attack types of the traffic data, and then label the traffic data, including... K - One attack type and one normal traffic type; the number of tags in the traffic data is: K .

[0006] Step S200 includes the following steps: S201 – Data cleaning, including deduplication of data, correction of erroneous labels in the data, and filling in missing attribute values ​​if any are missing; S202 – Numericalization, using one-hot encoding to quantify character-type data in attribute values, and using Label-Encoding to quantify labels; S203 – Standardization, since the attribute values ​​in the numerically quantified dataset are a mixture of discrete and continuous values, data of different specifications are transformed to the same specification. The standardization formula is... ,in, These are attribute values ​​that have undergone standardization. x For attribute values ​​before standardization processing, The mean of the attribute values. The standard deviation of the attribute values.

[0007] Step S300 uses an autoencoder to reduce the dimensionality of the preprocessed traffic data, constructing a homogeneous feature space. The preprocessed source and target domain data are input into the encoder, and the decoder outputs source and target domain data of the same dimension. The autoencoder consists of an encoder and a decoder. Assuming the input is F-dimensional data, i.e. An encoder is a function that learns functions. The decoder completes the encoding of the input data; the decoder is the learning function. The input features are reconstructed by decoding the encoded features. During the training process of the autoencoder, the error between the controlled input features and the reconstructed input features is minimized.

[0008] Step S400 constructs the feature extractor into a CNN-BiLSTM network structure, using CNN for feature extraction and BiLSTM to aggregate information from different views. The specific process is as follows: S401 - The source domain data and target domain data processed in S300 are input into the fully connected layer, and the output of the fully connected layer is rendered to V different views. These views will be treated as sequences, and CNN is used to extract features from different views; S402 - The features extracted by CNN from different views are passed to BiLSTM, BiLSTM aggregates these features, and then passes them to the fully connected layer; S403 - The features are summed using the fully connected layer, which is the extracted feature vector.

[0009] The attention mechanism model in step S500 is a simple gate mechanism activated by the sigmoid function: ,in S is the ReLU function; S is the output of the attention mechanism, also known as the influence score; the first expansion A(X,W): A refers to the attention model; X is the feature vector obtained from S400; W is the dimensionality reduction layer and the dimensionality expansion layer; the second expansion : This refers to the Sigmoid activation function; g(X,W) is the attention model excluding the Sigmoid function; third expansion : Select two fully connected layers as dimensionality reduction layers, i.e. Where r is the dimensionality reduction ratio parameter; a ReLU activation layer and a dimension expansion layer restore its dimension to D dimensions, i.e. .

[0010] The specific steps of step S600 are as follows: The label predictor... The classification loss is quantified as follows: ,in, For source domain sample instances, For label predictors, For feature extractor, The feature vectors input to the label predictor Influence score, For point The domain tag.

[0011] In step S700: Construct multiple domain discriminant models , Each discriminator corresponds to a label in the source domain data; the loss function for multiple domain discriminators is: ,in, Let K be the k-th domain discriminator, where K represents the total number of domain discriminators and k represents the index of the domain discriminator. For its cross-entropy loss, The feature vector of the input domain discriminator Influence score, For feature extractor, For source domain data, , For point The field label, i, represents the sequence number of the data sample.

[0012] The loss function for quantizing the entire network in step S800 is:

[0013]

[0014] =

[0015] -

[0016] Where S represents the source domain G, t represents the target domain W, and n is the number of samples, n=n s +n t D is the dataset. , For balance coefficient, Feature extractor The parameters, Label predictor The parameters, The parameters of the domain discriminator, These represent the parameters of the attention models corresponding to the label predictor and the domain discriminator, respectively.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] 1. This invention renders traffic data into multiple views and extracts and aggregates features through a CNN-BiLSTM network, thereby improving the expressive power of features and making traffic data more adaptable.

[0019] 2. This invention introduces an attention mechanism in domain adversarial training, enabling the network to perform dynamic feature calibration, simplifying the model learning process, and significantly enhancing the network's representational ability.

[0020] 3. This invention introduces the setting of multiple domain discriminators, and achieves fine-grained alignment of cross-domain data based on multiple domain discriminators by capturing the multi-mode structure of data distribution, thereby maximizing the matching of data distributions of different categories. Attached Figure Description

[0021] Figure 1 This is a flowchart of the industrial control intrusion detection system proposed in this invention;

[0022] Figure 2 This is the autoencoder dimensionality reduction network used in this invention;

[0023] Figure 3 This is the multi-view feature extractor model proposed in this invention;

[0024] Figure 4 This is the model training process for the adversarial domain adaptation method proposed in this invention;

[0025] Figure 5 This is the model testing process for the adversarial domain adaptation method proposed in this invention. Detailed Implementation

[0026] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments, but this does not constitute a limitation thereof.

[0027] This example provides a method for constructing an industrial control system intrusion detection system based on adversarial domain adaptation. The overall process is summarized as follows: Figure 1 As shown, the process begins with capturing flow data from two different industrial control systems, corresponding to step 1 below. Next, the flow data needs to be preprocessed and its feature space isomorphized, corresponding to steps 2 and 3 below. Then, a multi-view feature extractor and attention model are constructed, corresponding to steps 4 and 5 below. Finally, the loss function of the label predictor is quantized, and multiple domain discriminators are constructed. Combining the feature extractor and attention model from steps 4 and 5, the loss function of the entire adversarial domain adaptation network is quantized, and the model is trained, corresponding to steps 6, 7, and 8 below.

[0028] The specific implementation steps for this example are as follows:

[0029] Step 1, as follows Figure 1 As shown in the data capture section, this invention captures flow data at the SCADA network layer of the natural gas pipeline control system (G) and the water storage tank control system (W), both of which use the Modbus RTU protocol for communication. The SCADA network layer data from the natural gas pipeline control system contains 26 features, and the SCADA network layer data from the water storage tank control system contains 23 features. These two data sources each contain 7 types of attack traffic and 1 type of normal traffic, i.e., the number of tags is [number missing]. K =8. This invention uses the traffic data captured from G as the source domain data and the traffic data captured from W as the target domain data to obtain the migration task. .

[0030] Step 2, as follows Figure 1As shown in the data preparation section, this invention performs preprocessing operations on the crawled data: 1) Data cleaning, including deduplication, correction of erroneous labels, and filling in missing attribute values ​​using mean, median, mode, and constant as appropriate. 2) Numericalization, using one-hot encoding to quantify character-type data in attribute values, and using Label-Encoding to quantify labels. 3) Standardization, since the attribute values ​​in the numerically quantified dataset are a mixture of discrete and continuous values, this invention uses a standardization method to transform data of different specifications to the same specification. The standardization formula is: ,in, These are attribute values ​​that have undergone standardization. For attribute values ​​before standardization processing, The mean of the attribute values. The standard deviation of the attribute values.

[0031] Step 3: In this invention, an autoencoder is used to reduce the dimensionality of the payload characteristics of the traffic data from G and W. The structure of the autoencoder is as follows: Figure 2 As shown. The traffic data from G contains 26 features, namely... It contains 7-dimensional payload features, which are mapped by the encoder. and decoder mapping We obtain a 20-dimensional feature space; the traffic data from W contains 23 features, namely... It contains 4-dimensional payload features, which are mapped by the encoder. and decoder mapping A 20-dimensional feature space is obtained. By reconstructing the feature spaces of the source and target domains, the isomorphic feature space is constructed.

[0032] Step 4: Construct the feature extractor as a CNN-BiLSTM network structure. CNN is used for feature extraction, and BiLSTM is used to aggregate information from different views. The specific process is as follows: S401 – The source and target domain data processed in S300 are input into a fully connected layer. The output of the fully connected layer is rendered to V different views, which are treated as sequences. CNN extracts features from these views. S402 – The features extracted by CNN from different views are passed to BiLSTM, which aggregates these features and then passes them to the fully connected layer. S403 – The fully connected layer sums the extracted features to obtain the extracted feature vector. Figure 3 Let's take a look. The structure of the feature extractor is a fully connected layer (multi-view rendering) + CNN (feature extraction) + bidirectional LSTM + fully connected layer (output).

[0033] like Figure 3As shown. First, the traffic data from the source and target domains is input into the fully connected layer, and then the output of the fully connected layer is rendered to... V The system uses three distinct views, treated as sequences, and a CNN extracts features from each view. Then, a BiLSTM is used to aggregate information from all views in the forward and backward directions. Finally, a fully connected layer sums the information from both directions to obtain the extracted feature vector. After step 3, the feature spaces of the source and target domain data are isomorphically transformed into 20 dimensions. The goal of the feature extraction network is to learn a function... Compress the input 20-dimensional vector into one D ( D A more compact expression of <20) dimensions, in which For ReLU function, These are the weight matrix and bias matrix of the feature extraction network, respectively.

[0034] Step 5, considering the time complexity of domain adversarial training, the attention mechanism model in step S500 is a simple gate mechanism activated by the sigmoid function: ,in S is the ReLU function; S is the output of the attention mechanism, also known as the influence score; the first expansion A(X,W): A refers to the attention model; X is the feature vector obtained from S400; W is the dimensionality reduction layer and the dimensionality expansion layer; the second expansion : This refers to the Sigmoid activation function; g(X,W) is the attention model excluding the Sigmoid function; third expansion : Select two fully connected layers as dimensionality reduction layers, i.e. Where r is the dimensionality reduction ratio parameter; a ReLU activation layer and a dimension expansion layer restore its dimension to D dimensions, i.e. The output of the attention model S Let be the influence score vector, where each dimension corresponds to the influence score of each dimension of the feature vector on the classification task.

[0035] In general adversarial domain adaptation methods, two different classification models, label predictor and multiple domain discriminator, need to fit the same set of feature vectors from the source domain. This means that feature dimensions that are weakly correlated with the corresponding classification task will interfere with the model's learning process and increase the burden on the model's gradient convergence process. In addition, the model may also suffer information loss during gradient descent due to insufficient calibration of features by the network.

[0036] Therefore, this invention introduces an attention mechanism in domain adversarial training, enabling the network to perform dynamic feature calibration, simplifying the model learning process, and significantly enhancing the network's representational capabilities.

[0037] S= In general adversarial domain adaptation methods, the two different classification models, label predictor and multiple domain discriminator, need to fit the same set of feature vectors from the source domain. In this way, the feature dimension that is weakly related to the corresponding classification task will interfere with the model's learning process and increase the burden on the model's gradient convergence process. In addition, the model will also suffer information loss during the gradient descent process due to insufficient calibration of features by the network.

[0038] Step 6: This invention employs a gated recurrent unit (GRU) as the label predictor. Feature extraction needs to be considered during the quantization of the label predictor loss function. And multiply by the influence score output by the attention model, so given a pair of source instances Tag predictor The classification loss is: ,in, For label predictors, The feature vectors input to the label predictor Influence score, For point The domain tag.

[0039] Step 7, the present invention addresses the number of tag types obtained in step 1. K =8 Constructing multiple domain discriminant models , Each discriminator corresponds to a label in the source domain data, such as Figure 4 The image shows multiple predictor components. (Label predictor) For each source domain data The output is a probability distribution over the label space of the eight exception types, which nicely describes the... The probability is assigned to each of the eight exception types. Therefore, this invention employs... Introduced as semantic information, it is multiplied by the corresponding domain discriminator to represent each data point. Probability weights aligned with each source domain label. This is achieved by using probability... Its characteristics Weighting can be applied to each data point. The attention level of the domain discriminators is modeled and applied to all eight domain discriminators. , The loss function for multiple domain discriminators can be obtained as follows: ,in, For the k-th domain discriminator, For its cross-entropy loss, The feature vector of the input domain discriminator Influence score, For point The domain tag.

[0040] Step 8: The information-enhanced adversarial domain adaptation network proposed in this invention consists of the multi-view feature extractor constructed in Step 4, the attention mechanism model constructed in Step 5, the label predictor in Step 6, and multiple domain discriminators in Step 7, as follows: Figure 4 As shown. Based on the consideration of the loss of each of the above parts during training, the loss function for the entire network is quantified as follows:

[0041]

[0042] =

[0043] -

[0044] Where S represents the source domain G, t represents the target domain W, and n is the number of samples, n=n s +n t D is the dataset. , This is the balancing coefficient. After training, the domain classifier is no longer used in the model testing phase, such as... Figure 5 As shown, this is the final intrusion detection model.

Claims

1. A method for constructing an industrial control intrusion detection system based on adversarial domain adaptation, characterized in that: Includes the following steps, S100 captures traffic data in two different industrial control communication networks, labels the type of traffic data, and uses the traffic data in one industrial control network as the source domain data and the other as the target domain data. S200 – Performs numerical preprocessing on the traffic data following the label; S300 – Constructs a homogeneous feature space for the preprocessed flow data; S400 – Construct a multi-view feature extractor to extract feature vectors from source domain data and target domain data; Step S400 constructs the feature extractor into a CNN-BiLSTM network structure, using CNN for feature extraction and BiLSTM to aggregate information from different views. The specific process is as follows: S401 inputs the source domain data and target domain data processed by S300 into the fully connected layer, and renders the output of the fully connected layer to V different views. These views will be treated as sequences, and CNN is used to extract features from different views. The S402 CNN extracts features from different views and passes them to the BiLSTM. The BiLSTM aggregates these features and then passes them to the fully connected layer. S403 – The features are summed using a fully connected layer to obtain the extracted feature vector. S500 – A model that introduces an attention mechanism into feature vectors; The attention mechanism model in step S500 is a simple gate mechanism activated by the sigmoid function: ,in S is the ReLU function; S is the output of the attention mechanism, also known as the influence score; the first expansion A(X,W'): A refers to the attention model; X is the feature vector obtained from S400; W' is the dimensionality reduction layer and the dimensionality expansion layer; the second expansion : This refers to the Sigmoid activation function; g(X,W') is the attention model excluding the Sigmoid function; third expansion : Select two fully connected layers as dimensionality reduction layers, i.e. Where r is the dimensionality reduction ratio parameter; a ReLU activation layer and a dimension expansion layer restore its dimension to D' dimensions, i.e. ; S600 – Loss function for quantized label predictor; The specific steps of step S600 are as follows: The label predictor... The classification loss is quantified as follows: ,in, For source domain sample instances, For label predictors, G f For feature extractors, The feature vectors input to the label predictor Influence score, yi as a point The domain tag; S700 – Designs multiple domain discriminators for the tags of captured traffic data; In step S700: Construct multiple domain discriminant models , Each discriminator corresponds to a label in the source domain data; the loss function for multiple domain discriminators is: ,in, Let K be the k-th domain discriminator, where K represents the total number of domain discriminators and k represents the index of the domain discriminator. For cross-entropy loss, The feature vector of the input domain discriminator Influence score, For feature extractors, For source domain data, , For point The field label, i, represents the sequence number of the data sample; S800 – Quantizes the loss function of the entire adversarial domain adaptation network to complete model training; The loss function for quantizing the entire network in step S800 is: = - Where S represents the source domain G, t represents the target domain W, n is the number of samples, n = ns + nt, and D is the dataset. , For balance coefficient, Feature extractor The parameters, Label predictor The parameters, The parameters of the domain discriminator, These represent the parameters of the attention models corresponding to the label predictor and the domain discriminator, respectively.

2. The method for constructing an industrial control intrusion detection system based on adversarial domain adaptation according to claim 1, characterized in that: The specific process of step S100 is as follows: Select two different industrial control communication networks, capture their SCADA network layer traffic data, analyze the attribute values, data characteristics, and attack types of the traffic data, and then label the traffic data, including... K - One attack type and one normal traffic type; the number of tags in the traffic data is: K .

3. The method for constructing an industrial control intrusion detection system based on adversarial domain adaptation according to claim 1, characterized in that: Step S200 Includes the following steps, S201 – Data cleaning, which includes deduplicating data, correcting erroneous labels in the data, filling in missing attribute values ​​if any are missing, and otherwise leaving them unprocessed. S202 - Numericalization Processing: One-hot encoding is used to quantify the character data in the attribute values, and Label-Encoding is used to quantify the tags. S203 – Standardization: The attribute values ​​in the quantified dataset are a mixture of discrete and continuous values. This process transforms data from different specifications to the same specification. The standardization formula is: ,in, These are attribute values ​​that have undergone standardization. x For attribute values ​​before standardization processing, The mean of the attribute values. The standard deviation of the attribute values.

4. The method for constructing an industrial control intrusion detection system based on adversarial domain adaptation according to claim 1 or 3, characterized in that: Step S300 uses an autoencoder to reduce the dimensionality of the preprocessed traffic data and constructs a homogeneous feature space. The preprocessed source domain data and target domain data are input into the encoder, and the decoder outputs source domain data and target domain data of the same dimension.

Images