A Lightweight Anomaly Detection Method for Unsupervised Domain Adaptive Industrial Control Networks

By adopting a lightweight anomaly detection method with unsupervised domain adaptation, combined with sparse convolutional autoencoder (SRCAE) and cross-dimensional interactive spatial channel attention mechanism (SCA), the problems of data label dependence and large computational load in industrial control networks are solved, achieving efficient and accurate anomaly detection and adapting to feature differences caused by operating condition drift.

CN118861712BActive Publication Date: 2026-11-17SHENYANG INSTITUTE OF CHEMICAL TECHNOLOGY
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202410970499.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-19
Publication Date
2026-11-17
Estimated Expiration
2044-07-19

AI Technical Summary

Technical Problem

Existing methods for detecting anomalies in industrial control networks rely on labeled data, which involves a large amount of computation and cannot effectively cope with differences in data characteristics caused by shifts in operating conditions, resulting in a decline in detection performance.

Method used

We employ an unsupervised, domain-adaptive, lightweight anomaly detection method that combines a sparse convolutional autoencoder (SRCAE) and a cross-dimensional interactive spatial channel attention mechanism (SCA). We use maximum mean difference (MMD) for domain adversarial training and an improved real-time streaming clustering method (DenStream) to establish a detection baseline, thereby reducing computational cost and adapting to data changes.

Benefits of technology

Improve model performance with low computational overhead, enhance generalization ability to industrial control networks, improve detection efficiency and accuracy, reduce dependence on labeled data, and adapt to feature differences caused by operating condition drift.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118861712B_ABST
    Figure CN118861712B_ABST
Patent Text Reader

Abstract

A lightweight, unsupervised, domain-adaptive anomaly detection method for industrial control networks is disclosed. This invention trains a model using industrial control network data from both historical and online phases. During the domain adversarial training phase, an unsupervised spatiotemporal feature extractor (SCA-SRCAE) is applied to improve model detection performance. Maximum Mean Discrepancy (MMD) is introduced as a regularization method to train a domain-shared spatiotemporal feature extractor. Simultaneously, source and target domain data are used to train a domain discriminator, seeking domain-invariant feature representations to improve model generalization ability. An improved Denstream is used to form the contour baseline of normal communication messages, enhancing the efficiency of anomaly detection in industrial control networks.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to an anomaly detection method for industrial control networks, and more particularly to a lightweight anomaly detection method for unsupervised domain adaptive industrial control networks. Background Technology

[0002] Industrial control systems (ICS) are a crucial component of critical infrastructure, and their secure operation is vital to societal development. With the rapid advancement of computer and network technologies, ICS are evolving towards intelligent and networked architectures. While improving the production efficiency of ICS, this also exposes their network systems to security threats. Against this backdrop, intrusion detection for industrial control networks has been widely studied as an effective method for addressing cybersecurity threats. Among these methods, anomaly-based intrusion detection can detect unknown attacks, offering advantages such as strong adaptability and flexibility. It does not require predefined attack signatures and is particularly effective in detecting novel attacks.

[0003] Existing anomaly detection algorithms are powerful, but when faced with real-world ICS network environments, the performance of these models in practical applications still faces three main challenges.

[0004] Anomaly detection models rely on a large amount of labeled data during training, while real-world industrial control network data is unlabeled. Domain-adversarial anomaly detection methods require high detection efficiency, but previous models have overemphasized detection accuracy while neglecting computational complexity.

[0005] 2. Anomaly detection models are trained using historical data. However, due to varying operating conditions, historical and online data may differ, impacting the model's practical performance. Most anomaly detection methods use historical data for training, but actual industrial control network data exhibits operating condition drift. Therefore, unclear or inconsistent data features can lead to decreased detection performance.

[0006] 3. Current adversarial anomaly detection models in this field have high requirements for detection efficiency. However, previous models have focused excessively on detection accuracy while neglecting computational complexity. Therefore, an efficient anomaly detection solution that is suitable for real-world industrial control networks is needed to address these issues. Training a large number of models requires labeled data, and in practical industrial control scenarios, obtaining these labels excessively consumes manual labor. Summary of the Invention

[0007] The purpose of this invention is to provide a lightweight anomaly detection method for industrial control networks using unsupervised domain adaptive methods. This method applies the domain-shared unsupervised spatiotemporal feature extractor SCA-SRCAE to a domain adversarial neural network, addressing the model's dependence on data labels and improving model performance with low computational overhead. Maximum mean difference (MMD) is introduced during the domain adversarial training phase to further optimize the feature extractor and domain discriminator, addressing the problem of feature distribution differences in the data. An improved real-time streaming clustering method, DenStream, is used to enhance the efficiency of anomaly detection in industrial control networks.

[0008] The objective of this invention is achieved through the following technical solution: A lightweight anomaly detection method for unsupervised domain adaptive industrial control networks, comprising the following steps: S1: Attack the network security attack and defense exercise range of oil and gas gathering and transportation in the key laboratory for information security in the petrochemical industry. Use Wireshark network packet analysis software to obtain abnormal and normal data, and preprocess the data to make it meet the model input requirements. S2: A lightweight spatial channel attention (SCA) mechanism for capturing cross-dimensional interactions is applied to the sparse convolutional autoencoder SRCAE. The input tensor is fed into the spatial attention mechanism and the channel attention mechanism respectively, and finally new attention weights are generated and applied to the input tensor to obtain a weighted tensor with cross-dimensional interaction function. SCA-SRCAE is used to select spatiotemporal features in an unsupervised manner from the historical normal source domain data and the online target domain normal data collected by Wireshark, thereby improving the model performance. S3: Perform domain adversarial training, backpropagating the Maximum Mean Discrepancy (MMD) during the training to train the domain shared spatiotemporal feature extractor SCA-SRCAE. Then, train a domain discriminator using source and target domain data to acquire domain-invariant features. S4: Pass the spatiotemporal features of normal data from the source and target domains to the improved real-time stream clustering model Denstream, aiming to achieve a baseline profile of normal communication messages based on valuable spatiotemporal features; S5: Based on the contour baseline of normal data, the improved Denstream gradually adapts to changes in normal behavioral information, adapts to dynamic changes in data while ensuring the accuracy of model prediction, updates clustering results in real time, and detects abnormal network changes in actual industrial control systems.

[0009] The aforementioned lightweight anomaly detection method for unsupervised domain adaptive industrial control networks uses anomaly detection data packets in S1 that originate from attacks on the oil and gas gathering and transportation network security attack and defense exercise range in the key laboratory for information security in the petrochemical industry. Wireshark network packet analysis software is used to obtain abnormal and normal data, and the data is preprocessed to make it meet the model input requirements.

[0010] The aforementioned lightweight anomaly detection method for unsupervised domain adaptive industrial control networks integrates the unsupervised sparse convolutional autoencoder (SRCAE) in S2 with the lightweight channel attention mechanism (SCA) for capturing cross-dimensional interactive spaces, thereby improving model performance.

[0011] The aforementioned lightweight anomaly detection method for unsupervised domain adaptive industrial control networks introduces the maximum mean difference as regularization during domain adversarial training in S3 to train the domain shared spatiotemporal feature extractor SCA-SRCAE. When the discriminator cannot identify whether the data belongs to the source domain or the target domain, it indicates that the domain shared feature extractor has extracted domain-invariant features.

[0012] The aforementioned lightweight anomaly detection method for unsupervised domain adaptive industrial control networks utilizes an improved real-time stream clustering model, Denstream, in steps S4 and S5 to gradually adapt to changes in normal behavior information, dynamically detecting abnormal changes in the industrial control network while ensuring the accuracy of model predictions.

[0013] The aforementioned unsupervised domain adaptive lightweight anomaly detection method for industrial control networks utilizes the unsupervised spatiotemporal feature extractor SCA-SRCAE, as described in steps S2, S3, S4, and S5, combined with a lightweight cross-dimensional spatial channel attention SCA mechanism to improve model performance. During the domain adversarial training phase, Maximum Mean Difference (MMD) is used as a regularization in unsupervised learning to train the spatiotemporal feature extractor, reducing the distribution difference between the source and target domains and addressing the issue of data distribution discrepancies to extract domain-invariant features. Finally, an improved real-time stream clustering method, DenStream, is used to form a detection baseline based on data features of normal behavior, thereby improving the efficiency of anomaly detection in industrial control networks.

[0014] The advantages and effects of this invention are: This invention utilizes data from the oil and gas gathering and transportation network security attack and defense exercise range in the Key Laboratory of Information Security in the Petrochemical Industry. Compared with traditional anomaly detection models, it applies domain adversarial neural network (SRCAE) knowledge to the anomaly detection system, effectively addressing the feature space differences in industrial control network data caused by operational condition drift and improving the model's generalization ability. The SRCAE model uses SCA-SRCAE to capture the enhanced spatiotemporal features of cross-dimensional spatial channel interactions in an unsupervised manner, significantly reducing the time spent annotating industrial data packets and achieving model performance improvement with low computational overhead. Furthermore, in the real-time stream clustering stage, an improved Denstream is used to establish a detection baseline based on the spatiotemporal features of normal data, improving the efficiency of anomaly detection in industrial control networks. Attached Figure Description

[0015] Figure 1 This is a flowchart of a lightweight anomaly detection method for unsupervised domain adaptive industrial control networks according to the present invention. Figure 2 This is a schematic diagram of a sparse convolutional autoencoder (SRCAE). Figure 3 A schematic diagram of the channel attention mechanism (SCA) for capturing cross-dimensional interactive spaces; Figure 4 This is a schematic diagram of the unsupervised adversarial training process. Figure 5 A schematic diagram of the improved real-time stream clustering model. Detailed Implementation

[0016] To make the objectives, contents, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0017] Model Design While existing anomaly detection algorithms are powerful, their performance in practical applications still falls short. This invention provides a lightweight anomaly detection method for industrial control networks using unsupervised domain adaptive methods, primarily addressing the following issues: Existing methods for detecting anomalies in industrial control networks rely on historical data from the industrial control network for model training, which cannot effectively address the discrepancies between online and historical data caused by operational condition drift. Currently, training domain adversarial models typically requires a large amount of labeled data, and anomaly detection methods based on labeled data are difficult to apply to anomaly detection in real industrial control networks. Current adversarial anomaly detection methods in the field have high requirements for detection efficiency, but previous models have focused too much on detection accuracy and ignored the computational cost of the model.

[0018] To address the existing problems, the method of the present invention is as follows: Figure 1This paper demonstrates the integration of domain adversarial neural network (DNS) knowledge with anomaly detection knowledge. A lightweight, cross-dimensional interactive spatial channel attention (SCA) mechanism is applied to the unsupervised domain-shared spatiotemporal feature extractor (SRCAE) as a domain-shared autoencoder. Maximum mean difference (MMD) is used as a regularization in unsupervised learning to reduce the distributional differences between the source and target domains, addressing the issue of distributional discrepancies between training and test data. Based on the spatiotemporal characteristics of normal data, an improved real-time stream clustering method (DenStream) is used for anomaly detection, improving the anomaly detection efficiency of industrial control networks.

[0019] 1) Unsupervised Domain Shared Spatiotemporal Feature Extractor SCA-SRCAE: like Figure 2 The SRCAE shown mainly consists of an extended causal convolutional layer, a batch normalization layer, and the LeakyReLU activation function. The convolutional layer is used for extracting local spatial features, and the batch normalization layer is used to normalize the data to balance each batch of data. The extended causal deconvolutional layer of the decoder module completes the functions of upsampling and data reconstruction. The data is input into the encoder and encoded using the encoding function, which is shown in Equation (1): (1) Input sample Mapped to hidden layer In this process, valuable feature mapping from high-dimensional to low-dimensional dimensions is achieved. The decoder aims to utilize the decoding function to... China Resumption Reconstruct the input data.

[0020] The decoding function is shown in equation (2): (2) (3) (4) and Represents the weights and biases in the encoder. and These are the weights and biases in the decoder. and This represents the output of the encoder and decoder. The model uses two convolutional modules to extract valuable features for the clustering layer. The information dilation value of the first convolutional module is set to 1 to retain sufficient important information, obtaining a one-dimensional time series feature. In the second convolutional block, the dilation value is set to 2 to extract temporal features. The model revisits multiple historical time steps to construct intermediate features with spatiotemporal correlation.

[0021] To improve the model's ability to capture important spatial and temporal information from the input data, Figure 3The lightweight capture-across-dimensional interactive spatial channel attention mechanism (SCA) shown is applied in the encoder of a sparse convolutional autoencoder after the first convolutional layer.

[0022] The two branches of the spatial channel attention mechanism are responsible for capturing the channel dimension. With spatial dimension or This involves cross-dimensional interactions between the input tensors. SCA rearranges the input tensor through a permutation operation. The input tensor is then sequentially fed into a convolutional layer and a sigmoid activation layer, generating two attention maps for cross-dimensional interactions. Finally, the attention maps are permuted again and applied to the input tensor via multiplication to obtain feature representations that enhance interactions across all dimensions, improving model performance with low overhead.

[0023] 2) Domain adversarial model training: Unsupervised adversarial training process as follows Figure 4 As shown, this not only addresses the impact of feature distribution differences on model detection performance in practical applications but also eliminates the dependence on data labels. This module primarily uses a sparse convolutional autoencoder (SCA-SRCAE) combined with SCA as a domain-shared unsupervised spatiotemporal feature extractor to extract temporal and spatial features. During domain adversarial training, maximum mean difference (MMD) is introduced as a regularization method to train the domain-shared spatiotemporal feature extractor to extract domain-invariant features. Simultaneously, source and target domain data are used to train a domain discriminator, making it unable to determine the data source. The domain adversarial training loss function is shown in equation (7): (7) In the formula Represents source domain characteristics, This represents the features of the target domain. Among them... It is a discriminator The weights are such that, in order to make the discriminator more discriminative, formula (7) is maximized.

[0024] 3) Improved real-time stream clustering Denstream: Improved DenStream, such as Figure 5 This illustrates a streaming clustering algorithm that can execute in real-time as new data is continuously generated, making it suitable for real-time applications and scenarios involving processing real-time data streams. Density-based clustering algorithms do not require a predefined number of clusters and can form clusters of arbitrary shapes. To construct... Microcluster Two important parameters need to be defined. Minimum number of points. Dataset and points The distance is less than or equal to The set of all points in a dataset is defined as a neighborhood. Calculate the neighborhood of a point in the dataset. The neighborhood of is shown in equations (8) and (9): (8) (9) In equation (9) Point and points The Euclidean distance between them. Calculate the Euclidean distance for a single point. The neighborhood is used to determine the core object when the number of points in the neighborhood exceeds the minimum number of points. Once the core object is determined, the algorithm repeats the above steps until all points have been traversed. This is based on the temporal and spatial characteristics of normal data. Micro-clustering removes data points from the Denstream abnormal buffer and models them for normal data.

[0025] The present invention can be further understood from the following embodiments using the above method. Example 1

[0026] This invention provides a lightweight anomaly detection method for unsupervised domain adaptive industrial control networks, the flowchart of which is shown below. Figure 1 As shown, it includes the following steps: Step 1: Data Acquisition. Before launching the attack, Namp was used for network sniffing to obtain information on device deployment and open ports within the LAN of the attack and defense exercise platform. A detailed scan of the SIEMENS S7-300 within the test range was performed using PLC Scan. Information such as the PLC's CPU model, hardware version, and batch was obtained. Based on this information, a denial-of-service vulnerability was discovered. The Hping3 DOS attack tool, included with the Kali system, was used to launch an attack, causing abnormal behavior in the industrial control system and generating abnormal data. Wireshark network packet analysis software was used to obtain both abnormal and normal data, and the packets were preprocessed to ensure they matched the model input.

[0027] Step 2: Data Processing. Input 500 unlabeled historical data points from the source domain and 500 online data points from the target domain into the shared feature extractor SCA-SRCAE. After passing through the first convolutional module, a time-series feature tensor is obtained. This output tensor is then passed to the two branches of the lightweight SCA to obtain a channel-spatial cross-dimensional interaction weighted tensor. Multiplying the weighted tensor with the output tensor yields the interaction-enhanced features. Reviewing historical time steps, spatiotemporally relevant intermediate features are constructed, ultimately obtaining the spatiotemporal features.

[0028] Step 3: Domain Adversarial Training. During SCA-SRCAE training, backpropagation of the maximum mean difference (MMD) minimizes the feature space difference between historical data from the source domain and online data from the target domain, resulting in the minimum MMD loss. Initially, the domain discriminator loss is high. As the discriminator loss decreases, it eventually becomes unable to identify the domain from which the data originates, indicating that the feature extractor has learned domain-invariant feature representations.

[0029] Step 4: Anomaly Detection. During the training phase, the improved DenStream continuously updates the core P micro-clusters to form the baseline of normal communication messages. During the testing phase, the algorithm iterates through the Euclidean distance between the center of each micro-cluster and the current point, finds the nearest p micro-cluster, and determines whether to merge it. Points that cannot be merged are classified as anomalous data and placed in the anomaly buffer. Over time, micro-clusters with weights below a threshold are removed to ensure that the model retains only micro-clusters with normal communication messages. This ensures efficient and accurate anomaly detection.

[0030] The method of this invention detected 442 abnormal data points. When using SRCAE as a feature extractor, the model only detected 395 abnormal data points. Compared with using a semi-supervised domain adaptive model, the method of this invention can save 221 seconds of training time. Therefore, it is proven that this method improves accuracy and detection efficiency when processing real-world industrial control network data, ensuring the stability and security of ICS.

[0031] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. An unsupervised domain adaptation method for lightweight anomaly detection in industrial control networks, characterized in that, The method includes the following steps: S1: Attack the network security attack and defense exercise range of oil and gas gathering and transportation in the key laboratory for information security in the petrochemical industry. Use Wireshark network packet analysis software to obtain abnormal and normal data, and preprocess the data to make it meet the model input requirements. S2: A lightweight Spatial Channel Attention (SCA) mechanism for capturing cross-dimensional interaction is applied to the Sparse Convolutional Autoencoder (SRCAE). SCA is applied after the first convolutional layer in the encoder of the SRCAE. SCA-SRCAE is used to select spatiotemporal features in an unsupervised manner from historical source domain data and online target domain data collected by Wireshark, thereby improving model performance. Historical source domain data and online target domain data are input into SCA-SRCAE, and the first convolutional module generates a time-series feature tensor. This output tensor is then passed to the two branches of the lightweight SCA to obtain a channel-space cross-dimensional interaction weighted tensor. The weighted tensor is multiplied by the output tensor to obtain the interaction enhancement features. Historical time steps are reviewed to construct spatiotemporally relevant intermediate features, ultimately obtaining the spatiotemporal features. S3: Perform domain adversarial training, backpropagate the maximum mean discrepancy (MMD) during domain adversarial training to train the domain shared spatiotemporal feature extractor SCA-SRCAE; and train the domain discriminator using source domain data and target domain data to obtain domain invariant features; S4: Pass the spatiotemporal features of normal data from the source and target domains to the improved real-time stream clustering model Denstream, aiming to achieve a baseline profile of normal communication messages based on valuable spatiotemporal features; S5: Based on the contour baseline of normal data, the improved Denstream gradually adapts to changes in normal behavioral information, adapts to dynamic changes in data while ensuring the accuracy of model prediction, updates clustering results in real time, and detects abnormal network changes in actual industrial control systems.

2. The unsupervised domain adaptation lightweight anomaly detection method for industrial control network according to claim 1, characterized in that, The anomaly detection data packets in S1 originate from attacks on the oil and gas gathering and transportation network security attack and defense exercise range in the key laboratory for information security in the petrochemical industry. Abnormal and normal data are obtained using Wireshark network packet analysis software, and the data is preprocessed to make it meet the model input requirements.

3. The unsupervised domain adaptation lightweight anomaly detection method for industrial control network according to claim 1, characterized in that, The fusion of the unsupervised sparse convolutional autoencoder SRCAE and the lightweight capture cross-dimensional interactive spatial channel attention mechanism (SCA) in S2 improves model performance.

4. The unsupervised domain adaptation lightweight anomaly detection method for industrial control network according to claim 1, characterized in that, During the domain adversarial training in S3, the maximum mean difference is introduced as a regularization to train the domain shared spatiotemporal feature extractor SCA-SRCAE. When the discriminator cannot identify whether the data belongs to the source domain or the target domain, it indicates that the domain shared feature extractor has extracted domain-invariant features.

5. The unsupervised domain adaptation lightweight anomaly detection method for industrial control network according to claim 1, characterized in that, The improved real-time stream clustering model Denstream used in S4 and S5 gradually adapts to changes in normal behavioral information, dynamically detecting abnormal changes in the industrial control network while ensuring the accuracy of model prediction.

6. The unsupervised domain adaptation lightweight anomaly detection method for industrial control network according to claim 1, characterized in that, The S2, S3, S4 and S5 processes employ the unsupervised spatiotemporal feature extractor SCA-SRCAE to improve model performance. In the domain adversarial training phase, the maximum mean difference (MMD) is used as a regularization in unsupervised learning to train the spatiotemporal feature extractor, reduce the distribution difference between the source domain and the target domain, and solve the problem of data distribution difference to extract domain-invariant features. Finally, an improved real-time streaming clustering method, DenStream, is used to form a detection baseline based on data characteristics of normal behavior, thereby improving the efficiency of anomaly detection in industrial control networks.

Citation Information

Patent Citations

  • Unsupervised cross-domain self-adaptive medical image segmentation method based on deep adversarial learning

    AU2020103905A4

  • Methods for determining amount of NAD metabolites from sample and methods and uses related thereto

    CN116249904A