Anomaly Detection in Industrial Control Systems Based on Cross-Domain Representation Learning
Dongyang Zhan*, Wenqi Zhang, Lin Ye, Xiangzhan Yu, Hongli Zhang, and Zheng He · 2025
Read Anomaly Detection in Industrial Control Systems Based on Cross-Domain Representation Learning by Dongyang Zhan*, Wenqi Zhang, Lin Ye, Xiangzhan Yu, Hongli Zhang, and Zheng He (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-11786.txt.
Introduction
Dongyang Zhan and 5 co-authors detect anomalies in industrial control systems by joining network and sensor evidence in a 2025 security paper. Such systems run physical processes from afar, and attacks on them can cause serious damage. Most detectors examine only one domain at a time.
Network traffic and physical sensor behavior are linked, so single-domain study misses anomalies that cross domains. The authors build a cross-domain graph that stands for behaviors from many domains. A graph neural network learns their joint features. Because anomalies show up differently per domain, a multi-task setup detects them domain by domain while training jointly. Tests showed better results than past approaches at spotting anomalies. The method targets the security and stability of industrial work.
Catalog record
| CREATOR | Dongyang Zhan*, Wenqi Zhang, Lin Ye, Xiangzhan Yu, Hongli Zhang, and Zheng He |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 65 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-11786
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Anomaly Detection in Industrial Control Systems Based on Cross-Domain Representation Learning Dongyang Zhan*, Wenqi Zhang, Lin Ye, Xiangzhan Yu, Hongli Zhang, and Zheng He Abstract Industrial control systems (ICSs) are widely used in industry, and their security and stability are very important. Once the ICS is attacked, it may cause serious damage. Therefore, it is very important to detect anomalies in ICSs. ICS can monitor and manage physical devices remotely using communication networks. The existing anomaly detection approaches mainly focus on analyzing the security of network traffic or sensor data. However, the behaviors of different domains (e.g., network traffic and sensor physical status) of ICSs are correlated, so it is difficult to comprehensively identify anomalies by analyzing only a single domain. In this paper, an anomaly detection approach based on cross-domain representation learning in ICSs is proposed, which can learn the joint features of multi-domain behaviors and detect anomalies within different domains. After constructing a cross-domain graph that can represent the behaviors of multiple domains in ICSs, our approach can learn the joint features of them by leveraging graph neural networks. Since anomalies behave differently in different domains, we leverage a multi-task learning approach to identify anomalies in different domains separately and perform joint training. The experimental results show that the performance of our approach is better than existing approaches for identifying anomalies in ICSs. I Introduction Industrial control systems (ICSs) are critical components of many modern industrial and infrastructure systems, such as power grids, and manufacturing plants. They enable remote monitoring and control of physical sensors or actuators based on network communication, which is very important for industry. However, ICSs are vulnerable to cyberattacks, and the consequences of an attack can be serious. In recent years, lots of attacks against ICSs have triggered serious consequences. For instance, the Stuxnet worm [ 1 ] attacked Iranian nuclear facilities, and it demonstrated the potential for cyberattacks to cause physical damage to industrial systems. In 2021, the Colonial Pipeline was hit by a ransomware attack, causing widespread fuel shortages [ 2 ] . In 2015 and 2016, Ukraine’s power grid was attacked, which caused power outages [ 3 ] . In the past, ICSs are isolated from the Internet and run in physically secure locations, so the security mechanisms are not well designed [ 4 ] . Currently, as ICSs are increasingly connected to the Internet, ICSs are facing more and more security threats. Therefore, it is crucial to secure ICSs and ensure the security and reliability of critical infrastructure systems. There are several challenges of securing ICSs. First, many legacy protocols are not designed with modern security features, making them vulnerable to attacks [ 5 ] such as denial of service, man in the middle, etc. For instance, MQTT is a widely used messaging protocol used for machine-to-machine communication in IoT networks that has been found to have many vulnerabilities [ 6 ] . Second, the devices (e.g., sensors and actuators) are mostly low-power and have limited computing power for security analysis and protection functions [ 7 ] . Therefore, it is difficult for administrators to enhance security directly based on existing ICSs. Many approaches have been proposed to secure ICSs. First, intrusion detection systems (IDSs) [ 8 ] can be applied to secure ICSs, which typically identify the network behavioral characteristics of known attacks but face the problems of identifying only few known attacks and the inability to detect unknown attacks. In addition, an important difference between ICS anomaly detection and network anomaly detection is that ICS managers have access to data from each sensor in the network (e.g., pressure, flow, etc.), based on which anomalies can be more accurately identified. This is because some anomalies do not cause changes in network traffic, but rather in anomalies in sensor data. Second, security analysis approaches based on behavior models are proposed, which construct the behavior/data models of every sensor, actuator and the interaction between them. If a physical sensor value or network packet does not fit the model, an anomaly alert is generated. But it is difficult to build comprehensive models of complex industrial processes systematically, so these approaches lack generality and are not widely used in ICSs [ 9 ] . To address the limitations of traditional approaches, deep learning is utilized for anomaly detection in ICSs. Utilizing the capabilities of deep neural networks, these techniques can learn features that determine anomaly scores, resulting in improved accuracy in anomaly detection. Specifically, the Recurrent Neural Network (RNN) [ 10 ] offers the potential to model the time series values of ICS physical devices, such as temperature sensors and pressure sensors. It captures the dependencies between inputs and outputs by extracting partial sequences from relevant inputs and scores anomalies based on the predicted or reconstructed features of ICS physical device values. Further building on the foundations of RNN, attention mechanism-imbued time series prediction models such as LSTM-NDT [ 11 ] and openGauss [ 12 ] have been introduced, showcasing superior performance in identifying enduring correlations in ICS physical device values. Models such as TranAD [ 13 ] and Anomaly Transformer [ 15 ] adapt the Transformer architecture for time-series anomaly detection, which effectively capture rich associations across entire sequences using their self-attention weight distribution. Besides, the Generative Adversarial Network (GAN) [ 16 ] also stands out for its prowess in multivariate anomaly detection through bias reconstruction, as demonstrated in models such as TAnoGAN [ 17 ] and GAN-AD [ 18 ] . However, it is difficult for these approaches to detect anomaly points in the time series values of ICS physical devices because they cannot model the topological and interaction relationships between nodes in ICSs. Graph Neural Networks (GNNs) [ 19 ] have shown significant promise in graph representation learning. Graph Convolutional Network (GCN) [ 20 ] is a convolutional neural network-based model that generates node embeddings by applying convolutional operations on graphs. Consequently, GCN performs well in graph embeddings and can incorporate inter-node correlations into the learning modeling process. Based on GCN, the Graph Deviation Network (GDN) [ 21 ] utilizes embedding vectors to capture the physical device features of ICSs (i.e., the values of each sensor or actuator in the real world). It encodes and learns the relationships between pairs of devices as graph edges. It predicts future device behaviors by applying an attention function to the neighboring sensors in the graph. Additionally, it identifies and explains deviations from the learned sensor relationships. While hybrid models (NSIBF [ 22 ] and MTAD-GAT [ 23 ] ) use self-supervised methods to explicitly capture the correlation between features of different physical devices in time series data. Feng and Tian utilize Bayesian filtering to handle implicit relationships, while Zhao et al.’s approach involves direct learning through graph attention networks. However, both methods result in a high level of task complexity. GLIN [ 24 ] analyzes the joint representations of local nodes and the global network to perform anomaly detection. This method is used for node-level anomaly detection within industrial control systems. MMGAN [ 25 ] jointly learns the features of physical device time series in both the time and frequency aspects, enhancing the model’s ability to model distribution. However, these methods are designed