Anomaly Detection Method for Industrial Control System Operation Data Based on Time-Frequency Fusion Feature Attention Encoding

Jiayi Liu, Yun Sha, Wenchang Zhang, Yong Yan, Xuejun Liu2024 · DOI

TL;DR

This paper introduces TFANet, a novel method for detecting anomalies in industrial control systems (ICS) by fusing time-domain and frequency-domain features with an attention mechanism. This approach effectively captures complex temporal and cross-feature dependencies, outperforming existing methods on multiple ICS datasets.

What Problem Does This Paper Address?

Industrial control systems (ICS) are increasingly connected to the internet, making them vulnerable to cyberattacks. Anomalies in ICS operation data can signal system failures or security breaches. However, ICS data is characterized by its complexity, high dimensionality, and long-sequence nature, with features exhibiting varying periodicities and long-distance, often misaligned, temporal associations. Existing anomaly detection methods struggle to extract comprehensive features from such data, leading to insufficient anomaly detection performance. This paper aims to address this gap by developing a more robust feature extraction and anomaly detection method specifically for ICS data.

Methods

The proposed method, TFANet, employs a time-frequency fusion feature attention encoding strategy.

  • Time-Frequency Transformation: The original time-domain data is transformed into the frequency domain using Fourier transform to obtain amplitude and phase data. This is crucial because periodic variations are often more concentrated in the frequency domain, complementing the time-domain information.
  • Dual-Perspective Feature Extraction: The model uses two main networks:
  • T-Net: Extracts features from the time-domain data, considering both long-term temporal changes and long-distance correlations between features. It utilizes residual blocks composed of convolutional layers, ReLU activation, and pooling to capture features at different scales.
  • F-Net: Extracts features from the frequency-domain data (amplitude and phase), also considering temporal changes and correlations.
  • Feature Fusion: Six distinct feature maps are generated from both T-Net and F-Net. These features are then concatenated to form a comprehensive fusion feature matrix.
  • Attention Encoding (A-Net): A self-attention mechanism is applied to the fused features. This module learns weights for each feature map, allowing the model to focus on the most relevant features for anomaly detection. This attention mechanism helps in understanding the correlation changes between features of different scales.
  • Classification: The weighted features are then fed into a classifier for anomaly detection.

Datasets

The study was validated using three datasets:

  • SWaT (Secure Water Treatment) dataset: A public ICS security research dataset simulating a real water treatment system.
  • WADI (Water Distribution Network Dataset): Another public ICS security research dataset simulating a real water supply system.
  • OTS (Oil Transportation Simulation) dataset: A self-built dataset created by simulating an oil storage business process with a SCADA system model. This dataset includes 119 features and 13 simulated attack types.

Key Insights and Ideas

  • Frequency Domain is Key for Periodicity: Recognizing that periodic patterns are more evident in the frequency domain is a critical insight. Fusing this with time-domain data provides a more complete picture of system behavior.
  • Dual-Perspective Feature Extraction is Crucial: Simply looking at time-domain dependencies isn't enough. Capturing both temporal variations within features and correlations between features, across both time and frequency domains, is essential for robust anomaly detection in complex ICS.
  • Attention Mechanism for Feature Weighting: The attention mechanism effectively learns which extracted features are most indicative of anomalies, allowing the model to adaptively focus on the most informative signals.
  • Small Convolution Kernels Excel at Capturing "Jumps": The ablation studies suggest that smaller convolution kernels (1x3) are more effective at detecting the "instantaneous jumps" characteristic of ICS start-stop operations, which are common in business logic.

Key Results

TFANet demonstrated superior performance across three datasets compared to popular time-series anomaly detection methods like iTransformer, TimesNet, and Crossformer.

  • SWaT Dataset: Achieved an average improvement of approximately 19% in accuracy, 37% in precision, 31% in recall, 35% in F1-score, and 22% in ROC-AUC.
  • WADI Dataset: Showed significant improvements, with TFANet achieving accuracy of 0.80152, precision of 0.80743, recall of 0.80256, F1-score of 0.80089, and ROC-AUC of 0.80254, outperforming others.
  • OTS Dataset: TFANet achieved accuracy of 0.80160, precision of 0.78486, recall of 0.69823, F1-score of 0.70926, and ROC-AUC of 0.83673, again showing substantial gains.
  • Ablation Studies: Confirmed that a 1x3 convolution kernel generally performs better than larger kernels (1x11) for ICS anomaly detection, especially for capturing sudden changes. The number of attention layers and feedforward dimensions also impact performance, with specific configurations showing optimal results.

Limitations and Future Work

  • Convolution Kernel Size: While smaller kernels (1x3) are shown to be effective for capturing instantaneous jumps, the paper notes that larger kernels might be better for capturing broader "patterns" of business logic. The optimal kernel size might depend on the specific dataset and anomaly types.
  • Impact of Attention Layers and Dimensions: The ablation studies suggest that while attention layers can improve performance, excessively deep networks or overly large feedforward dimensions can lead to information dilution and gradient disappearance. Future work could explore more efficient ways to manage network depth and dimensionality.
  • Generalizability: The paper suggests that when applying this method to other datasets, factors like anomaly types, system business logic, and data dimensions need to be considered to fine-tune parameters like the convolution kernel size. This implies that while the method is robust, further adaptation might be needed for diverse ICS environments.

Summary updated

BibTeX
@misc{liu2024anomaly,
  author = {Jiayi Liu and Yun Sha and Wenchang Zhang and Yong Yan and Xuejun Liu},
  title = {{Anomaly Detection Method for Industrial Control System Operation Data Based on Time-Frequency Fusion Feature Attention Encoding}},
  year = {2024},
  doi = {10.3390/s24186131},
}
Figure