TL;DR
This paper introduces a novel multivariate time series anomaly detection method that combines Graph Attention Networks (GAT) for short-term forecasting with the Informer model for long-term forecasting. This dual-forecasting approach effectively captures complex inter-series relationships and future trends, leading to improved anomaly detection accuracy, especially on high-dimensional industrial control system data.
What Problem Does This Paper Address?
Industrial Control Systems (ICSs) are increasingly connected to external networks, making them vulnerable to cyberattacks. Detecting anomalies in the time series data generated by sensors and controllers within these systems is crucial for maintaining security. While many anomaly detection algorithms exist, their performance often degrades significantly as the number of features (dimensions) in the time series increases. Furthermore, existing methods tend to focus on either short-term predictions or ignore the influence of future trends, limiting their effectiveness in complex, dynamic environments. This paper aims to address these limitations by proposing a method that leverages both short-term and long-term forecasting to accurately detect anomalies in high-dimensional multivariate time series data.
Methods
The proposed method integrates two key deep learning architectures:
- Graph Attention Network (GAT) for Short-Term Forecasting: GAT is used to learn the interrelationships between different time series (features). It constructs a graph structure where nodes represent time series and edges represent their correlations. Graph attention mechanisms then aggregate information from related time series to learn their short-term characteristics. A Gated Recurrent Unit (GRU) is also employed to further enhance the learning of these short-term temporal features. The output is a prediction of the next time step's value.
- Informer for Long-Term Forecasting: Informer, a variant of the Transformer architecture, is utilized for its efficiency in long sequence time series forecasting. It captures longer-term dependencies and future trends in the data.
The core novelty lies in the joint optimization of these two forecasting modules. The short-term forecast (from GAT+GRU) and the long-term forecast (from Informer) are combined to produce a final prediction. This combined prediction is then used to calculate an anomaly score. The loss function is a weighted sum of the losses from both short-term and long-term forecasting, encouraging the model to excel at both tasks. Anomaly scoring is performed by calculating the error between the observed and predicted values, followed by normalization and thresholding.
Datasets
The proposed method was evaluated on two real-world industrial control system datasets:
- SWaT (Singapore Water Treatment Plant): This dataset contains 51 features and simulates attack scenarios in a water treatment plant.
- WADI (Water Distribution Network): An extension of SWaT, WADI is more complex, featuring 127 dimensions, and also simulates cyber-physical attacks.
Both datasets are commonly used benchmarks for time series anomaly detection. The paper notes that for faster training, data was sampled every 10 seconds, with the most frequent label within that interval used as the ground truth.
Key Insights and Ideas
- Synergy of Short- and Long-Term Forecasting: The paper demonstrates that combining short-term predictions (capturing immediate dynamics) with long-term forecasts (understanding future trends) significantly improves anomaly detection accuracy, especially in complex systems.
- Graph Attention for Inter-Series Relationships: GAT effectively models the dependencies between different sensors or features in a multivariate time series, which is crucial for understanding system-wide behavior and detecting anomalies that arise from interactions.
- Informer's Efficiency for Long Sequences: Informer's ability to handle long time series efficiently is vital for capturing the broader context and potential future impacts of anomalies.
- Performance on High-Dimensional Data: The proposed method shows particular strength in handling high-dimensional datasets (like WADI), where many traditional methods struggle.
- Interpretability: The anomaly scoring mechanism, based on prediction errors, allows for the localization and interpretation of anomalies, providing insights into when and where anomalies occur.
Key Results
- Competitive Performance: The proposed method achieved competitive results on both SWaT and WADI datasets, outperforming several advanced baseline methods.
- Improved Recall and F1: On the SWaT dataset, the method improved recall by 1.26% and F1 score by 1.23% compared to the best baseline. For the more challenging WADI dataset, recall improved by 30.33% and F1 by 14.03% compared to the suboptimal baseline.
- High-Dimensional Data Advantage: The method demonstrated superior performance on the high-dimensional WADI dataset, highlighting its scalability.
- Accuracy and Interpretability: The model not only achieved high accuracy in detecting anomalies but also provided interpretable anomaly scores, allowing for better understanding and localization of abnormal events.
- Ablation Studies: Experiments showed that removing any component (joint optimization, Informer, GRU) led to a decrease in performance, validating the necessity of each part of the proposed architecture.
Limitations and Future Work
- Model Size and Parameters: The authors acknowledge that integrating two distinct models (GAT and Informer) can lead to an increase in model size and the number of parameters, potentially impacting operational efficiency.
- Future Work: To address the size limitation, the authors plan to explore knowledge distillation techniques to create a more lightweight version of the model without sacrificing accuracy. They also aim to further optimize the model for more efficient deployment across various environments and to extract even richer time series features for improved anomaly detection.
Summary updated