Graph Attention Network and Informer for Multivariate Time Series Anomaly Detection.

Area of Science:

• Cybersecurity
• Artificial Intelligence
• Industrial Control Systems

Background:

• Time series anomaly detection is crucial for industrial control system (ICS) security.
• Existing algorithms often struggle with high-dimensional data, leading to performance degradation.

Purpose of the Study:

• To propose a robust anomaly detection scheme for ICSs that overcomes the limitations of high-dimensional data.
• To enhance the security and reliability of industrial control systems through advanced anomaly detection.

Main Methods:

• A novel scheme combining Graph Attention Network (GAT) for sequential characteristics and Informer for long time series prediction.
• Utilizing both long-time and short-time forecasting losses for multivariate time series anomaly detection.
• Experimental validation on SWaT and WADI industrial control system datasets.

Main Results:

• Achieved competitive results compared to state-of-the-art methods, particularly on higher-dimensional datasets.
• Demonstrated the method's ability to accurately locate anomalies within time series data.
• The proposed approach offers interpretability in anomaly detection.

Conclusions:

• The GAT and Informer-based scheme provides an effective solution for time series anomaly detection in ICSs.
• The method shows significant improvements in handling high-dimensional data, enhancing system security.
• The approach offers both accurate detection and interpretability for anomalies.