Adversarial Danger Identification on Temporally Dynamic Graphs.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Multivariate time series forecasting is crucial for applications like finance and healthcare.
• Temporal graph neural networks (GNNs) excel at capturing complex patterns but are vulnerable to adversarial attacks.
• Existing defenses are inadequate for dynamic forecasting due to generalization and contradiction issues.

Purpose of the Study:

• To develop an effective defense mechanism for GNN-based multivariate time series forecasting models against adversarial attacks.
• To address the overlooked challenge of defending these models in real-world applications.

Main Methods:

• A three-step adversarial danger identification method for temporally dynamic graphs.
• Utilizing a hybrid GNN-based classifier to pinpoint dangerous time points.
• Employing approximate linear error propagation to identify critical variables.
• Implementing a scatter filter to refine time series data with minimal feature loss.

Main Results:

• The proposed method effectively protects GNN-based forecasting models against various adversarial attacks.
• Experiments demonstrated significant improvements in model robustness across multiple state-of-the-art forecasting models.
• The method successfully identifies dangerous times and variates, mitigating adversarial influence.

Conclusions:

• The developed adversarial danger identification method provides a robust defense for GNN-based forecasting models.
• This research bridges the gap in adversarial defense for multivariate time series forecasting.
• The findings enhance the reliability and security of deep learning models in critical forecasting applications.