Leveraging Vulnerabilities in Temporal Graph Neural Networks via Strategic High-Impact Assaults.

Area of Science:

• Graph Neural Networks
• Network Security
• Data Science

Background:

• Temporal Graph Neural Networks (TGNNs) are vital for dynamic graph analysis in various critical sectors.
• Existing adversarial attacks on Spatio-Temporal Dynamic Graphs (STDGs) are often simplistic and lack strategic targeting.
• The robustness of TGNNs against sophisticated temporal attacks remains a significant research challenge.

Purpose of the Study:

• To introduce a novel restricted black-box attack framework, High Impact Attack (HIA), to expose vulnerabilities in TGNNs.
• To develop an attack that overcomes limitations of existing methods by strategically targeting influential nodes and edges.
• To demonstrate the effectiveness of HIA in degrading TGNN performance through stealthy and impactful perturbations.

Main Methods:

• HIA utilizes a data-driven surrogate model to identify structurally and dynamically important nodes.
• A hybrid perturbation strategy combines strategic edge injection and targeted edge deletion.
• The attack minimizes the number of perturbations to enhance stealth and reduce detectability.

Main Results:

• HIA significantly reduces TGNN accuracy on link prediction tasks, with up to a 35.55% decrease in Mean Reciprocal Rank (MRR).
• Experiments on five real-world datasets and four TGNN architectures (TGN, JODIE, DySAT, TGAT) validate HIA's effectiveness.
• The attack demonstrates substantial performance degradation compared to state-of-the-art baselines.

Conclusions:

• Current STDG models exhibit fundamental vulnerabilities to sophisticated adversarial attacks.
• There is an urgent need for robust defenses that address both structural and temporal dynamics in TGNNs.
• The HIA framework provides a critical tool for evaluating and improving the security of dynamic graph analysis.