D4A: An efficient and effective defense across agnostic adversarial attacks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph Neural Networks (GNNs) are increasingly vulnerable to structure adversarial attacks.
• Existing defenses lack robustness against diverse attacks and graph properties.
• GNN vulnerability stems from local graph smoothing and a phenomenon called unfitting.

Purpose of the Study:

• To investigate the abnormal behaviors of GNNs under structure perturbations.
• To develop a robust defense mechanism against structure adversarial attacks in graph data.
• To mitigate the 'unfitting' phenomenon and posterior distribution shifts in GNNs.

Main Methods:

• Analyzing GNN abnormal behaviors from a posterior distribution perspective.
• Proposing smooth-less message passing to enhance perturbation tolerance and mitigate unfitting.
• Introducing distribution shift constraints to prevent other abnormal behaviors.

Main Results:

• The proposed method significantly improves defense performance across various attacks.
• Experimental results on six datasets demonstrate superior robustness compared to 11 baselines.
• The approach achieves a favorable trade-off between accuracy and adversarial robustness.

Conclusions:

• The novel defense strategy effectively addresses GNN structural vulnerability.
• Smooth-less message passing and distribution shift constraints offer intrinsic protection.
• The findings provide a more robust and reliable approach for GNNs in adversarial settings.