From Bi-Level to One-Level: A Framework for Structural Attacks to Graph Anomaly Detection.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Mining

Background:

• Graph neural networks have advanced graph mining and anomaly detection.
• Existing graph mining methods are susceptible to structural manipulation by attackers.
• Attackers can alter graph structures to evade detection of anomalous nodes.

Purpose of the Study:

• To investigate the structural vulnerability of graph anomaly detection (GAD) systems.
• To develop an effective attack algorithm against GAD systems.
• To demonstrate the efficacy of the proposed attack against unsupervised and supervised GAD methods.

Main Methods:

• Formulating structural poisoning attacks as bi-level optimization problems.
• Transforming the bi-level problem into a one-level optimization problem using regression.
• Optimizing the one-level problem in the discrete domain by leveraging gradient information.

Main Results:

• The proposed attack algorithm, BinarizedAttack, effectively exploits structural vulnerabilities in GAD systems.
• Demonstrated the attack's effectiveness against both unsupervised FeXtra-based GAD and supervised graph convolutional network (GCN)-based GAD.
• Comprehensive experiments validated the proposed attack's performance.

Conclusions:

• GAD systems exhibit significant structural vulnerability to malicious attacks.
• The BinarizedAttack algorithm provides a novel and effective method for attacking GAD systems.
• The findings highlight the need for more robust GAD methods resistant to structural manipulations.