Backdoor attacks on unsupervised graph representation learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Analytics

Background:

• Unsupervised graph learning methods, often using mutual information maximization, generate node and graph representations.
• These techniques are increasingly popular but lack robust defenses against data poisoning attacks.
• Existing backdoor attacks are typically designed for supervised settings and are unsuitable for unlabeled graph data.

Purpose of the Study:

• To investigate the vulnerability of unsupervised graph learning to backdoor attacks.
• To propose a novel attack method, GRBA, specifically for unsupervised graph learning settings.
• To demonstrate the effectiveness and versatility of the proposed attack across various downstream tasks.

Main Methods:

• Introduced GRBA (Gradient-based Representation Backdoor Attack), a first-order, gradient-based attack.
• Designed GRBA to target node and graph representations directly, without needing downstream task information.
• Applied GRBA to poison unlabeled graph data by introducing triggers in node features and graph structure.

Main Results:

• Demonstrated that unsupervised graph learning models are susceptible to backdoor attacks.
• Showcased GRBA's effectiveness in disrupting learned representations and impacting downstream tasks.
• Validated GRBA's performance on state-of-the-art unsupervised learning models for node and graph-level tasks.

Conclusions:

• Unsupervised graph learning methods present a new attack surface for backdoor adversaries.
• GRBA offers a pioneering and effective method for backdoor attacks in this domain.
• The attack's independence from downstream task specifics highlights its broad applicability and potential risk.