Lurker: Backdoor attack-based explainable rumor detection on online media.

Area of Science:

• Graph Neural Networks
• Computational Social Science
• Network Analysis

Background:

• Graph neural networks (GNNs) excel at graph-level classification, including rumor detection.
• Online rumors spread rapidly via manipulated networks (e.g., bots), exhibiting distinct early-stage propagation structures.
• Identifying these structures is key to distinguishing rumors from credible information.

Purpose of the Study:

• To develop an interpretable rumor detection algorithm inspired by backdoor attacks.
• To identify critical online users and their interaction patterns that define rumor propagation structures.
• To transform rumor detection into a problem of detecting specific propagation structures.

Main Methods:

• Utilizing causal analysis to identify the causal subgraph determining graph classification (rumor vs. normal).
• Extracting critical online users and exploring their specific propagation patterns.
• Implementing a backdoor-based approach by planting the identified special propagation structure into the detection model as a trigger.

Main Results:

• The proposed algorithm, Lurker, demonstrates effectiveness in detecting special rumor propagation structures.
• Experimental results on three real-world datasets validate the algorithm's performance.
• Lurker achieved significant improvements over baseline methods, with up to 33.1% increase in attack success rate and 61.8% in clean accuracy drop.

Conclusions:

• The study successfully proposes an interpretable, backdoor-inspired algorithm for rumor detection.
• The method effectively identifies critical users and propagation structures crucial for rumor identification.
• Lurker offers a promising approach for enhancing the accuracy and interpretability of rumor detection systems.