Dispelling the Fake: Social Bot Detection Based on Edge Confidence Evaluation

Area of Science:

• Computer Science
• Artificial Intelligence
• Network Security

Background:

• Online social networks (OSNs) face threats from sophisticated social bots.
• Graph Neural Networks (GNNs) are promising for bot detection but struggle with unreliable interaction edges.
• Advanced bots camouflage by interacting with genuine accounts, creating noisy graph structures.

Purpose of the Study:

• To propose a novel social bot detection method, Bot detection method based on Edge Confidence Evaluation (BECE).
• To enhance the reliability of GNN-based bot detection by evaluating edge confidence.
• To improve the differentiation between bot and human representations in OSN graphs.

Main Methods:

• Developed an edge confidence evaluation module to assess and identify unreliable edges.
• Designed edge features based on user node representations.
• Utilized parameterized Gaussian distributions for edge embeddings and Kullback-Leibler (KL) divergence for optimization.

Main Results:

• BECE demonstrated superior effectiveness in social bot detection across three real-world datasets.
• The edge confidence evaluation module improved detection performance when integrated with six widely used GNN architectures.
• The method successfully mitigates the impact of unreliable edges in GNN-based bot detection.

Conclusions:

• BECE offers a robust solution for social bot detection in OSNs.
• The edge confidence evaluation module is a versatile plug-in for enhancing existing GNN models.
• This approach significantly improves the integrity and safety of online social platforms.