FedKG: A Knowledge Distillation-Based Federated Graph Method for Social Bot Detection

Area of Science:

• Computer Science
• Cybersecurity
• Artificial Intelligence

Background:

• Malicious social bots threaten social network security by spreading misinformation.
• Data scarcity and labeling costs hinder centralized bot detection.
• Federated learning offers a decentralized approach to train models without sharing raw data.

Purpose of the Study:

• To develop an effective federated social bot detection model.
• To address data heterogeneity challenges in federated learning for bot detection.
• To improve the accuracy and efficiency of detecting malicious social bots.

Main Methods:

• Combined federated learning with Relational Graph Convolutional Neural Networks (RGCN).
• Utilized class-level cross-entropy loss for local model training to handle class imbalance.
• Applied knowledge distillation techniques, including a global generator and server-side knowledge integration, to manage data heterogeneity.

Main Results:

• The proposed approach demonstrated effectiveness in social bot detection within heterogeneous data scenarios.
• Achieved a 3-10% improvement in detection accuracy compared to baseline methods, especially with higher data heterogeneity.
• Reached specified accuracy with minimal communication rounds, indicating efficiency.

Conclusions:

• The federated RGCN model with knowledge distillation is a robust solution for social bot detection.
• The method successfully mitigates challenges posed by data imbalance and heterogeneity.
• This approach offers a promising direction for enhancing social network security against malicious bots.