Federated learning for network attack detection using attention-based graph neural networks.

Area of Science:

• Cybersecurity
• Machine Learning
• Network Security

Background:

• Federated learning (FL) addresses data isolation and privacy concerns in machine learning.
• Network attacks pose a significant threat to the security of FL systems.
• Existing methods struggle to detect sophisticated cross-level and cross-department attacks within FL architectures.

Purpose of the Study:

• To propose an effective method for detecting network attacks in federated learning environments.
• To enhance the security of network devices and architectures deploying federated learning.
• To enable collaborative model training while preserving data privacy.

Main Methods:

• Developed an attention-based Graph Neural Network (GNN) for network attack detection.
• Organized network traffic data chronologically and constructed graph structures based on log density.
• Introduced a Federated Graph Attention Network (FedGAT) model incorporating an attention mechanism to analyze node interactivity.

Main Results:

• The proposed FedGAT model accurately detects cross-level and cross-department network attacks.
• The method achieves accuracy and robustness comparable to traditional detection techniques.
• Experimental results validate the effectiveness of the FedGAT model in enhancing network security within FL.

Conclusions:

• The FedGAT model offers a robust solution for detecting network attacks in federated learning.
• This approach prioritizes privacy protection and data security in distributed machine learning environments.
• The attention-based GNN effectively improves the precision of internal network interaction analysis for attack detection.