Graph Convolution Neural Network and Deep Q-Network Optimization-Based Intrusion Detection with Explainability...

Area of Science:

• Cybersecurity
• Artificial Intelligence
• Network Security

Background:

• Traditional network intrusion detection models struggle with the increasing complexity and scale of modern networks and IoT systems.
• Graph Convolutional Networks (GCNs) show promise in analyzing network structures but require enhanced methods to capture nuanced correlations.
• The variability in correlations between network nodes and edges necessitates adaptive weighting for improved accuracy.

Purpose of the Study:

• To develop an advanced intrusion detection model that accurately captures intricate correlations in large-scale networks.
• To enhance the expressiveness and accuracy of GCNs for network intrusion detection tasks.
• To integrate attention mechanisms and deep reinforcement learning for adaptive weight optimization.

Main Methods:

• Proposed the GCN-DQN model, combining GCN with a multi-head attention mechanism and Deep Q Network (DQN).
• Implemented adaptive attention weighting to prioritize nodes and edges based on similarity.
• Validated the model using the UNSW NB15 and CIC-IDS2017 datasets for intrusion detection.

Main Results:

• The GCN-DQN model demonstrated superior classification accuracy compared to baseline models.
• Adaptive weighting significantly improved the model's ability to detect network intrusions.
• Experimental results confirmed the effectiveness of the proposed approach on benchmark datasets.

Conclusions:

• The GCN-DQN model offers a significant advancement in network intrusion detection capabilities.
• Adaptive attention mechanisms are crucial for handling complex network correlations.
• The model's explainability was enhanced using LIME and SHAP techniques.