Reconstructed Graph Neural Network With Knowledge Distillation for Lightweight Anomaly Detection.

Area of Science:

• Computer Science
• Network Security
• Artificial Intelligence

Background:

• The proliferation of Internet-of-Things (IoT) devices generates massive data, necessitating secure and efficient communication.
• Limited computational power of IoT devices demands lightweight models for intelligent services.
• Ensuring data security and efficient communication in IoT networks is crucial for anomaly detection.

Purpose of the Study:

• To design a lightweight anomaly detection model for IoT communication networks.
• To develop a graph representation learning model integrating graph neural networks (GNN) and knowledge distillation (KD).
• To achieve secure and efficient data exchange in IoT environments.

Main Methods:

• A novel graph network reconstruction strategy is devised, treating data communications as nodes.
• Graph neural network (GNN) and knowledge distillation (KD) techniques are integrated into the reconstructed graph with global-local distillation (RG-GLD) model.
• Graph attention network (GAT), multilayer perceptron (MLP), and self-attention mechanisms are employed for feature extraction and information preservation.

Main Results:

• The RG-GLD model demonstrates effectiveness in lightweight anomaly detection across IoT networks.
• Experiments show improved knowledge transfer efficiency with higher classification accuracy.
• The model achieves a lower computational load compared to four baseline methods.

Conclusions:

• The proposed RG-GLD model is effective for lightweight anomaly detection in IoT environments.
• The model offers a balance between classification accuracy and computational efficiency.
• RG-GLD is suitable for deployment in sustainable IoT computing environments for enhanced security.