GAT-SFFS: A Novel Feature Selection Method Using Graph Attention Networks and Sequential Forward Approach for Machine Learning-Based IoT Intrusion Detection System

The rapid expansion of the Internet of Things (IoT) has greatly enhanced convenience and service quality in modern society. However, the proliferation of IoT devices results in massive daily data collection, leading to significant information transmission and privacy challenges. Recognizing the urgent need for effective cyber attack mitigation in IoT networks, we propose a novel feature selection method called GAT-SFFS, which combines the strengths of Graph Attention Networks (GAT) and Sequential Forward Feature Selection (SFFS). This method is designed to extract the optimal feature set (OFS) from the TON-IoT dataset, which is collected from practical IoT environments and used to develop an effective Intrusion Detection System (IDS). The selected OFS is also validated using multiple machine learning classifiers and compared with previous works to assess the efficiency of the proposed method. Our results show that the XGBoost algorithm, utilizing a combination of only 13 features, achieves the highest detection accuracy at 98.91%, surpassing other current approaches. Furthermore, the OFS enhances detection performance and reduces training time. Numerical results demonstrate that the IDS developed with the OFS exhibits exceptional performance, underscoring its potential to address network security challenges in IoT environments effectively.