A Hybrid Approach to Feature Selection for Cyber Threat Detection in IoT Networks

The quick expansion of Internet of Things (IoT) devices have greatly impact multiple sectors, including healthcare, smart agriculture, and environmental monitoring, thereby impacting human lives. The heterogeneous nature and limited resources of these devices make them highly vulnerable to cyber threats. To address these security challenges, machine learning plays a significant role in the development of adaptive intrusion detection systems. Machine learning helps in identifying patterns from the vast amounts of data produced through different IoT devices, particularly for enhancing security. It facilitates in identifying assaults by analyzing trends and abnormalities in data that indicate threats, offering a proactive strategy for cyber security. In this study, we have explored the impact of different Methods of feature selection on the effectiveness of machine learning models in detection of IoT attacks. Feature Selection Methods such as Principal Component Analysis (PCA), Autoencoders (AE), Linear Discriminant Analysis (LDA) and a hybrid selection method combining SelectKBest and Recursive Feature Elimination (RFE) are employed to asses different machine learning models. The UNSW-NB15 dataset is used to evaluate the effectiveness of Random Forest (RF), XGBoost and Support Vector Machine (SVM) classifiers. The experimental results indicate that the hybrid feature selection approach considerably improves classification efficiency, particularly for the RF model which outperform the other having accuracy 99.91%, surpassing traditional dimensionality reduction techniques. The suggested approach effectively eliminates redundant attributes while retaining the most informative elements, leading to improved detection accuracy and other performance metrics. These findings underscore the practical applicability of hybrid feature selection techniques in cyber security, offering a scalable and efficient solution for IoT attack detection.