Exploring optimal set of features in machine learning for improving IoT multilayer security

The growing use of Internet of Things (IoT) systems in industry and healthcare has raised serious concerns about their security because IoT-based devices are highly vulnerable to multilayer attacks. Alongside cross-layer interactions, machine learning methods are used for detecting IoT multilayer attacks. However, there is a lack of research findings to explore the optimal set of feature conditions in the machine learning process for detecting multilayer IoT attacks. In this paper, we have incorporated several feature selection methods and hyperparameter tuning of classification algorithms to optimize the overall process. The paper also presents a detailed strategy consisting of data collection, pre-processing, feature selection, dataset splitting, and binary classification. A range of feature selection techniques such as mutual information, information gain, decision tree entropy, correlation, chi-square, and principal component analysis (PCA) is implemented to identify the most significant features. The performance of the classification models is evaluated using different feature sets with standard (70:30, train: test) dataset-splitting scenarios. The results demonstrate that the information gain feature selection method with the highest 31 score features is effective in improving the accuracy of machine learning models in the field of multilayer attack detection in IoT networks. The Artificial Neural Network (ANN) model, which is a powerful deep-learning model, achieved the highest accuracy of 98.88%. The Decision Tree model and the Naïve Bayes model performed lower, however, they may still be useful for certain applications.