A Comparative Study for IoT Attack Detection Using Machine Learning Algorithms

Prior studies on behaviour-based threat detection on Internet of Things (IoT) device networks have generated machine-learning models with a limited and frequently unproven capacity to learn from unseen data. In this study, we provide a generalizability-focused modelling technique for IoT network assaults that also improves detection and performance. Firstly, we develop a multi-step feature selection technique that minimizes overfitting and provides an enhanced rolling window strategy for feature extraction. Second, in order to prevent frequent data leaks that have restricted the generalizability of earlier models, we develop and test our models using separate train and test datasets. Third, we employ a wide range of machine learning models, assessment measures, and datasets to assess our approach thoroughly. Lastly, we employ explainable AI approaches to strengthen the models' confidence, enabling us to pinpoint the characteristics that support precise attack detection. Models are updated gradually by use of algorithms such as Online Naive Bayes and Stochastic Gradient Descent (SGD).