Feature Engineering Considerations in IoT: A Case Study

Since the emergence and integration of the Internet of Things, homes have become increasingly intelligent and communicative. These connected homes require special control and security from conception, as they can expose people through confidential data sharing and system attacks. To cope with this, Intrusion Detection Systems remain the best solution, despite the need for improvement and adaptation, since these technologies frequently monitor enormous volumes of data flow with unnecessary and duplicated capabilities, this has a detrimental effect on how well they work. Current work on the Internet of Things shows a real willingness on the part of researchers to propose lightweight, accurate IDS-IoTs with reduced functionality. This study aims to provide an overview of the design of a security solution, by experimenting with Feature Engineering extraction and selection techniques. As such, the PCA, IG, ANOVA, LDA and RFE algorithms were evaluated on the TON_IoT dataset. The features obtained for each technique were evaluated through the Random Forest model using indicators, such as ROC, Accuracy, selection time for each algorithm, and training time for each group of features obtained. Each algorithm was trained and evaluated using differ parameters to detect optimal thresholds of 1, 15 and 20. We obtained an accuracy of 98.18% for RFE with a training time of 74.8 s, and the PCA algorithm gave us the best feature acquisition time 2.8 s.