Machine-Learning Based Wireless Device Identification

Abstract

Internet of Things (IoT) is becoming increasingly ubiquitous by the day. Nowadays, the adoption of IoT devices has spread widely amongst people and the accurate identification of IoT devices on the network has become a significant area of concern to maintain network security and protect the network from suspicious devices. In this research, ANOVA, XGBoost, Random Forest and Kendall feature selection methods are investigated to see how each type calculates the top performing features. This research also investigates the classification accuracy of four different supervised machine learning algorithms namely, XGBoost, Random Forest, Decision Trees, and k-nearest neighbours (KNN) when feature selection methods are applied. XGBoost was found to be the best performing classifier. After further hyperparameter tuning on the XGBoost classifier, the classification accuracy obtained was 88%. In addition, this research presents a small testbed that was created to capture packets from the network to create a dataset. Four different IoT and non IoT devices were captured and manually labelled to form a dataset. XGBoost was used to train the data and was able to obtain an accuracy of 96.5%.