IoT Intrusion Detection using Auto-Encoder and Machine Learning Techniques

Abstract

IoTnetwork refers to the capability of connecting smart and various devices to asingle network for the sake of performing a particular task. Similar toconventional networks, IoT networks are vulnerable to several attacks.Therefore, IoT Intrusion Detection has caught much research attention. Severalstudies have examined the task of intrusion detection for IoT networks. Withinsuch studies, the focus was set to accommodate a feature selection process foridentifying the most relevant features per the intrusions. Yet, the featureselection techniques used in the literature were based on feature selectionrather than a reduction in which individual solutions are being selected. Thiscould lead to a fall in local minima problems where the optimal solution is notdetermined but instead, another near-optimal solution is identified. This studyproposes a dimensionality reduction approach rather than feature selectionusing Auto-Encoder architecture for IoT intrusion detection. A benchmarkdataset of UNSW-NB15 has been used within the experiment. In addition, a datapreparation process of feature transformation has been applied to convert thecategorical features into numeric ones. Then, the proposed autoencoder has beencarried out upon the transformed data for the sake of dimensionality reduction.The reduced dimension produced by the proposed autoencoder has been utilized byfour classifiers including DT, LR, NN, and RF for conducting the intrusiondetection. Results showed that the proposed autoencoder with RF classifier hasobtained the highest F1-score of 99% and the lowest FAR value of 0.78%. Suchresults are competitive in terms of the state of the art.