Demystifying machine learning models for IOT attack detection with explainable AI

Abstract

Internet of things (IoT) dramatically is changing our lives with its newly invented devices and applications which leads to various emerging cybersecurity challenges or threats. The rapid growth of IoT arouses security and privacy issues that need more attention to ensure the safety of human personal data, saving from serious damages. Over the year, several techniques have been conducted to establish IoT attack detection model so that it can detect attacks e ciently. Unfortunately, it is di cult to identify a good model that can detect both binary and multi-type attacks with accurately. The prediction result of models provides very little knowledge to the users or experts how the model classify attacks for detection which can not be understood through a simple output. Thus, it is getting necessary to understand the reasons behind the prediction to make people trust on the model by providing the insight view of the model. In the paper, we have introduced Explainable Arti cial Intelligent on our proposed model for making the model faithful enough and human understandable, by explaining the strategy of the detection model for predicting the attacks and representing the features or properties in uence of respective prediction. For this, we have establish an IoT attack detection model by using Xg-boost classi er on a dataset, name, IoT Intrusion Dataset[11], that supports both binary and multi-class classi cation to classify the attacks for detection. We have also used Explainable AI tools, named, Shap, Lime, and ELI5 to validate the performance of the model through analyzing the property of the established model by representing each feature's contribution and action of the model, for each prediction to give a clear idea how e cient the model is, for detecting the IoT attacks.