Detecting IoT Cyberattacks: Advanced Machine Learning Models for Enhanced Security in Network Traffic

Abstract

The IoT is one of the most revolutionary technological advancements of the contemporary era, embedding networked devices into nearly every aspect of human life, from smart homes and wearables to industrial systems and healthcare applications in the U.S.A. The immediate need for better cybersecurity in the U.S.A. arises from the increasing sophistication and frequency of cyberattacks on IoT systems. Machine learning and AI have emerged as promising technologies to deal with the security challenges IoT systems pose. Unlike traditional rule-based systems, ML models learn from large datasets to identify deviations from the normal behavior pattern that signifies malicious activity.  The prime objective of this research is to design, curate, evaluate, and deploy state-of-the-art machine learning models that improve the detection of cyberattacks over IoT network traffic. This research used a well-established dataset that emulates IoT network traffic consisting of benign and malicious activities. Benchmarks like the UNSW-NB15, CICIDS2017, and TON_IoT have been in extensive use by researchers in this domain because they contain a rich variety of network traffic created by various IoT devices and systems along with corresponding labels that classify normal and associated with specific types of cyberattacks: DDOS, MITM, and botnet attacks. Data preprocessing and cleaning ensured that the dataset was consistent, complete, and in a format that helps machine learning algorithms learn from it. Imputation techniques used the feature's mean/median/mode to handle missing values. In this research project, two machine learning algorithms were used in the experiment, notably, Logistic Regression and Random Forest. In this study, the machine learning algorithms used in the experiment undertaken for the current research project are Logistic Regression and Random Forest. The performance of Random Forest was superior to Logistic Regression in almost all metrics. While Logistic Regression provided a strong baseline, it struggled with detecting attacks, as evidenced by its lower recall and higher number of false negatives. This implied that Logistic Regression was less reliable in detecting cyberattacks, which could be critical in real-world cybersecurity settings. By contrast, Random Forest attained impressive accuracy and significantly diminished the number of false negatives. Its higher precision and recall demonstrate that it is better suited for detecting attacks in this dataset, offering a more reliable solution for cyberattack detection.