Machine-Learning-Based Darknet Traffic Detection System for IoT Applications

Abstract

The massive modern technical revolution in electronics, cognitive computing, and sensing has provided critical infrastructure for the development of today’s Internet of Things (IoT) for a wide range of applications. However, because endpoint devices’ computing, storage, and communication capabilities are limited, IoT infrastructures are exposed to a wide range of cyber-attacks. As such, Darknet or blackholes (sinkholes) attacks are significant, and recent attack vectors that are launched against several IoT communication services. Since Darknet address space evolved as a reserved internet address space that is not contemplated to be used by legitimate hosts globally, any communication traffic is speculated to be unsolicited and distinctively deemed a probe, backscatter, or misconfiguration. Thus, in this paper, we develop, investigate, and evaluate the performance of machine-learning-based Darknet traffic detection systems (DTDS) in IoT networks. Mainly, we make use of six supervised machine-learning techniques, including bagging decision tree ensembles (BAG-DT), AdaBoost decision tree ensembles (ADA-DT), RUSBoosted decision tree ensembles (RUS-DT), optimizable decision tree (O-DT), optimizable k-nearest neighbor (O-KNN), and optimizable discriminant (O-DSC). We evaluate the implemented DTDS models on a recent and comprehensive dataset, known as the CIC-Darknet-2020 dataset, composed of contemporary actual IoT communication traffic involving four different classes that combine VPN and Tor traffic in a single dataset covering a wide range of captured cyber-attacks and hidden services provided by the Darknet. Our empirical performance analysis demonstrates that bagging ensemble techniques (BAG-DT) offer better accuracy and lower error rates than other implemented supervised learning techniques, scoring a 99.50% of classification accuracy with a low inferencing overhead of 9.09 µ second. Finally, we also contrast our BAG-DT-DTDS with other existing DTDS models and demonstrate that our best results are improved by (1.9~27%) over the former state-of-the-art models.