An accurate IoT Intrusion Detection Framework using Apache Spark

TL;DR

This paper presents an IoT intrusion detection system built on Apache Spark, demonstrating high accuracy in classifying attacks using MLlib with the BoT-IoT dataset.

Contribution

It introduces a scalable IoT intrusion detection framework leveraging Apache Spark and MLlib, with comprehensive evaluation on partial and full datasets.

Findings

Random Forest achieved 99.7% f1 in binary classification on partial data.

Decision Tree scored 97.9% f1 in binary classification on full dataset.

High accuracy in classifying IoT attacks demonstrates effectiveness of the proposed system.

Abstract

The internet has caused tremendous changes since its appearance in the 1980s, and now, the Internet of Things (IoT) seems to be doing the same. The potential of IoT has made it the center of attention for many people, but, where some see an opportunity to contribute, others may see IoT networks as a target to be exploited. The high number of IoT devices makes them the perfect setup for staging denial-of-service attacks (DoS) that can have devastating consequences. This renders the need for cybersecurity measures such as intrusion detection systems (IDSs) evident. The aim of this paper is to build an IDS using the big data platform, Apache Spark. Apache Spark was used along with its ML library (MLlib) and the BoT-IoT dataset. The IDS was then tested and evaluated based on F-Measure (f1), as was the standard when evaluating imbalanced data. Two rounds of tests were performed, a partial dataset for minimizing bias, and the full BoT-IoT dataset for exploring big data and ML capabilities in a security setting. For the partial dataset, the Random Forest algorithm had the highest performance for binary classification at an average f1 measure of 99.7%, as well as 99.6% for main category classification, and an 88.5% f1 measure for sub category classification. As for the complete dataset, the Decision Tree algorithm scored the highest f1 measures for all conducted tests; 97.9% for binary classification, 79% for main category classification, and 77% for sub category classification.