# Machine learning for Internet of Things (IoT) device identification: a comparative study

## TL;DR

This paper compares machine learning methods for identifying IoT devices, showing that a new algorithm called BGWO reduces features while maintaining high accuracy.

## Contribution

The study introduces and evaluates a Binary Green Wolf Optimizer (BGWO) for IoT device identification, demonstrating its superior feature reduction and classification accuracy.

## Key findings

- BGWO reduced feature sets by 85.11% and 73.33% on two datasets while maintaining high classification accuracy.
- Wrapper-based feature selection methods were shown to be effective in reducing feature sets for IoT device identification.
- BGWO achieved classification accuracies of 98.51% and 99.8% on two widely used datasets.

## Abstract

The rapid deployment of millions of connected devices brings significant security challenges to the Internet of Things (IoT). IoT devices are typically resource-constrained and designed for specific tasks, from which new security challenges are introduced. As such, IoT device identification has garnered substantial attention and is regarded as an initial layer of cybersecurity. One of the major steps in distinguishing IoT devices involves leveraging machine learning (ML) techniques on device network flows known as device fingerprinting. Numerous studies have proposed various solutions that incorporate ML and feature selection (FS) algorithms with different degrees of accuracy. Yet, the domain needs a comparative analysis of the accuracy of different classifiers and FS algorithms to comprehend their true capabilities in various datasets. This article provides a comprehensive performance evaluation of several reputable classifiers being used in the literature. The study evaluates the efficacy of filter-and wrapper-based FS methods across various ML classifiers. Additionally, we implemented a Binary Green Wolf Optimizer (BGWO) and compared its performance with that of traditional ML classifiers to assess the potential of this binary meta-heuristic algorithm. To ensure the robustness of our findings, we evaluated the effectiveness of each classifier and FS method using two widely utilized datasets. Our experiments demonstrated that BGWO effectively reduced the feature set by 85.11% and 73.33% for datasets 1 and 2, respectively, while achieving classification accuracies of 98.51% and 99.8%, respectively. The findings of this study highlight the strong capabilities of BGWO in reducing both the feature dimensionality and accuracy gained through classification. Furthermore, it demonstrates the effectiveness of wrapper methods in the reduction of feature sets.