An Evaluation of Machine Learning Models for Threat Classification in IoT Devices

This study presents a comparative analysis of machine learning models for threat detection in Internet of Things (IoT) devices using the CICIoT2023 dataset.We evaluate Logistic Regression, K-Nearest Neighbors, and Random Forest algorithms across three classification granularities: binary (benign vs. attack), multiclass (8 categories), and fine-grained (34 subtypes).Our methodology incorporates comprehensive preprocessing including feature engineering, variance thresholding, correlation filtering, and dimensionality reduction.Performance assessment focuses on accuracy, precision, recall, and F1-score, along with model scalability when trained on small datasets and tested on larger ones.Results demonstrate that Random Forest consistently outperforms other models across all classification tasks (binary: F1 = 0.710, 8-class: F1 = 0.629, 34-class: F1 = 0.590).All models show performance degradation as classification granularity increases, with notable challenges in detecting BruteForce and Web attacks.Feature importance analysis reveals protocol-specific characteristics and TCP flag information as crucial for attack identification.Scalability testing indicates significant performance decline when models trained on limited data (0.1%) are applied to larger datasets (0.5%, 1%), though Random Forest demonstrates superior generalization capabilities.An unsupervised autoencoder approach achieves moderate success for anomaly detection (accuracy = 0.881) but struggles with recall (0.070).These findings highlight the trade-off between detection granularity and accuracy in IoT security implementations and suggest hierarchical classification approaches for resource-constrained environments.The study provides valuable guidance for selecting appropriate machine learning techniques for real-world IoT security applications.