Comparative Analysis of Malicious Traffic Detection in Iot Network Using Machine Learning and Deep Learning Approaches
Rosya Satria Firdhaust, Kurnia Mustika Wati, Alfiah Zalfa Tsabitah, Inung Wijayanto · 2025
The rapid adoption of the Internet of Things (IoT) has introduced new security challenges, as resourceconstrained devices are increasingly susceptible to cyberattacks. This study compares the performance of Support Vector Machine (SVM) and Multi-Layer Perceptron (MLP) models in detecting malicious traffic using the IoT-23 dataset. To address the issue of class imbalance, the Synthetic Minority Oversampling Technique (SMOTE) was applied prior to training the model. Model performance was evaluated using accuracy, precision, recall, and$\mathbf{F 1}$-score. The results show that both models achieved 90 % accuracy, with the MLP slightly outperforming the SVM in precision at$\mathbf{9 4 \%}$compared to$\mathbf{9 3 \%}$. However, both models demonstrated low recall for benign traffic (30 %), indicating difficulty in correctly identifying nonmalicious samples. This limitation, often caused by overlapping features and imbalance, highlights the need for further optimization. Potential improvements include adjusting the decision threshold, implementing hybrid sampling methods, and utilizing cost-sensitive learning. The novelty of this work lies in its integration of deep learning with oversampling strategies tailored explicitly for IoT anomaly detection. Additionally, the study provides empirical justification for the architectural configuration of the MLP model and highlights its relevance for addressing security challenges in emerging innovative environments. These findings contribute to the growing body of research on machine learning applications in IoT security, demonstrating the potential of hybrid approaches to enhance detection reliability in real-world scenarios.