The rapid proliferation of Internet-of-Things (IoT) devices has heightened the need for resource-efficient network monitoring solutions capable of supporting critical tasks such as intrusion detection, application identification, and asset management. While recent neural network–based traffic analysis methods deliver high accuracy, their computational complexity and reliance on specialized hardware hinder widespread deployment on resource-constrained edge gateways. In this paper, we introduce a suite of robust yet lightweight modeling techniques that operate directly on raw IoT traffic packets, eliminating the need for manual feature engineering. We propose simple packet-encoding schemes that capture both temporal and statistical characteristics of network flows and apply four distinct machine learning algorithms—Random Forest, Decision Tree, Multilayer Perceptron (MLP), and XGBoost—to classify network behaviors. Through extensive experiments on benchmark IoT traffic datasets, we demonstrate that our approach achieves classification performance on par with deep learning baselines, while significantly reducing inference latency and memory footprint. Notably, tree-based models offer interpretable decision rules and require only commodity hardware for real-time operation, making them particularly well-suited for deployment in heterogeneous IoT environments. Our findings highlight the practical viability of lightweight ML models for scalable, high-throughput IoT network analytics.

Keywords: IoT network behavior modeling; raw traffic packet encoding; lightweight machine learning; Random Forest; Decision Tree; Multilayer Perceptron; XGBoost; edge deployment; intrusion detection..