Cognitive Topology Learning for IoT: Interpretable and Fair Device Classification with Continuous Behavioral Insight

Abstract

The rapid proliferation of Internet of Things (IoT) devices in smart environments has introduced significant challenges in identifying and securing heterogeneous network entities. Traditional rulebased classification systems are often inadequate due to the diversity and dynamic behavior of IoT devices. To address this issue, this study proposes a machine learning–driven IoT device classification platform that leverages network traffic fingerprints to accurately categorize devices. The system analyzes behavioral features such as packet frequency, byte count, and communication duration to map devices to their respective profiles. A key challenge in IoT classification is class imbalance, where frequently occurring devices dominate the dataset while rare devices are underrepresented. To mitigate this issue, the Adaptive Synthetic Sampling (ADASYN) technique is employed to generate synthetic samples for minority classes, ensuring balanced and unbiased model training. The proposed framework includes a comparative machine learning pipeline with classifiers such as Gaussian Naive Bayes (GNB), Multinomial Naive Bayes (MNB), Decision Tree Classifier (DTC), and a novel Greedy Tree Classifier (GTC). The GTC model, optimized using interpretable modeling techniques, achieves a balance between high classification accuracy and explainability through rule-based decision structures. Implemented using a modular Flask-based architecture, the platform provides both exploratory data analysis and real-time inference capabilities. Experimental results demonstrate that the GTC model outperforms baseline approaches in terms of F1-score, effectively distinguishing various IoT devices. The system offers a scalable and automated solution for enhancing network management and IoT security.