Key points are not available for this paper at this time.
Message Queuing Telemetry Transport (MQTT) is essentialfor resource-constrained Internet of Things (IoT) environments; however, its widespread adoption has introduced significant security vulnerabilities. Although machine learning (ML) offers a promising solution for anomaly detection, existing models are often hindered by unrealistic data, severe class imbalances, and high computational costs. To address these limitations, we present a comprehensive ML framework for MQTT anomaly detection benchmarked on MQTTEEB-D, a high-fidelity dataset from a physical IoT testbed. Our framework evaluates a diverse suite of algorithms, including tree ensembles and boosting methods, on both original imbalanced and balanced data. We assessed performance using standard metrics, imbalance-stable metrics such as the Matthews Correlation Coefficient (MCC), and a Performance-Efficiency Score (PES) to quantify the trade-off between predictive power and computational cost. Our results establish a new state-of-the-art, with the top models achieving over 98.8% accuracy and F1-score. These models also yielded dramatic efficiency gains, including a 43-fold reduction in training time and a 299-fold speedup in inference latency over previous benchmarks. Critically, we found that a model’s resilience to class imbalance is more vital for real-world deployment than its peak performance on artificially balanced data. Simpler tree-based models remained robust under imbalanced conditions, where more complex algorithms failed. These findings provide a new benchmark and reorient model selection toward efficient, reliable, and deployable IoT security systems.
Allaga et al. (Tue,) studied this question.