Resource-Efficient Hybrid Ensemble ML Framework for Anomaly Detection in IoT Smart Homes

Abstract

The Internet of Things (IoT) technologies are used to support smart home systems through device and sensor connectivity for data exchange. However, the growth in adoption increases exposure to cyber-attacks and device faults, which puts system reliability and user safety at risk. This study proposes a framework that uses a pre-trained hybrid ensemble model to detect and separate attacks and faults while supporting timely mitigation. Firstly, the study evaluates models on the CICIoT2023 and IntelLab fault-injected datasets using ensemble learning methods and traditional supervised classifiers. Extreme Gradient Boosting shows the strongest intrusion detection performance. Random Forest shows the strongest fault detection performance. Secondly, both models were fine-tuned and combined within a hybrid meta-model. The results show high accuracy, strong F1 scores, and low false positive rates. The framework was implemented as a web application using Flask and Streamlit to support real-time simulations of attack, fault, and normal events. Evaluation reports latency under 5 seconds and memory use under 400 KB, which supports deployment on resource constrained IoT devices. It was optimized using quantisation and compression. The paper proposes a hybrid ensemble approach for joint fault and intrusion detection, a deployable prototype for constrained environments, and methods to enhance model performance.