Anomaly Detection in IoT Sensor Energy Consumption Using LSTM Neural Networks and Isolation Forest
Quôc Vo, Philippe Ea, Selma Benzouaoua, Osman Salem, Ahmed Mehaoua · 2024
The widespread deployment of Internet of Things (IoT) sensors across various domains, necessitates efficient energy management and robust security measures. This paper introduces a novel approach to simultaneously predict energy consumption and detect anomalies in IoT sensor networks, specifically to identify potential security threats. Our key contributions include the development of a machine learning framework that integrates energy consumption forecasting with unsupervised anomaly detection. By employing Long Short-Term Memory (LSTM) networks to model and predict future energy usage based on historical data, we capture temporal dependencies crucial for accurate forecasts. The contribution lies in analyzing the residuals from these predictions using an unsupervised Isolation Forest algorithm, which effectively identifies deviations indicative of abnormal behavior or security breaches. Among various tested models, the combination of LSTM and Isolation Forest yielded the highest accuracy of 0.98 in anomaly detection. Extensive experimental validation on IoT sensor data confirms the efficacy of our approach, particularly in critical applications such as healthcare, where maintaining continuous, secure sensor operation is paramount. Our method not only enhances energy efficiency but also provides a robust mechanism for early threat detection in IoT networks.