Imposters Among Us: A Supervised Learning Approach to Anomaly Detection in IoT Sensor Data

Abstract

Internet of Things (IoT) technology has made smart homes more prevalent in everyday lives. However, anomalies in IoT data may be emblematic of potential cybersecurity risks like false data injection attacks or physical security risks like house fires. In this paper, we propose a segmentation based anomaly detection method that converts unsupervised time-series data into a supervised format, and then trains a Long-Short Term Memory (LSTM) neural network to detect anomalies. The LSTM network is trained to predict un-sound statistical properties, which gets combined with sound statistical properties, to detect anomalies in IoT sensor data. Data smoothing using Holt Winters Exponential Smoothing is also performed, without loss of information, to improve anomaly detector performance. Using Precision, Recall, and F-Measure scores as metrics, results show efficient anomaly detection performance on IoT temperature sensor data. We, additionally, test performance by varying specific parameters. Lastly, results also show that performing data smoothing, to a certain extent, can improve anomaly detection performance over data that didn’t undergo any smoothing.