Scalable and Robust Outlier Detector using Hierarchical Clustering and Long Short-Term Memory (LSTM) Neural Network for the Internet of Things

Abstract

The emerging centralized entities, like cloud, edge, or Software-Defined Network (SDN), make automated decisions for the Internet of Things (IoT) applications based on the measured data from several sensors. However, the malicious injection of the anomalies or outliers in measured sensor data may disrupt the automated decision making capabilities of the applications running at the centralized location. Therefore, the detection of such outliers is an essential problem for IoT that needs to be researched out. This paper presents a scalable outlier detector that uses Hierarchical clustering in conjunction with Long Short-Term Memory (LSTM) neural network. Hierarchical clustering provides scalability to the outlier detector by finding correlated sensors. The LSTM neural network is coupled with the robust statistics, M-estimator, to accurately detect outliers in time-series data. The simulation results on different data-sets show that the proposed method has an accuracy of more than 90% for different attack strength. Also, the model parameter can be tuned according to the application requirement so that the outlier detector can be tailored to either precision or recall sensitive.