Anomaly Detection Using Hybrid Learning for Industrial IoT

Abstract

The industrial internet of things (IIoT) enhances industrial and manufacturing operations by using smart sensors and actuators. However, it is hampered due to the energy efficiency requirements, real time performance requirements in a dynamic environment, and maintaining the security of applications. Security is a serious issue nowadays and is mostly caused by abnormal traffic of some nodes. For detecting those abnormalities, there are two basic machine learning approaches, namely Federated and Centralized Learning. Centralized Learning has better performance, but it has a privacy issue since edge devices send data to the server. On the other hand, Federate Learning obviates privacy issues, but it has less performance due to the resource limitation of edge devices. In this study, a typical hybrid learning based abnormality detection framework has been proposed in which edge devices undertake Federated Learning with a limited number of datasets and the edge server will use the periodically collected aggregated data from edge devices. For security reasons, edge devices share their data after a certain period of time when the time value of the data has declined. We have used Long Short Term Memory (LSTM) Autoencoders with two different datasets (a smaller for edge devices and a larger for the edge server) for the demonstration. The experimental result shows that the size of the dataset affects the predicting performance and resource utilization in an anomaly detection model.