Anomaly Detection in WBANs Using CNN-Autoencoders and LSTMs

Abstract

Wireless Body Area Networks (WBANs) are wireless networks that consist of microscopic sensors that are either placed on a subject’s body or attached to their clothes. These low-power devices track the physiological readings from a subject and relay them to a server over the Internet. Since WBANs find various uses, from remote medical monitoring to early disease detection in the medical field, their readings must be accurate. Anomalies can occur in the data being transmitted for various reasons, such as sensor faults or malicious attacks.

In some instances, however, what is considered an anomaly may not be an anomaly, and the readings are correct. To mitigate such cases, we must detect anomalies promptly. The anomaly detection process developed here consists of point and contextual anomaly detection. Point anomaly detection deals with checking whether, given a set of readings, some of the readings are anomalous in an isolated sense. Contextual anomaly detection then checks whether readings from other sensors can corroborate the reading of our “faulty” sensor. This paper has employed a CNN Autoencoder and an LSTM-based classifier to do this two-step process. Our model shows an accuracy of 94% and a loss of 14%.