Edge Mining on IoT Devices Using Anomaly Detection

With continuous monitoring and sensing, millions of Internet of Things sensors all over the world generate tremendous amounts of data every minute. As a result, recent studies start to raise the question as whether to send all the sensing data directly to the cloud (i.e., direct transmission), or to preprocess such data at the network edge and only send necessary data to the cloud (i.e., preprocessing at the edge). Anomaly detection is particularly useful as an edge mining technique to reduce the transmission overhead in such a context when the frequently monitored activities contain only a sparse set of anomalies. This paper analyzes the potential overhead-savings of machine learning based anomaly detection models on the edge in three different IoT scenarios. Our experimental results prove that by choosing the appropriate anomaly detection models, we are able to effectively reduce the total amount of transmission energy as well as minimize required cloud storage. We prove that Random Forest, Multilayer Perceptron, and Discriminant Analysis models can viably save time and energy on the edge device during data transmission. K-Nearest Neighbors, although reliable in terms of prediction accuracy, demands exorbitant overhead and results in net time and energy loss on the edge device. In addition to presenting our model results for the different IoT scenarios, we provide guidelines for potential model selections through analysis of involved tradeoffs such as training overhead, prediction overhead, and classification accuracy.