An Efficient Framework for Unsupervised Anomaly Detection over Edge-Assisted Internet of Things

Abstract

The detection of anomaly status plays a pivotal role in the maintenance of public transportation and facilities in smart cities. Owing to the pervasively deployed sensing devices, one can collect and apply multi-dimensional sensing data to detect and analyze potential anomalies and react promptly. Current efforts concentrate on offline manners and fail to fit the situation in smart cities, where efficient and online solutions are expected. In this paper, a novel framework is designed for anomaly detection over edge-assisted Internet-of-Things (IoTs). The framework allows periodical data collection from sensors and continuous anomaly detection at the edge node. A novel efficient and unsupervised deep learning model is designed to balance the resource consumption and accuracy for anomaly detection, based on the combination of a convolutional autoencoder and adversarial training. Meanwhile, the proposed framework also adopts an adaptive strategy for continuous anomaly detection to reduce the overall resource consumption. According to theoretical analysis and evaluation on several real-world datasets, the proposed framework can discover the potential correlation features among multi-dimensional sensing data, and efficiently detect the abnormality of public transportation and facilities in smart cities.