Energy Efficient Distributed Anomaly Detection using Semi-Supervised Models in IoT

Utilizing IoT technologies for monitoring large-scale smart facilities such as power, water and gas distribution networks has been the subject of many studies recently. The aim is to detect anomalous events in the network due to elements’ failure, bad designs, attacks or abuses of the network and alert the network operators in a timely manner. As the centralized cloud-based approaches are impractical in time-critical and real-time anomaly detection applications due to 1) high sensor-to-cloud transmission latency 2) high communication cost and 3) high energy consumption at the sensor nodes, the distributed anomaly detection methods based on Deep Neural Networks (DNN) have been applied in past studies vastly. In these methods, in order to detect anomalies in real-time, copies of the anomaly detection model are placed at the sensor nodes (rather than placing one at the cloud node) reducing the sensor-to-cloud transmissions significantly. Nevertheless, new normal samples collected at the sensor nodes still need to be transmitted to the cloud node at predefined intervals to re-train the distributed anomaly detection DNNs. In order to minimize these sensor-to-cloud transmissions during the retraining process, in this paper, two well-known lossless coding algorithms: Huffman Coding and Arithmetic Coding were studied and it was observed that the Huffman and Arithmetic Coding were able to reduce the transmission traffic up to 50% and 75% respectively using two IoT benchmark datasets of pipeline measurements. Besides, the Huffman Coding shown to be computationally feasible on resource limited sensors and resulted in up to 10% saving in energy consumption on each sensor resulting in longer network longevity. Moreover, the experimental results showed that the auto-encoder DNN could outperform the one-class SVM in the iterative distributed anomaly detection method.