Data-centric anomalies in sensor network deployments: analysis and detection

Recent real-world sensor network deployments have helped decision makers and sensor data analysts draw highly precise conclusions by relying on finer-grained raw sensory data. A small number of anomalous sensor readings, however, lead to misinterpretations and false conclusions. Hence, detecting and eliminating anomalies in sensor network deployments play a major role during the decision making process by assuring sensor data quality, integrity, and trustworthiness. A sensor network should be able to accurately identify malicious and faulty observations in a timely manner while maintaining a low communication overhead. In this work, we first perform an analysis of anomalies prevalent in sensor network deployments and then propose a distributed data-centric anomaly detection framework for sensor networks where each node in the collection tree is asked to maintain a statistical data summary over a specified period of time. Rather than transmitting raw sensory data, each node only communicates the behavioral summary to the base station. The base station, in turn, learns the behavior of each sensor node and instructs parent `intermediate´ nodes in the tree to stop forwarding readings of misbehaving nodes. Given the summaries obtained from all sensor nodes, learning can be performed using existing machine learning techniques. We show that boosting is a good candidate for data classification in sensor networks over a short period of time compared to SVM.