SNMiner: A Rapid Evaluator of Anomaly Detection Accuracy in Sensor Networks

Modeling faults and malicious activities in sensor networks can be challenging. Designing and re-evaluating a “good” classifier to detect abnormalities imposes yet another challenge once the sensor network is deployed in the field. Common approaches among researchers involve obtaining publicly accessible labeled datasets, generating synthetic sensor data, or collecting sensor readings from a real-world deployment. Then, an anomaly detection/classification algorithm is applied over the dataset and its accuracy is further validated (after labeling data instances of unknown classes). These approaches, however, do not directly take into account the emergence of new types of abnormalities under the dynamics of a deployed sensor network. In order to incorporate new fault models, it is essential to reconfigure the network without interrupting its operation. To this end, we introduce SNMiner, a tool that facilitates fault modeling, and constantly re-evaluates the accuracy of existing classifiers or detection algorithms over real-time sensor streams.