We designed, implemented, and evaluated FEATS to analyze sensor-related data and assess the trustworthiness of sensors within a sensor network. FEATS allows a sensor network analyst to track the trustworthiness of sensors without the need for any dedicated diagnostic channels. Instead, data classifiers compare new data to previously observed behavior, to identify anomalous and potentially untrustworthy behavior. The framework enables rapid customization of classifiers for a specific sensor network deployment. To increase overall classification accuracy, multiple classifiers can be combined in parallel or series. This versatile approach to classifier composition also allows a FEATS-based application to compare results between sensors and groups of sensors, which allows the application to scale to very large networks. A trustworthiness assessor aggregates the output of the classifiers to infer the trustworthiness of each sensor or group of sensors over time. Because the trustworthiness of a sensor may need to be inferred from multiple classifiers, the assessor performs a weighted aggregation of classifier results based on the confidence of each classification. Finally, a human-computer interface allows the network analyst to monitor the sensor trustworthiness or any other data being communicated across the framework.

The figure shows an overview of the FEATS architecture. Classifiers at each node (e.g., sensor, hub) monitor incoming data for anomalous events. In this example, the topology is hierarchical, so each classification tier processes data output from the tier beneath it. Classifiers at higher levels can detect patterns of behavior across multiple sensors or groups of sensors.