CTOD: Collaborative Tree-Based Outlier Detection in Wireless Sensor Networks

Outlier detection is a well studied problem in various fields. The unique challenges of wireless sensor networks such as limited bandwidth, memory, energy, and unreliable communication make this problem especially difficult. Sensors can detect outliers for a plethora of reasons and these reasons need to be inferred in real time. Here, we present a new robust communication framework for the unsupervised in-network detection of outliers in a wireless sensor network. First, communication is minimized through an ad-hoc collaborative communication scheme which controls sensor behavior to increase overall visibility of individual streaming data sets. Second, an outlier detection algorithm is tailored to fit within this communication model. At the same time, minimal assumptions are made about the nature of the data set as to minimize the loss of generality in the architecture. We also build on our previous foundation to introduce the concept of trust to model anomalous behavior caused by security compromises and hardware failures. We not only prove the convergence of our method but also evaluate the performance which shows the usefulness of our model in comparison to other recent work.