Skip to Main Content
In this work we evaluate the feasibility of both classical machine learning algorithms and bio-inspired algorithms for misbehavior detection in sensor networks, since recent works in that field seem to concentrate mainly on bio-inspired approaches, without a convincing rational reason. As a first step, we analyze the packet traffic of a simulated sensor network in order to find relevant features that distinguish normal network operation from misbehaving nodes. This kind of data analysis is often missing in previous studies. Using these features acquired by the systematic data analysis we study the suitability of classical machine learning algorithms as well as bio-inspired learning algorithms for the given classification problem. We conclude which algorithms perform best in this special scenario, considering classification success and resource-friendliness of the algorithms. As result we can say that classical algorithms have equal or even better detection capabilities compared to some bio-inspired algorithms. It turns out that it is even possible to detect different levels of misbehavior with nearly 100% accuracy.