BANBAD
Anomaly detection is an important aspect of any security mechanism. We present an efficient anomaly detection algorithm, named BANBAD. Using Belief Networks (BNs), the algorithm identifies abnormal behavior of a feature, like inappropriate energy consumption of a node in a network. By applying structure learning techniques to training dataset, BANBAD establishes a joint probability distribution among relevant features, such as average velocity, displacement, local computation and communication time, energy consumption, and response time of a node of the network. A directed acyclic graph (DAG) is used to represent the features and their dependencies. Using a training process, BANBAD maintains dynamic, updated profiles of network node behaviors and uses specific Bayesian inference algorithm to distinguish abnormal behavior during testing. BANBAD works especially well in ad hoc networks. Extensive simulation results demonstrate that a centralized BANBAD achieves low false alarm rates, below 5%, and high detection rates, greater than 95%. We also show that BANBAD detects anomaly efficiently and accurately in two real datasets. The key for achieving such high performance is bounding the false alarm rate at certain predefined threshold value. By fine-tuning at the threshold, we can achieve high detection rate as well.