Handling dynamic network behavior and unbalanced datasets for WSN anomaly detection

Abstract

In this paper an extreme learning machine (ELM) based approach is presented for the anomaly detection in dynamic wireless sensor network (WSN) under extremely imbalanced class distribution conditions. The imbalanced class distribution and dynamic nature of the network result in the classifier's inferior performance. The proposed approach tackles the imbalanced classification of datasets by dividing the class with the majority of samples into multiple sub-classes. This is achieved by splitting the majority-class samples evenly into multiple segments using Affinity Propagation (AP). The initial preferences for AP are set to influence the total number of segments such that the number of samples in any segment closely matches with the numbers of minority-class samples. Such sections of the dataset together with the minor class are then viewed as different classes and used to train the ELM. The synthetic datasets required to evaluate the performance of the proposed technique are generated using network simulator version 2 (NS2). The dynamic WSN is simulated for different attacks (considering only one attack at a time) using network simulator version 2 (NS2). During the simulation, all required network data is dumped into trace files. These dumped trace files are later processed through MATLAB to extract the required features in a usable format to generate a proper dataset. This process is repeated for two different WSN configurations, one with lower node density, and the other with higher node density, and the generated datasets for these two WSN configurations are denoted as Dataset-1, and Dataset-2 respectively. The experimental findings support the substantial improvement in performance from previous approaches by the new technique.