Anomaly Detection in Streaming Environment by Evolving Neural Network with Interim Decision

Abstract

Algorithms in a streaming environment recognise a concept in real time since patterns are presented continuously; yet, in the presence of anomalies, algorithms fail to recognise the underlying concept, making anomaly identification a critical task. Neural networks detect anomalies efficiently; however, to use neural networks in a streaming environment, the architecture must be adaptive to learn ideas that vary over time. The proposed architecture places each node of the neural network in individual sites, and each node is made up of a perceptron. These perceptrons have two activation functions: one is used for architecture training, while the other is utilised for classification. Each layer has a decision-making node that takes decisions from the last layer nodes and decides the anomalous nature by majority voting. These layer-wise decisions help to select the training phase and appropriate architecture based on the desired performance. To demonstrate its usefulness, the proposed structure is tested on real-world data sets and compared to conventional and alternative neural networks.