Dependability Evaluation of Wireless Sensor Networks

Abstract

The performance of a network is dependent on the qualitative and quantitative features, which are closely tied to the Quality of Service (QoS). The QoS determines the characteristics of a network required for its effective functioning. The QoS encompasses many aspects of the network such as dependability, scalability, fault recovery, energy efficiency, packet loss ratio. The most important aspect of QoS is dependability and hence dependability evaluation of a network is obligatory to investigate the perilous aspect that affects the faultless functioning of the network. This research work focuses on the reliability and security aspect of dependability. Reliability is defined as the “measure of the continuity of correct service”. It is the most quantifiable feature of network design. Security is defined as the “judgment of how likely it is that the network can resist accidental or deliberate intrusions”.

The Wireless Sensor Networks (WSNs) are capable of monitoring the dynamically changing environment in a particular timespan. The data collected by this network consists of unexpected and complex patterns. To understand these patterns, Machine Learning plays a vital role. ML algorithms facilitate in discovering important correlations in the collected data and hence provide improved deployment strategy. The main focus of this research work is dedicated to the analysis of various dependability evaluation techniques in WSN. Also, an ML-based framework is proposed to optimize the data flow parameter of the network. The data flow is a vital parameter that affects the QoS of a network.

This dissertation proposes a novel ML-based framework, which predicts the overall reliability of WSN in terms of performance metrics such as, sent packets, received packets, packets forfeit, packet delivery ratio, and throughput. Ten Machine Learning models namely, Cubist, Random Forest, Support Vector Machine (SVM), Neural Networks, Weka Lazy Model, Conditional Inference Tree, Bayesian Regularized Neural Networks, Bagged Multivariate Adaptive Regression Splines, Bagged Classification and Regression Trees, and Tree Model from Genetic Algorithms are used to predict the Data Flow, Number of Nodes and Protocol Name. Also, an ensemble model is proposed, which yields an optimum result for prediction.

Further, we considered the security aspect taking into account the flooding attack on the WSN. The node deployment has been carried out in two ways: randomized and normalized deployment. Also, an Intrusion Detection System (IDS) is designed to diagnose any suspicious activity in the network traffic flow. This IDS analyzes the traffic patterns both for the randomized and normalized deployment of sensor nodes. Different Machine Learning approaches namely, Linear Tree, Decision Tree, Extreme Learning Machine, Tree Model from Genetic Algorithms, Generalized Additive Model, Model Tree, Projection Pursuit Regression, Bayesian Regularized Neural Network, PartyKit, Generalized Linear Model, and Linear Regression are used for predicting data flow patterns. These models perform differently according to the training-testing partition size.

Also, traffic flow prediction has been carried out with the help of intelligent soft computing techniques such as, Neural Network, Bayesian Regularized Neural Network, Neural Network using Model Averaging, Multi-Layer Perceptron, Multi-Layer Perceptron with Multiple Layers, Quantile Regression Neural Network, and Stacked Autoencoder Deep Neural Network. These methods prove to be very effective and substantially enhance the prediction efficiency.