Evidence theory based Machine Learning approaches for network security

In recent years, there has been an immense research interest in applying Machine Learning (ML) for defending networked systems from cyber threats. Two particular challenges in this domain are i) the identification and selection of appropriate features that ensure prompt and correct cyber threat detection and, ii) increasing the robustness of ML models against Adversarial Machine Learning (AML) attacks. AML attacks refer to changing, or perturbing, the input data of an ML model in such a way as to change the output to benefit the attacker.

The main contributions of this thesis are the application of evidential ML classifiers in network security environments for Feature Selection (FS) and in addressing AML attacks. This thesis critically analyses an evidential approach to FS against state-of-the-art approaches, such as Decision Tree (DT), Random Forest (RF), L1 Regularisation (Lasso) and Analysis of Variance (ANOVA), by augmenting a Logistic Regression (LR) model to measure uncertainty within each individual feature. This thesis also analyses an evidential approach to minimise the effect of AML perturbation attacks on a Neural Network (NN) model with evidence theory, to indicate when feature values deviate through perturbation.

The results of the experiments in this thesis have shown that an evidential approach to FS can create a feature subset that matches or improves on the state-of-the-art, as well as performing the training and testing in a faster time frame. The results indicate an F1 Score of 0.99 on a large, realistic network security dataset, while performing the classification in a third of the average time required across the state-of-the-art. Furthermore, an evidential approach to address the challenges of AML attacks has shown a decrease of the misclassification rates on the two perturbed malicious classes from 70.53% to 13.09%, and from 99.67% to 1.33%, respectively.