Abstract
Recent advances in artificial intelligence (AI) have seen generative modelling evolve into a
significant research area. The ability to train algorithms using AI techniques to explore the
fundamental structure of a dataset to generate realistic samples has influenced various fields of
research, such as computer vision, natural language processing, procedural content generation,
computer graphics and cybersecurity. Apart from generating creative art or language, a
significant amount of research in generative modelling explores the generation of synthetic
samples to address the class imbalance inherent in some datasets. Class imbalance is a
phenomenon where the number of samples for one or more classes in a dataset is significantly
less than the average number of samples per class. Class imbalance is a significant problem in
machine learning because it impedes a model from efficiently generalising the classes from the
dataset. One area of research where class imbalance is common in datasets is intrusion
detection. The fundamental reason for a significant amount of class imbalance in intrusion
detection is that the amount of attack traffic in the real world is significantly less than that of
normal traffic. Therefore, if data is collected from such an environment for training an intrusion
detection model, the data would be highly imbalanced due to the low number of samples
representing attack traffic. Such is also true in Industrial Internet of Things (IIoT)
environments. This thesis document investigates the realisation of a novel immunologically
inspired generative model for synthetic data generation. The proposed model is called a
generative adversarial artificial immune network (GAAINet). GAAINet can be applied to
different problem domains. However, in this thesis document GAAINet is specifically applied
to IIoT intrusion detection with the following purpose: (1) generating synthetic attack samples
for IIoT intrusion detection to address class imbalance and (2) performing intrusion detection
by training an immunologically inspired intrusion detector. The typical use of artificial immune
networks (AINs) is to learn the structure of a dataset for clustering or classification purposes
and the research conducted in this thesis document aims to propose a novel standalone
generative AIN model for generating synthetic samples. Furthermore, the generator AIN learns
to generate synthetic samples without exposure to the original dataset. This is achieved through
adversarial training inspired by generative adversarial networks (GANs). GAAINet is not
trained through backpropagation, nor does it use artificial neural networks (ANNs). However,
the training is achieved through immunologically inspired mechanisms such as clonal
expansion, hypermutation and immune network dynamics (i.e. B Cell stimulation and
suppression by neighbours). The main benefits provided by GAAINet is the ability to
synthesize attack samples using an immunologically inspired approach that has the advantage
of self-stabilisation, self-adaptation and self-organisation. Moreover, GAAINet also makes use
of immunologically inspired intrusion detector agents capable of self-improvement as well as
knowledge sharing over time.