Traditional networks are designed to be hard on the outside and soft on the inside. It is this soft inside which has made the traditional perimeter model laughable to attackers, who can easily breach a network and run away with the data without even having to deal with the hardened perimeter. The zero-trust security model, created by John Kindervag in 2010, addresses the security flaws of the traditional perimeter model and asserts that all network traffic on the inside should not be trusted by default. Other core principles of zero trust include verification and continuous monitoring of all communication, as well as encryption of all data in transit and data at rest, since the goal of zero trust is to focus on protecting data. Although the zero-trust model was created in 2010, with some of the associated security practices existing even before that, many healthcare organisations are still choosing to focus primarily on securing the perimeter instead of focusing on the vulnerabilities within them. The current COVID-19 pandemic which healthcare providers are struggling with further highlights the need for improvements to security within the network perimeter, as many healthcare providers and vaccine developers are still using vulnerable, outdated legacy systems which could become compromised and indirectly have a detrimental effect on patient care. Legacy systems which are technologically limited, as well as medical devices which cannot be controlled or managed by network administrators, create boundaries to transitioning to a zero-trust architecture. It is challenges like this that have been explored during the research phase of this project in order to gain a better understanding of how a health organisation can adopt zero-trust practices despite the limitations of their current architecture. From the information gathered during this research, a framework was developed to allow a health organisation to transition to a more secure architecture based on the concept of zero-trust. Aspects of the proposed framework were tested in Cisco Modelling Labs (CML), and the results were evaluated to ensure the validity of some of the recommendations laid out in the framework. The main objective of this research was to prove that if a host within the local area network (LAN) were to be compromised, the damage would be limited to that host and would not spread throughout the rest of the network. This was successful after the qualitative research performed in CML. One of the other takeaways from testing the framework in CML was that medical devices could be secured by placing firewalls directly in front of them. This placement of firewalls may seem like an unorthodox approach and was shown to increase latency, but the blocking of all unnecessary traffic on the rest of the network will result in a performance boost and should balance it out in a real-world application.

Using technology to prevent cyber-attacks has allowed organisations to somewhat automate cyber security. Despite solutions to aid organisations, many are susceptible to phishing and spam emails which can make an unwanted impact if not mitigated. Traits that make organisations susceptible to phishing and spam emails include a lack of awareness around the identification of malicious emails, explicit trust, and the lack of basic security controls. For any organisation, phishing and spam emails can be received and the consequences of an attack could result in disruption. This research investigated the threat of phishing and spam and developed a detection solution to address this challenge. Deep learning and natural language processing are two techniques that have been employed in related research, which has illustrated improvements in the detection of phishing. Therefore, this research contributes by developing Phish Responder, a solution that uses a hybrid machine learning approach combining natural language processing to detect phishing and spam emails. To ensure its efficiency, Phish Responder was subjected to an experiment in which it has achieved an average accuracy of 99% with the LSTM model for text-based datasets. Furthermore, Phish Responder has presented an average accuracy of 94% with the MLP model for numerical-based datasets. Phish Responder was evaluated by comparing it with other solutions and through an independent t-test which demonstrated that the numerical-based technique is statistically significantly better than existing approaches.

Annotation tools are an essential component in the creation of datasets for machine learning purposes. Annotation tools have evolved greatly since the turn of the century, and now commonly include collaborative features to divide labor efficiently, as well as automation employed to amplify human efforts. Recent developments in machine learning models, such as Transformers, allow for training upon very large and sophisticated multimodal datasets and enable generalization across domains of knowledge. These models also herald an increasing emphasis on prompt engineering to provide qualitative fine-tuning upon the model itself, adding a novel emerging layer of direct machine learning annotation. These capabilities enable machine intelligence to recognize, predict, and emulate human behavior with much greater accuracy and nuance, a noted shortfall of which have contributed to algorithmic injustice in previous techniques. However, the scale and complexity of training data required for multimodal models presents engineering challenges. Best practices for conducting annotation for large multimodal models in the most safe and ethical, yet efficient, manner have not been established. This paper presents a systematic literature review of crowd and machine learning augmented behavioral annotation methods to distill practices that may have value in multimodal implementations, cross-correlated across disciplines. Research questions were defined to provide an overview of the evolution of augmented behavioral annotation tools in the past, in relation to the present state of the art. (Contains five figures and four tables).

Smart city infrastructure and the related theme of critical national infrastructure have attracted growing interest in recent years in academic literature, notably how cyber-security can be effectively applied within the environment, which involves using cyber-physical systems. These operate cross-domain and have massively improved functionality and complexity, especially in threat modelling cyber-security analysis—the disparity between current cyber-security proficiency and the requirements for an effective cyber-security systems implementation. Analysing risk across the entire analysed system can be associated with many different cyber security methods for overall cyber risk analysis or identifying vulnerability for individually modelled objects. One method for performing risk analysis proposed in the literature is by applying Bayesian-based threat modelling methodologies. This paper performs a systematic literature review of Bayesian networks and unique alternative methodologies for smart city infrastructure analysis and related critical national infrastructures. A comparative analysis of the different methodological approaches, considering the many intricacies, metrics, and methods behind them, with suggestions made for future research in the field of cyber-physical threat modelling for smart city infrastructure.

Lung cancer (LC) is the most common cause of cancer-related deaths in the UK due to delayed diagnosis. The existing literature establishes a variety of factors which contribute to this, including the misjudgement of anatomical structure by doctors and radiologists. This study set out to develop a solution which utilises multiple modalities in order to detect the presence of LC. A review of the existing literature established failings within methods to exploit rich intermediate feature representations, such that it can capture complex multimodal associations between heterogenous data sources. The methodological approach involved the development of a novel machine learning (ML) model to facilitate quantitative analysis. The proposed solution, named EMM-LC Fusion, extracts intermediate features from a pre-trained modified AlignedXception model and concatenates these with linearly inflated features of Clinical Data Elements (CDE). The implementation was evaluated and compared against existing literature using F1 score, average precision (AP), and area under curve (AUC) as metrics. The findings presented in this study show a statistically significant improvement (p < 0.05) upon the previous fusion method, with an increase in F-Score from 0.402 to 0.508. The significance of this establishes that the extraction of intermediate features produces a fertile environment for the detection of intermodal relationships for the task of LC classification. This research also provides an architecture to facilitate the future implementation of alternative biomarkers for lung cancer, one of the acknowledged limitations of this study.

This study introduces a novel architecture for value, preference, and boundary alignment in large language models (LLMs) and generative AI systems, accompanied by an experimental implementation. It addresses the limitations in AI model trustworthiness stemming from insufficient comprehension of personal context, preferences, and cultural diversity, which can lead to biases and safety risks. Using an inductive, qualitative research approach, we propose a framework for personalizing AI models to improve model alignment through additional context and boundaries set by users. Our framework incorporates user-friendly tools for identification, annotation, and simulation across diverse contexts, utilizing prompt-driven semantic segmentation and automatic labeling. It aims to streamline scenario generation and personalization processes while providing accessible annotation tools. The study examines various components of this framework, including user interfaces, underlying tools, and system mechanics. We present a pilot study that demonstrates the framework’s ability to reduce the complexity of value elicitation and personalization in LLMs. Our experimental setup involves a prototype implementation of key framework modules, including a value elicitation interface and a fine-tuning mechanism for language models. The primary goal is to create a token-based system that allows users to easily impart their values and preferences to AI systems, enhancing model personalization and alignment. This research contributes to the democratization of AI model fine-tuning and dataset generation, advancing efforts in AI value alignment. By focusing on practical implementation and user interaction, our study bridges the gap between theoretical alignment approaches and real-world applications in AI systems.