Large language models (LLMs), such as ChatGPT, have rapidly penetrated into people's work and daily lives over the past few years, due to their extraordinary conversational skills and intelligence. ChatGPT has become the fastest-growing software in terms of user numbers in human history and become an important foundational model for the next generation of artificial intelligence applications. However, the generations of LLMs are not entirely reliable, often producing content with factual errors, biases, and toxicity. Given their vast number of users and wide range of application scenarios, these unreliable responses can lead to many serious negative impacts. This thesis introduces the exploratory works in the field of language model reliability during the PhD study, focusing on the correctness, non-toxicity, and fairness of LLMs from both software testing and natural language processing perspectives. First, to measure the correctness of LLMs, we introduce two testing frameworks, FactChecker and LogicAsker, to evaluate factual knowledge and logical reasoning accuracy, respectively. Second, for the non-toxicity of LLMs, we introduce two works for red-teaming LLMs. Third, to evaluate the fairness of LLMs, we introduce two evaluation frameworks, BiasAsker and XCulturalBench, to measure the social bias and cultural bias of LLMs, respectively.

Diffusion models have emerged from various theoretical and methodological perspectives, each offering unique insights into their underlying principles. In this work, we provide an overview of the most prominent approaches, drawing attention to their striking analogies - namely, how seemingly diverse methodologies converge to a similar mathematical formulation of the core problem. While our ultimate goal is to understand these models in the context of graphs, we begin by conducting experiments in a simpler setting to build foundational insights. Through an empirical investigation of different diffusion and sampling techniques, we explore three critical questions: (1) What role does noise play in these models? Our findings aim to enhance the understanding of diffusion models and in the long run their application in graph machine learning. The forward process is modelled by a Markov process. The reverse process is unknown and needs to be approximated; this is usually done with a neural network. Consider the analogy of dropping a small amount of paint into a glass of water. Initially, the paint is concentrated in one location, but over time, it diffuses throughout the water until it reaches a state of equilibrium.

Large Language Models (LLMs) have shown remarkable progress, but their real-world application necessitates reliable calibration. This study conducts a comprehensive analysis of calibration degradation of LLMs across four dimensions: models, calibration metrics, tasks, and confidence extraction methods. Initial analysis showed that the relationship between alignment and calibration is not always a trade-off, but under stricter analysis conditions, we found the alignment process consistently harms calibration. This highlights the need for (1) a careful approach when measuring model confidences and calibration errors and (2) future research into algorithms that can help LLMs to achieve both instruction-following and calibration without sacrificing either.

We introduce a novel neural network architecture, termed Taylor-neural Lyapunov functions, designed to approximate Lyapunov functions with formal certification. This architecture innovatively encodes local approximations and extends them globally by leveraging neural networks to approximate the residuals. Our method recasts the problem of estimating the largest region of attraction - specifically for maximal Lyapunov functions - into a learning problem, ensuring convergence around the origin through robust control theory. Physics-informed machine learning techniques further refine the estimation of the largest region of attraction. Remarkably, this method is versatile, operating effectively even without simulated data points. We validate the efficacy of our approach by providing numerical certificates of convergence across multiple examples. Our proposed methodology not only competes closely with state-of-the-art approaches, such as sum-of-squares and LyZNet, but also achieves comparable results even in the absence of simulated data. This work represents a significant advancement in control theory, with broad potential applications in the design of stable control systems and beyond.

Knowledge Tracing (KT) is a critical component in online learning, but traditional approaches face limitations in interpretability and cross-domain adaptability. This paper introduces Language Model-based Code Knowledge Tracing (CodeLKT), an innovative application of Language model-based Knowledge Tracing (LKT) to programming education. CodeLKT leverages pre-trained language models to process learning data, demonstrating superior performance over existing KT and Code KT models. We explore Domain Adaptive Pre-Training (DAPT) and Task Adaptive Pre-Training (TAPT), showing enhanced performance in the coding domain and investigating cross-domain transfer between mathematics and coding. Additionally, we present an theoretically-informed integrated system combining CodeLKT with large language models to generate personalized, in-depth feedback to support students' programming learning. This work advances the field of Code Knowledge Tracing by expanding the knowledge base with language model-based approach and offering practical implications for programming education through data-informed feedback.

Graphs are ubiquitous for modeling complex relationships between objects across various fields. Graph neural networks (GNNs) have become a mainstream technique for graph-based applications, but their performance heavily relies on abundant labeled data. To reduce labeling requirement, pre-training and prompt learning has become a popular alternative. However, most existing prompt methods do not differentiate homophilic and heterophilic characteristics of real-world graphs. In particular, many real-world graphs are non-homophilic, not strictly or uniformly homophilic with mixing homophilic and heterophilic patterns, exhibiting varying non-homophilic characteristics across graphs and nodes. In this paper, we propose ProNoG, a novel pre-training and prompt learning framework for such non-homophilic graphs. First, we analyze existing graph pre-training methods, providing theoretical insights into the choice of pre-training tasks. Second, recognizing that each node exhibits unique non-homophilic characteristics, we propose a conditional network to characterize the node-specific patterns in downstream tasks. Finally, we thoroughly evaluate and analyze ProNoG through extensive experiments on ten public datasets.

Deep supervised learning models require high volume of labeled data to attain sufficiently good results. Although, the practice of gathering and annotating such big data is costly and laborious. Recently, the application of self supervised learning (SSL) in vision tasks has gained significant attention. The intuition behind SSL is to exploit the synchronous relationships within the data as a form of self-supervision, which can be versatile. In the current big data era, most of the data is unlabeled, and the success of SSL thus relies in finding ways to improve this vast amount of unlabeled data available. Thus its better for deep learning algorithms to reduce reliance on human supervision and instead focus on self-supervision based on the inherent relationships within the data. With the advent of ViTs, which have achieved remarkable results in computer vision, it is crucial to explore and understand the various SSL mechanisms employed for training these models specifically in scenarios where there is less label data available. In this survey we thus develop a comprehensive taxonomy of systematically classifying the SSL techniques based upon their representations and pre-training tasks being applied. Additionally, we discuss the motivations behind SSL, review popular pre-training tasks, and highlight the challenges and advancements in this field. Furthermore, we present a comparative analysis of different SSL methods, evaluate their strengths and limitations, and identify potential avenues for future research.

Artificial intelligence models encounter significant challenges due to their black-box nature, particularly in safety-critical domains such as healthcare, finance, and autonomous vehicles. Explainable Artificial Intelligence (XAI) addresses these challenges by providing explanations for how these models make decisions and predictions, ensuring transparency, accountability, and fairness. Existing studies have examined the fundamental concepts of XAI, its general principles, and the scope of XAI techniques. However, there remains a gap in the literature as there are no comprehensive reviews that delve into the detailed mathematical representations, design methodologies of XAI models, and other associated aspects. This paper provides a comprehensive literature review encompassing common terminologies and definitions, the need for XAI, beneficiaries of XAI, a taxonomy of XAI methods, and the application of XAI methods in different application areas. The survey is aimed at XAI researchers, XAI practitioners, AI model developers, and XAI beneficiaries who are interested in enhancing the trustworthiness, transparency, accountability, and fairness of their AI models.

In the realm of cyber-security, detecting Advanced Persistent Threats (APTs) remains a formidable challenge due to their stealthy and sophisticated nature. This research paper presents an innovative approach that leverages Convolutional Neural Networks (CNNs) with a 2D baseline model, enhanced by the cutting-edge Cat Swarm Optimization (CSO) algorithm, to significantly improve APT detection accuracy. By seamlessly integrating the 2D-CNN baseline model with CSO, we unlock the potential for unprecedented accuracy and efficiency in APT detection. The results unveil an impressive accuracy score of $98.4\%$, marking a significant enhancement in APT detection across various attack stages, illuminating a path forward in combating these relentless and sophisticated threats.

Over the recent years, the protection of the so-called `soft-targets', i.e. locations easily accessible by the general public with relatively low, though, security measures, has emerged as a rather challenging and increasingly important issue. The complexity and seriousness of this security threat growths nowadays exponentially, due to the emergence of new advanced technologies (e.g. Artificial Intelligence (AI), Autonomous Vehicles (AVs), 3D printing, etc.); especially when it comes to large-scale, popular and diverse public spaces. In this paper, a novel Digital Twin-as-a-Security-Service (DTaaSS) architecture is introduced for holistically and significantly enhancing the protection of public spaces (e.g. metro stations, leisure sites, urban squares, etc.). The proposed framework combines a Digital Twin (DT) conceptualization with additional cutting-edge technologies, including Internet of Things (IoT), cloud computing, Big Data analytics and AI. In particular, DTaaSS comprises a holistic, real-time, large-scale, comprehensive and data-driven security solution for the efficient/robust protection of public spaces, supporting: a) data collection and analytics, b) area monitoring/control and proactive threat detection, c) incident/attack prediction, and d) quantitative and data-driven vulnerability assessment. Overall, the designed architecture exhibits increased potential in handling complex, hybrid and combined threats over large, critical and popular soft-targets. The applicability and robustness of DTaaSS is discussed in detail against representative and diverse real-world application scenarios, including complex attacks to: a) a metro station, b) a leisure site, and c) a cathedral square.