The growing complexity and diversity of network traffic have made the detection of previously unseen cyberattacks a critical challenge. While supervised learning models perform well on known threats, they often fail to generalise to novel attack types. In earlier work, we introduced the Unknown Network Attack Detector (UNAD), an unsupervised ensemble-based framework trained exclusively on benign traffic to detect anomalies. This paper presents an enhanced version of UNAD, referred to as UNAD+, which incorporates three key improvements. First, a Weighted Majority Voting (WMV) mechanism replaces majority voting to prioritise stronger detectors and eliminate ambiguous predictions. Second, a supervised refinement stage is introduced, where pseudo-labelled anomalies are used to train a secondary classifier that improves detection accuracy and reduces false positives. Third, a post-hoc explainability layer is added, combining LIME and surrogate tree modelling to provide both local and global interpretability of the system’s decisions. Evaluations on CICIDS2017 and NSL-KDD show that UNAD+ substantially improves detection performance compared with the original UNAD baseline, achieving an F1-score of up to 98.25
Deepfake technology poses significant threats to digital security and media integrity, necessitating robust detection methods. This study introduces a novel approach for enhancing deepfake detection by leveraging unique techniques for analyzing audio data from multiple benchmark datasets. By extracting audio features from Mel, Delta Mel, and Delta-Delta Mel spectrograms, applying image processing techniques to these features, and employing Siamese networks, our method demonstrated unprecedented performances with FakeAVCeleb and MLAAD datasets. Through extensive experimentation, our approach achieved near-perfect accuracy in both validation and testing phases among these datasets. This significant improvement in accuracy, particularly through the focused analysis of audio characteristics, offers a promising direction for developing more robust and resilient deepfake detection systems. These findings highlight the potential of our technique as a promising solution for combating the malicious use of deepfake technology and provides a strong foundation for future research in multimedia forensics and cybersecurity.
Low- or unskilled labour is prevalent in the logistics sector, which simultaneously faces an acute shortage of skilled workers. As part of the RessourcE joint research project, we explore how to identify and promote the potential of employees in such roles. We present a concept for an AI-based tool that analyses competency profiles to predict employee development potential. This tool aims to support companies in recognizing individuals suited for broader responsibilities or career advancement. We also address implementation challenges and propose evidence-based solutions derived from a needs and requirements analysis. Our findings are grounded in early-stage experiments using synthetic human resource (HR) datasets and large language models.
Synthetic aperture radar (SAR) change detection identifies differences between co-registered images acquired at different times. While SAR is highly suitable for disaster monitoring due to its all-weather and day-night capability, severe speckle complicates the detection task. We present a hybrid pipeline comprising edge-preserving denoising, fusion of three standard difference maps into a unified indicator, two-cluster Fuzzy C-Means to obtain high-confidence change and unchanged seeds, and a compact three-level Mini-U-Net trained with weighted Binary Cross-Entropy and Dice loss to classify uncertain pixels through sliding-window inference. Across three SAR datasets, the proposed method surpasses recent baselines in overall accuracy, F1 score and intersection over union, achieving effective speckle suppression and accurate change delineation.
Autonomous agents offer a promising approach to managing complex, resource-intensive operations such as vehicle routing, scheduling, and offshore decommissioning, where fleets must coordinate interdependent tasks across geographical locations under environmental and logistical constraints. This paper presents a simulation–optimization framework in which autonomous agents make local task selections using meta-heuristic-weighted decision criteria. These criteria, along with schedules and task queue sizes, are co-evolved by a Genetic Algorithm to minimize KPIs. We applied our approach to a North Sea decommissioning case study involving pipeline and structure removal, achieving up to 36
Deep learning-based robotic grasping systems rely on vast quantities of labelled data for training, driving the need for synthetically generated data for training models. However, due to their black-box nature, understanding why grasp predictions succeed or fail is challenging. This work investigates the decision-making process of a state-of-the-art (SOTA) grasping model, GraspNet, on point cloud input data using explainable artificial intelligence (XAI) methods. Three methods were implemented: Feature-Based Importance (FBI), Local Interpretable Model-agnostic Explanations (LIME) and a novel correlation-based importance approach. Each method was adapted for the GraspNet model and tested on a point cloud image of household objects placed on a surface in a cluttered manner. Results show that FBI highlights discontinuities in the point cloud, while LIME performs inconsistently due to variations in the model’s point sampling. The correlation-based method provides the most intuitive insights, but its accuracy was limited due to the non-linearities present in the model’s feature-extraction from the point cloud. These methods provide a framework for future analysis of models trained on synthetic data, aiming to help quantify the sim-to-real gap and support the development of accurate and interpretable robotic grasping systems for dynamic industrial environments.
The paper focuses on the possibility of using transformer-based and generative models for mining imbalanced datasets. We briefly review recent developments in constructing transformers and generative models and their application in machine learning. Next, we explain how to use public domain libraries to build a local transformer-based classifier, that can be used solo or in conjunction with a combined transformer-based and VAE model for generating synthetic minority data or generative model for synthetic minority data generation. The approach is validated in the computational experiment showing that transformer-based classifiers perform well and can be an interesting alternative to classic methods of mining imbalanced datasets.
Virtual reality (VR) enhanced by Generative AI (GenAI) avatars offers opportunities for adaptive and engaging educational experiences. While prior studies emphasize the roles of performance expectancy (PE) and technological efficacy (TE) in learner engagement, few examined how AI scaffolding influences metacognition (META), agency, and motivational beliefs in VR. This study applied structural equation modeling (SEM) to investigate these dynamics in two experimental groups in low-immersive VR (desktops). The model demonstrated acceptable fit (CFI = 0.912, TLI = 0.887) and R ^2 values were highest for intention and technological efficacy (both > 0.70), with metacognition at 0.491. PE predicted both intention to continue using the system and TE, while GPT interaction increased META. However, META negatively predicted PE, suggesting a trade-off between reflection and system use. These results highlight the importance of designing AI scaffolding that supports metacognitive processes without undermining motivation. Integrating AI within VR environments can better align with learner expectations and improve both engagement and learning outcomes.
The N-Person Iterated Prisoners’ Dilemma (N-IPD) is an excellent environment to explore collaboration. This paper shows that the voting mechanism is crucial in determining whether sets of agents collaborate, or defect. When each agent can vote against each other agent individually, the agents become cooperative much more easily, ascending the Collaborative Hill. When the agents have only one vote each round, they tend to defect, descending into the Tragic Valley. This is shown with static decision policies, and with policies that learn using reinforcement learning. Fortunately, when agents retain enough history, they can become collaborative even with one vote each round. This voting policy difference is due to the shape of the reward space.
This paper presents a transformer-based framework for Natural Language Processing (NLP) analysis of university reading lists, addressing representational diversity in higher education curricula. We apply sentence embeddings, clustering, topic modelling, and Shannon entropy to examine thematic patterns within academic reading lists, a novel application domain for modern NLP methods. Our approach identifies latent topic structures and quantifies thematic diversity across institutional contexts. Testing on 40 semantically aligned readings from Islamic Studies programmes at UK and Middle Eastern universities, the framework reveals significant differences in curricular emphasis despite surface-level similarity. The methodology demonstrates how transformer-based analysis can detect subtle epistemological patterns that traditional curriculum auditing methods cannot capture. This NLP approach offers scalable tools for curriculum evaluation, supporting evidence-based discussions around academic inclusivity and knowledge representation in higher education.
Recommender systems are widely adopted in digital retail platforms, and stakeholders increasingly demand transparency in how and when they perform reliably. We introduce PS4XRS (Partial Solutions for Explainable Recommender Systems), as a novel XAI tool. Methodology: Using a dataset of user interactions and model performance, we can generate explanations in the form “When the user interacts with at least 3 of these groups of items, we expect the model performance to be ...”. These explanations are obtained via a multi-objective evolutionary algorithm, with objectives based on interpretability, performance and the knowledge from latent item representations derived from the deep recommender system. We performed experiments to determine the most effective parameters for the evolutionary process, and evaluate the trade-offs between explanation complexity and stakeholder usability. Source code for our work can be found at https://github.com/Giancarlo- Catalano/PSSearch .
In GNSS-denied environments such as indoor facilities, urban canyons, or forested terrains, Unmanned Aerial Vehicles (UAVs) must autonomously navigate and track human targets without relying on global positioning. This paper presents a decentralized reinforcement learning (DRL) framework for real-time UAV person tracking and obstacle avoidance using multi-modal sensing. The proposed system integrates RGB, thermal, and depth data through a YOLOv8-nano perception module, enabling robust object detection under low visibility and occlusion. A rule-based geometric model estimates the target’s relative position, which, along with sensor fusion inputs, forms the state space for a Proximal Policy Optimization (PPO) control policy. The agent operates over a discrete action set, allowing stable and sample-efficient learning while maintaining reactive and goal-directed behavior. The framework is validated in simulated scenarios demonstrating superior performance in navigation accuracy, tracking robustness, and energy efficiency. Our approach highlights the potential of decentralized DRL for scalable and resilient UAV autonomy in challenging operational conditions.
In recent years, large language model technology has been widely used, but its reasoning ability still has significant limitations. Especially when dealing with more complex logical reasoning problems, the accuracy of large language model reasoning is often unable to meet the requirements, and the resolution method can better ensure the accuracy of reasoning. In this paper, we propose Bi-Resolution, a novel method for reasoning about large language models. We introduce bidirectional reasoning into the improved resolution method and implement an automated reasoning process based on the generation of large language models through the design of prompt words. Technically, Bi-Resolution first converts the natural language problem into a symbolic representation of first-order logic, and selects the corresponding version of resolution algorithm according to the predicted reasoning result. This method can help the large language model to more accurately judge the reasoning problem with the conclusion of "not entirely true and not entirely false". In the process of resolution, the idea of bidirectional reasoning is used to instantiate the constraint variables, which removes the redundant conditions in the reasoning problem and reduces the complexity of reasoning. We conducted experiments on the FOLIO dataset, and the results show that the Bi-Resolution architecture successfully improves the accuracy of large language model reasoning.
The reliability of modern applications increasingly depends on the stability of underlying distributed web services. Monitoring these services through time series analysis is essential for detecting performance degradations and security anomalies. However, many existing anomaly detection models struggle to adapt to the dynamic nature of cloud-based environments, particularly when dealing with evolving service patterns or system reconfigurations. To help address these challenges, this paper proposes DeLSTM-AE, a novel framework that integrates time series decomposition with a Long Short-Term Memory (LSTM) and an Autoencoder (AE) to improve the detection of anomalies in univariate time series data. By separating seasonal, trend, and residual components prior to training, our model enables more targeted learning of distinct temporal structures, thereby improving its ability to capture subtle anomalies. We evaluate the proposed framework on the benchmark AIOps Challenge 2018 dataset, which includes real-time Key Performance Indicator (KPI) logs from major cloud service providers such as eBay, Sogou, and Tencent. DeLSTM-AE achieves superior performance compared to existing state-of-the-art methods, with a Precision of 99
This paper introduces a hybrid artificial intelligence (AI) system that combines deep learning (DL) and genetic programming (GP) to solve visual arithmetic problems. DL is used for digit recognition in images, while GP performs symbolic reasoning to compute arithmetic results. Four image-based datasets were created, each representing different levels of difficulty in addition and subtraction tasks, with and without carrying or borrowing. By extracting and encoding structured representations from visual data, the new hybrid approach can explicitly capture the symbolic relationships, enabling more efficient reasoning. The findings suggest that the hybrid system matches the accuracy of DL benchmarks, while offering enhanced interpretability and data-efficient generalisation. The proposed approach demonstrates the potential of combining sub-symbolic perception with symbolic reasoning to unify perceptual understanding and logical inference as a foundational step towards human-like artificial intelligence.
This paper introduces the first application of data science to the UK Honours system. We present a comprehensive Natural Language Processing methodology for evaluating public sentiment of Honours recipients. In order to form an opinion about applicants for the UK King’s Honours, we have evaluated two existing sentiment algorithms (Afinn, Vader) and then created our own novel algorithm (Minos). The promising results in this work indicate that this system can be used to augment human evaluation to better judge whether a current or future recipient has maintained the high standards of conduct demanded to retain an Honour. Our novel approach is generalisable to any individual with a sufficient internet footprint and has applications in many fields including recruitment, national security and investigative journalism.
Integrating domain-specific metadata and clinical variables into deep learning frameworks remains essential yet challenging for accurate biomedical predictions. Conventional early or late fusion techniques often fail to capture complex interactions between different patient modalities, limiting predictive power and clinical utility. This study presents a novel Adaptive Metadata Encoder (AME) that dynamically embeds structured covariates, such as cholesterol levels, age, race, and other clinical features, directly within deep electrocardiogram (ECG) models. This AME enables cardiovascular risk assessment, specifically targeting the detection of reduced left ventricular ejection fraction (LVEF <40
This study presents a data-driven approach to forecasting total precipitation in London using an Artificial Neural Network (ANN) within a spatio-temporal framework. Leveraging ERA5 data from 2010 to 2025, the methodology includes automated NetCDF extraction, feature engineering with lagged precipitation and cyclic time encodings, and dimensionality reduction via a trained Autoencoder. The ANN, designed in a GenCast-style architecture, was trained using the Adam optimiser over 50 epochs and achieved strong performance. SHAP analysis highlighted the importance of lag features and seasonal time variables, enhancing interpretability and supporting the model’s application in urban flood risk management and climate resilience.
Energy profiling plays a crucial role in optimising smart building operations, especially with the increasing popularity of personalised, user-centric AI applications. Current research lacks emphasis on interpretability, transparency, and accessibility for non-expert stakeholders, where decision-making either relies solely on machine learning insights or unstructured knowledge bases. Hence, this study aims to enhance the interpretability of energy profiling and generate tailored recommendations based on correlated data sources from various aspects. This approach combines data-driven and knowledge-driven techniques by integrating energy clustering insights and unstructured knowledge bases to provide tailored energy recommendations. By combining Large Language Models (LLMs) and Explainable AI (XAI), this approach leads to: (1) identifying new consumer personas based on contextualised cluster insights, (2) finding the most impactful features reflecting energy insights, and (3) turning those insights into clear, human-readable reports and recommendations. This transforms smart meters from passive data collectors into intelligent advisory tools for consumers, policymakers, and energy providers.
For underwater object detection, sonar imagery generated from Forward Looking Sonar (FLS) and Side Scan Sonar (SSS) systems is usually inspected by experts while AI algorithms are currently under investigation. The transparency and reasoning of AI-based object detection models are crucial for users of these detectors. However, the black-box nature of AI-based detectors limits their interpretability. For a few years, transformer networks have been employed more extensively for object detection and are also used together with sonar data. This paper investigates transformer models for object detection on sonar data and evaluates explanations for the detected objects. We test multiple transformer object detectors, which are evaluated on two datasets. One dataset is a FLS dataset and one is a SSS dataset. After training the transformer models, attribution maps are generated and used for explainability. Here, two XAI methods are used for post-hoc explanations for object detection. The post-hoc explanations are evaluated based on eXplainable Artificial Intelligence (XAI) metrics considering faithfulness and localisation. It turns out that transformer networks with standard backbones provide reasonable accuracies, together with useful attribution maps if used for sonar data. However, multiple XAI methods have to be used to get good explanations.
