Context: Placing Artificial Intelligence (AI) in control of cybersecurity compliance and auditing shifts its role from decision-support to direct execution of regulatory operational processes, where AI outputs may constitute compliance artefacts and audit evidence. This raises the problem of Meta-Compliance , in which not only the organisation but also the AI system must satisfy enforceable requirements. Yet existing frameworks provide no operational criteria for recognising AI as authoritative in such roles. Trustworthy AI principles define high-level Second-Layer requirements but remain non-binding, whereas First-Layer organisational requirements impose explicit justificatory and evidentiary duties. Objectives: This study investigates the minimal normative conditions under which AI systems can be recognised as authoritative in compliance and auditing, capable of producing evidence valid for assurance. Methods: Doctrinal analysis is conducted on binding “shall/must” provisions across PCI DSS, DORA, UK GDPR, NIS2, ISO/IEC 27001, and NIST SP 800-53. Provisions are normalised through the compliance–audit chain ( requirement → control → rule → evidence ) and mapped against Second-Layer AI governance requirements. The result is the Compliance–Audit Authority Benchmark (CAAB) , comprising six criteria: Traceability, Explainability, Evidence Integrity, Adaptability, Action Governance, and Reasoning. Results: Applying CAAB across AI model families and architectures shows that symbolic and knowledge-representation methods satisfy most criteria intrinsically, whilst neural, deep, and generative models do not unless supported by external governance mechanisms. This exposes a structural gap between First-Layer organisational requirements and Second-Layer AI requirements, clarifying that authority rests on evidentiary guarantees rather than statistical accuracy. Conclusion: The study formalises Meta-Compliance as the recursive structure in which both organisations and AI systems become subjects of assurance. CAAB defines the minimum conditions for recognising AI as authoritative, whilst the proposed Verifiable Reasoning Architecture (VRA) may offer a pathway toward AI systems anchored in secured evidence, reproducible inference, and symbolic governance, establishing audit-ready authority in high-risk contexts.

The ever-escalating prevalence of malware is a serious cybersecurity threat, often requiring advanced post-incident forensic investigation techniques. This paper proposes a framework to enhance malware forensics by leveraging reinforcement learning (RL). The approach combines heuristic and signature-based methods, supported by RL through a unified MDP model, which breaks down malware analysis into distinct states and actions. This optimisation enhances the identification and classification of malware variants. The framework employs Q-learning and other techniques to boost the speed and accuracy of detecting new and unknown malware, outperforming traditional methods. We tested the experimental framework across multiple virtual environments infected with various malware types. The RL agent collected forensic evidence and improved its performance through Q-tables and temporal difference learning. The epsilon-greedy exploration strategy, in conjunction with Q-learning updates, effectively facilitated transitions. The learning rate depended on the complexity of the MDP environment: higher in simpler ones for quicker convergence and lower in more complex ones for stability. This RL-enhanced model significantly reduced the time required for post-incident malware investigations, achieving a high accuracy rate of 94% in identifying malware. These results indicate RL’s potential to revolutionise post-incident forensics investigations in cybersecurity. Future work will incorporate more advanced RL algorithms and large language models (LLMs) to further enhance the effectiveness of malware forensic analysis.

In cyber security, it is critical that event data is collected in as near real time as possible to enable early detection and response to threats. Performing analytics from event logs stored in databases slows down the response time due to the time cost of database insertion and retrieval operations. The authors present a data collection framework that minimizes the need for long-term storage. Events are buffered in memory, up to a configurable threshold, before being streamed in real time using live streaming technologies. The framework deploys virtualized data collecting agents that ingest data from multiple sources including threat intelligence. The framework enables the correlation of events from various sources, improving detection precision. The authors have tested the framework in a real time, machine-learning-based threat detection system. The results show a time gain of 300 milliseconds in transmission time from event capture to analytics system, compared with storage-based collection frameworks. Threat detection was measured at 95%, which is comparable to the benchmark snort IDS.

The research develops a practical Machine Learning framework with a comparative and comprehensive approach to sequence-learn and then detect the online banking payment fraud. The integrated framework introduces exploratory analysis and feature engineering, multiple modelling and performance comparison, and model robustness, uncertainty and sensitivity analysis toward a systematic approach for Machine Learning applications. For demonstration purpose, the framework is implemented on a set of real-life online banking transaction datasets obtained from a UK-based bank through three models, i.e., Support Vector Machine, Markov Model and LSTM model, with various combinational features of the datasets evidenced in the exploratory analysis and modelling with noise ratios of datasets, range values of model parameters and confidence intervals of prediction results. The modelling results show that overall, the LSTM model achieves the best performance, with outcome accuracy of 97.7%, indicating its advantage in modelling sequential data such as customer behaviours.

This paper details the research into using audio signal processing methods to provide authentication and identification services for the purpose of enhancing cyber security in voice applications. Audio is a growing domain for cyber security technology. It is envisaged that over the next decade, the primary interface for issuing commands to consumer internet-enabled devices will be voice. Increasingly, devices such as desktop computers, smart speakers, cars, TV's, phones an Internet of Things (IOT) devices all have built in voice assistants and voice activated features. This research outlines an approach to securely identify and authenticate users of audio and voice operated systems that utilises existing cryptography methods and audio steganography in a method comparable to a PKI for sound, whilst retaining the usability associated with audio and voice driven systems.

Neuro Linguistic Programming (NLP) is one of the most utilised approaches for personality development and Meta model is one of the most important techniques in this process. Usually, when one speaks about a problem or a situation, the words that one chooses will delete, distort or generalize portions of their experience. Meta model, which is a set of specific questions or language patterns, can be used to understand and recover the information hidden behind the words used. This technique can be adopted to understand other people's problems or enable them to understand their own issues better. Applying the Meta Model, however, requires a great level of skill and experience for correct identification of deletion, distortion and generalization. Using the appropriate recovery questions is challenging for NLP practitioners and Psychologists. Moreover, the efficiency and accuracy of existing methods on the Meta model can potentially be hindered by human errors such as personal judgment or lack of experience and skill. This research aims to automate the process of using the Meta Model in conversation in order to eliminate human errors, thereby increasing the efficiency and accuracy of this method. An intelligent software has been developed using Natural Language Processing, with the ability to apply the Meta model techniques during conversation with its user. Comparisons of this software with performance of an established NLP practitioner have shown increased accuracy in identification of the deletion and generalization processes. Recovery of information has also been more efficient in the software in comparison to an NLP practitioner.

The Sea-Viewing Wide Field-of-View Sensor (SeaWiFS) data from 1997 to 2001 are adopted to monitor the radiometric stability of six pseudo-invariant calibration sites (PICSs) at the top of atmosphere (TOA). Cloud-free and homogeneous observations of the spectral TOA reflectance ρTOAat eight SeaWiFS channels over these sites are fitted to the Ross-Li bidirectional reflectance distribution function (BRDF) model, and the time series of BRDF-normalized spectral TOA reflectance RTOAis presented and analyzed afterward. Overall, good stability during the evaluated period is exhibited as more than half of the derived trends are statistically insignificant, whereas root mean square (RMS) of the BRDF modeling residuals reveal spectral dependence of the PICSs' stability at TOA, i.e., the uncertainty of RTOAappears to be larger at shortwave visible (SV) channels (~2.5%) compared with that of red/NIR bands (~1%). In addition, the early mission data adopted in our study shows favorable reliability thus is recommended to be applied for similar purposes.

Improvements of robustness of speech recognition is one of the hottest topics in speech signal processing, particularly when applied within a noisy environment. Most of the research efforts focused in combining audio and visual data to implement an audiovisual speech recognition (AVSR) system. Bimodal approach demonstrated that a superior performance can be gained compared to the separate audio or visual approach. This paper proposes a fuzzy logic-based data fusion method that combines the recognition capabilities of two independent working systems namely the automatic speech recognition system (ASR) and the automatic visual recognition system (AVR). The main purpose is to boost the whole system's performance keeping the ASR separate from the AVR. This approach provides a powerful method that enables simpler data fusion at decision level rather than the more complex at data and features level. Such complexity is also lowered due to the fuzzy logic-based implementation of the data fusion engine. Preliminary experimental results confirms the proposed approach.

Text entry from people is not only grammatical and distinct, but also noisy. For example, a user's typing stream contains all the information about the user's interaction with computer using a QWERTY keyboard, which may include the user's typing mistakes as well as specific vocabulary, typing habit, and typing performance. In particular, these features are obvious in disabled users' typing streams. This paper proposes a new concept called noisy language modeling by further developing information theory and applies neural networks to one of its specific application-typing stream. This paper experimentally uses a neural network approach to analyze the disabled users' typing streams both in general and specific ways to identify their typing behaviors and subsequently, to make typing predictions and typing corrections. In this paper, a focused time-delay neural network (FTDNN) language model, a time gap model, a prediction model based on time gap, and a probabilistic neural network model (PNN) are developed. A 38% first hitting rate (HR) and a 53% first three HR in symbol prediction are obtained based on the analysis of a user's typing history through the FTDNN language modeling, while the modeling results using the time gap prediction model and the PNN model demonstrate that the correction rates lie predominantly in between 65% and 90% with the current testing samples, and 70% of all test scores above basic correction rates, respectively. The modeling process demonstrates that a neural network is a suitable and robust language modeling tool to analyze the noisy language stream. The research also paves the way for practical application development in areas such as informational analysis, text prediction, and error correction by providing a theoretical basis of neural network approaches for noisy language modeling.