This article surveys convolution-based models - convolutional neural networks (CNNs), Conformers, ResNets, and CRNNs-as speech signal processing models and provide their statistical backgrounds and speech recognition, speaker identification, emotion recognition, and speech enhancement applications. Through comparative training cost assessment, model size, accuracy and speed assessment, we compare the strengths and weaknesses of each model, identify potential errors and propose avenues for further research, emphasising the central role it plays in advancing applications of speech technologies.

Open Radio Access Networks (O-RAN) enhance RAN flexibility, interoperability, and intelligence by introducing open interfaces, disaggregation of RAN components, and datadriven control. This flexibility has the potential to significantly reduce capital and operating expenses (CAPEX/OPEX) by enhancing efficiencies in network energy consumption. Advanced Sleep Modes (ASMs) are an advantageous energy-saving technique for 5G base stations (BS), with a set of sleep depths that progressively deactivate different RAN components. The deployment and effectiveness of ASMs in O-RAN raises a number of challenges that have yet to be investigated. This paper surveys and critically reviews RAN ASMs, as well as other energy-enhancing techniques applied in O-RAN. It presents the first granular, percomponent O-RAN Radio Unit (O-RU) power model, that will enable researchers to quantify the power savings gained from deploying ASMs. A calculation of the energy savings achieved at each ASM level per O-RU component for different 3GPP functional splits is provided. The power consumption of O-RU sub-components for functional splits 6, 7.2x, and 8 is shown. The default 3GPP functional split 7.2x as adopted by O-RAN shows that substantial power savings can be achieved by applying ASMs to the O-RU, achieving a minimum of 80% savings depending on the choice of SM as compared to power consumption at full load. Finally, a discussion of the open research challenges and opportunities in deploying ASMs in O-RAN is provided.

Wearable electroencephalography (EEG) devices allow non-invasive brain monitoring during everyday activities and are widely used in managing conditions like epilepsy. However, EEG signals are small, and often corrupted by artifacts during real-world recordings. Many artificial intelligence models for EEG artifact removal have been proposed recently, but their real-time deployment on edge hardware, suitable for embedding in an EEG device itself, has remained unrealized until now. This paper introduces the first implementation of a deep autoencoder network for EEG artifact removal on edge hardware, using three embedded systems: an Arduino Nano 33 BLE, a Coral Dev Board Micro, and a Coral Dev Board Mini. We compare these systems relative to their power consumption and inference time when processing 4 s EEG segments. The Coral Dev Board Mini demonstrated the fastest inference time (8.9 ms) but at the cost of high power consumption (1.7 W). The Coral Dev Board Micro balanced inference time (273 ms) with power consumption (0.6 W), while the Arduino Nano 33 BLE achieved the lowest power draw (0.1 W) but with a longer inference time (1.3 s). Our results highlight that edge AI for EEG artifact removal is possible, and likely not limited by inference time but by power consumption if long-term, battery-powered operation is required. There remains scope, and need, for further power optimization. Overall, this first-of-its-kind edge deployment of EEG processing marks a significant step towards artifact robust real-time, portable EEG monitoring solutions.

Maintaining contextual coherence in extended text generation remains a significant challenge in computational linguistics. The introduction of the Stochastic Hierarchical Embedding (SHE) mechanism offers a novel approach to this issue, incorporating stochastic variability within hierarchical embeddings to enhance adaptability and coherence in language models. This study details the integration of SHE into a contemporary open-source Large Language Model (LLM), providing a comprehensive mathematical framework and outlining the modifications required for effective implementation. Experimental evaluations demonstrate that SHE significantly improves contextual coherence and prompt convergence, as evidenced by higher Contextual Coherence Scores (CCS) and Prompt Convergence Rates (PCR). Additionally, the SHE-augmented LLM exhibits enhanced generalization capabilities across diverse linguistic tasks and increased robustness to input noise. These findings suggest that SHE presents a promising advancement in the development of more coherent and adaptable LLMs.

The escalating sophistication of ransomware attacks necessitates the development of innovative detection methodologies capable of identifying and mitigating such threats with high precision. The introduction of Quantum Heuristic Analysis (QHA) represents a significant advancement in this domain, leveraging quantum computational principles to enhance anomaly detection mechanisms. Through the integration of quantum heuristics, QHA effectively identifies complex data patterns and subtle anomalies, thereby offering a robust defense against evolving ransomware threats. Empirical evaluations demonstrate that QHA achieves a detection accuracy exceeding 98%, surpassing traditional signature-based approaches. Furthermore, the framework's design, which operates without human intervention, reduces the potential for human error and allows for continuous monitoring and detection, thereby enhancing the overall resilience of cybersecurity infrastructures. The scalability of QHA has been validated across varying dataset sizes, maintaining consistent performance metrics, and its adaptability to diverse encryption techniques demonstrates its robustness. These findings suggest that QHA holds substantial promise for integration into existing cybersecurity infrastructures, providing a proactive and efficient approach to ransomware detection.

The Efficient Binary Arithmetic Model (EBAM) introduces a novel approach to digital logic circuit design aimed at minimizing transistor usage in binary arithmetic operations such as addition and subtraction. By employing the Transformer Logic Section (TLS) for one-to-one mapping of arithmetic pairs and separating operations into standardized and carry states, EBAM reduces computational complexity while optimizing circuit space. This unique design, which relies solely on AND gates for operation, demonstrates the ability to handle both neutral and carry states efficiently. Results show that EBAM is versatile, with potential expansion for other arithmetic operations such as multiplication, making it an ideal candidate for light-based or quantum processing environments where efficient circuit design is crucial.

We investigate cooperative impersonation jamming on angle-based physical layer authentication (PLA) within 6G systems using hybrid antenna arrays. PLA leverages angleof-arrival (AoA) information to authenticate user equipment, however, it remains vulnerable to sophisticated jamming where multiple adversaries cooperate. We extend previous research by formulating a comprehensive model of PLA that integrates hybrid arrays and by developing optimized jamming strategies that consider energy and information constraints. Our results demonstrate that angle-based authentication in analog arrays are vulnerable to cooperative jamming, compared with hybrid arrays. Additionally, the combiner design in the hybrid array, along with the energy and information constraints on jamming strategies, significantly influences the success of jamming. By identifying vulnerabilities in PLA and studying effective countermeasures, this work contributes to advancing physical layer authentication in 6G systems.

Electricity theft detection (ETD) is a vital global concern that affects utility providers (UPs), and non-technical losses (NTLs) are a major issue. While smart metering has helped to reduce conventional technical losses (TLs), NTLs caused by theft are still difficult to identify and have serious effects. Consumers frequently underreport their power consumption, which complicates detection attempts. This paper provides a comprehensive review of various ETD approaches, categorizing them into five areas: (i) NTL detection with synthetic data, (ii) sequential data-based schemes, (iii) non-sequential data analysis, (iv) neighborhood area-based networks (NANs), and (v) IoT and hardware-based solutions. Each area is presented using case examples that have been statistically, mathematically, and visually analyzed. The article also includes a summary table of known problems and possible solutions, making it a useful resource for academics and developers. A comparison analysis utilizing F1 scores is presented to assess the efficacy of different detection strategies. This is the first review of its type to convert technical articles into real-world case studies, offering useful insights for selecting the best theft detection technologies in smart grids.

The rapid addition of diverse Internet-of-Things (IoT) assets to enterprise and operational networks has created significant challenges for network operators in effectively understanding and managing their attack surfaces. Networks can only be protected when all connected assets are accurately mapped and their intended behaviors are well understood. Researchers have attempted to use machine learning algorithms to model the network behavior of IoT devices by passively analyzing traffic packets and flows. However, existing approaches often focus on a limited subset of the available network traffic information, resulting in models that lack flexibility and generalizability across environments with diverse network compositions, computing resources, and device types and behaviors. We advocate an approach that utilizes deep learning to comprehensively capture both major and subtle patterns in network traffic, with the ability to adapt to various contexts in the future. Our specific contributions are threefold. (1) We develop a configurable data structure that represents network traffic in a fixed-size matrix, including flow metadata, packet direction, timing and raw payloads, and sequences of packets and flows. By analyzing a large dataset of IoT packet traces, we generate our structured traffic data, which we will publicly release, and draw insights into behavioral patterns exhibited by IoT devices; (2) We develop a two-dimensional convolutional neural network (CNN) architecture capable of decoding intra-flow and interflow patterns specific to various device types without requiring specialized supervision regarding the application-layer protocols. We will experiment with three representative strategies for traffic inference and highlight their pros and cons based on four metrics: accuracy, coverage, computational workload, and need for traffic selection; and, (3) We develop a method that utilizes confidence scores alongside Shapley values to explain and verify model predictions. When applied to real traffic traces, we demonstrate how the explainability insights derived from our model can help refine the neural network, leading to higher-quality predictions.

In this paper, an optimized hardware architecture is presented for a motor imagery-based Brain-Computer Interface (MI-BCI) tailored for real-time, wearable applications that demand high accuracy and low power consumption. The design captures and filters EEG signals to isolate motor imagery patterns, utilizing Linear Discriminant Analysis (LDA) for feature extraction followed by Euclidean distance-based classification. Key optimizations, including sequential matrix multiplication, pipelining, and an FSM-based control unit, achieve 91.67% accuracy with a power consumption of 2 mW on 180 nm technology, significantly enhancing performance compared to conventional designs.

AI-generated images (AIGIs) can be created by prompts with unlimited creativity. However, like deep-fake videos, generated contents could result in social problems. In this paper, we focus on disgusting AIGIs, which could be very emotionally negative to viewers. We show that disgusting images can be easily created, both intentionally and accidentally. We also analyze both types of disgusting images by measuring the DisgustRate, which is defined as the percentage of disgusting images over the total of generated images. Our findings show that the DisgustRate of intentionally disgusting images is higher than accidentally disgusting images. In addition, we built a dataset of disgusting images and investigated the potentials of AI models to detect this kind of AI-generated images. We found that CLIP performs exceptionally well in disgusting image classification, achieving a test accuracy of up to 86%.

The distributed ledger technology (DLT) landscape comprises a wide range of independent networks with little to none built-in interoperability. To be applied in traditional enterprises, these DLT systems must also interact with legacy systems with no support for the processes of a DLT. An important topic in DLT research is, therefore, to establish standardised protocols for cross-network transfer and exchange of data and assets. Ideally, these protocols should be general-purpose so that they can be applied on top of many different types of ledger systems. One such protocol, supporting asset transfer, is the Secure Asset Transfer Protocol (SATP or 'SAT protocol') in development by the Internet Engineering Task Force (IETF). The SATP Core protocol draft by Hargreaves et al. (2024) describes an interoperability protocol that can facilitate asset transfer between two DLT systems, as well as between a DLT system and a non-DLT system. In either case, the SAT protocol imposes no restrictions on the underlying system implementations. Building on this work, in this paper, we present an adaptation of SATP that facilitates cross-network asset exchanges. This asset exchange protocol, named the Secure Asset Exchange Protocol, inherits the key advantages of SATP but enables asset exchanges instead of asset transfers.

Society’s dependence on telecommunications networks has increased significantly over the past 30 years. The expectations of always-available access, coupled with the demands for real-time video communications has driven a significant increase in network complexity. At the same time, there has been a shift to move industrial and machine communications from dedicated networks to shared networks, and an explosion of devices accessing these networks for a multitude of consumer and enterprise applications. The increased reliance on very complex networks requires a holistic approach to understanding the potential critical failure points in order to build the required network resilience. Driven by, and driving, this complex network transformation has been the the adoption of Internet Protocol (IP) not only for data services, but also mobile voice, Voice over Internet Protocol (VoIP) and interactive video services. In addition, IP networks are being increasingly used for critical industrial and commercial applications. This transition, however, exposes the fragility of centralised IP routing protocols, common distributed IP functions with common configuration options and the respective management systems. This paper examines the shift from traditional telecommunication redundancy models, which emphasised physical redundancy and the geographic isolation of failures, to the current IP-centric networks where even minor configuration errors can precipitate national-scale outages. Through a detailed analysis of the critical functions employed by mobile and fixed-line operators, we identify specific IP and mobile functions that demand heightened scrutiny to prevent cascading failures or failures that are difficult to troubleshoot. We present case studies illustrating the severe consequences of such failures on network services and their broader economic and corporate ramifications. Furthermore, we explore potential mitigation strategies and highlight the necessity for ongoing research to develop more robust network architectures that can withstand the inherent vulnerabilities of IP networks. This study aims to contribute to the enhancement of service resiliency in the face of evolving mobile and fixed network demands and complexities.

Bionic limb control through myoelectric pattern recognition, offering intuitive decoding of motor intent, can improve the quality of life for individuals with amputations. However, most work on pattern recognition only uses a small subset of the myoelectric data generated during daily life, to train an Artificial Neural Network (ANN) via Supervised Learning (SL). Scenarios substantially different from the recording session e.g. different limb positions, can lead to misclassifications by the ANN during everyday usage of the bionic limb. Recording labeled data from all scenarios encountered in daily life could alleviate the problem, but would be prohibitively time consuming. Unsupervised Domain Adaptation (UDA) offers a solution by leveraging unlabeled data from a target domain, not represented in the labeled dataset i.e. the source domain, to calibrate ANNs for improved performance. In this study we explore the potential of two UDA algorithms for domain shifts in myoelectric pattern recognition: Domain Adversarial Neural Networks (DANN) and Sliced Wasserstein Discrepancy (SWD). Offline evaluation identified SWD as the best-performing algorithm, which was subsequently validated in online experiments with 11 participants. Using UDA substantially improved the performance, not only on the target domain, but also on the source domain. Indeed, it nearly matched the performance of an ANN trained through SL on labeled data from both the source and target domain. This is the first time UDA was validated in online experiments as a viable approach to overcome domain shift problems caused by changes in limb position.

This paper develops a non-linear composite similaritybased framework for generating univariate physiological vital signs data from an input multivariate counterpart. The framework is built on mixture random variate using information provided by the interrelationships among variables. This allows the latent unidimensional data to be generated as a weighted linear combination of the multivariate data, providing an easy way to model the weights in terms of desirable data features of interest. Using variable specific non-linear composite similarity statistic to handle short, medium-and long-term auto-relationships, the framework provides a unified context for easy quantification and assessment of both vital sign and observation level relative relevance. With the above formulation, a better calibration and indication of key vital signs in traumatic events is presented. An illustrative example using real physiological vital sign data on trauma patients provides evidence on its utility in handling traumatic vital signs.

Data centers increasingly depend on Operational Data Analytics (ODA) for real-time insights from vast streams of telemetry data generated by IoT devices. Effective ODA enables organizations to make informed decisions, optimize operations, and enhance system performance. While data centers typically utilize NoSQL databases for scalability and data diversity, this often results in unstructured data representation, which poses significant challenges for querying. The lack of standardization, combined with schema flexibility and complex data structures, makes it difficult for system administrators to write and execute queries, ultimately complicating the automation of data retrieval tasks. While large language models (LLMs) offer opportunities to simplify data retrieval through natural language input, they frequently generate inaccurate or hallucinated query code. This manuscript presents EXASAGE, the first ODA co-pilot utilizing a knowledge graph-based approach to address these limitations. EXASAGE employs an LLM agent to convert natural language into SPARQL queries (native to knowledge graphs), executed at a graph database endpoint. In evaluations on 1,000 prompts, EXASAGE achieved a 92.5% accuracy rate in generating correct SPARQL code, significantly outperforming the 25% accuracy of NoSQL/SQLite query queries, which exhibited frequent hallucinations. Furthermore, SPARQL queries demonstrated greater conciseness, faster execution times, and shorter inference durations, underscoring their suitability for managing real-time IoT data in data centers.

The importance of a measurement of power system inertia in the context of the continuing replacement of synchronous generators with renewable energy sources is considered. A detailed description of the frequency time-series associated with the power system response to the loss of a major source of power in-feed is discussed. The purpose of this is to identify what part of the response is operationally associated with inertia, such that it can be measured and subsequently used for grid control and planning. A novel reference method is proposed, in particular, a signal processing transform to decompose the frequency time-series in order to isolate the operational inertial response. Power system modeling of a major loss event on the GB grid is used to verify the method. A measurement campaign is described and measurements from 23 major infeed loss events in the GB grid are presented. Inertia results from the new method and associated uncertainties are compared with the electricity system operator's estimates for each loss event.

The growth of the Electric Vehicle (EV) fleet facilitates a shift towards carbon-neutral transport, yet the power grid faces challenges with the increasing number of chargers. Developing smart charging algorithms for conventional and bidirectional charging EVs requires detailed knowledge of their charging characteristics. Currently, there are only a few public datasets with real-world EV charging profiles, especially for bidirectional charging. This paper introduces an open-access dataset with 142 EV charging profiles, focusing on the static and dynamic aspects of charging and discharging. The dataset has various applications, such as providing a resource for academic research in simulation studies or for the planning and operation of charging points. In addition, it provides valuable insights for policymakers and regulators seeking to understand the impacts of EV charging on grids. In this paper, a numerical analysis of reactive power characteristics, current harmonics, and bidirectional DC charger efficiency is performed as illustrative examples of its application. Furthermore, the contribution of this dataset for comparing performance and validating state of charge estimation models is emphasised. The dataset is publicly available for download here: https://doi.org/10.5281/zenodo.14065331