Electricity theft is a major source of non-technical losses and poses a significant operational challenge for distribution system operators. While advanced metering infrastructure offers large-scale electricity consumption (EC) data, existing deep learning methods remain computationally expensive, prone to overfitting, and sensitive to class imbalance and high-dimensional inputs. This work proposes a computationally efficient surrogate-activated spiking neural network (SSNet) for electricity theft detection (ETD). SSNet employs leaky integrate-and-fire (LIF) neurons with surrogate gradients to capture temporal patterns in discrete EC measurements while substantially reducing computational overhead compared with conventional deep neural networks. To mitigate class imbalance, the proximity-weighted synthetic oversampling (ProWSyn) technique is integrated to generate informative minority-class samples. In addition, a set of discriminative time-series features is extracted using the time series feature extraction library (TSFEL) to address high-dimensionality and enhance the separability of theft and non-theft profiles. SSNet is evaluated against several deep learning baselines and demonstrates superior performance, achieving 96.64% accuracy, 99.34% area under the curve (AUC), and 93.28% Matthews Correlation Coefficient (MCC). Moreover, SSNet operations per second (OPs) comparison with the floating-point operations per second (FLOPs) of deep learning networks demonstrates the sparsity and energy efficiency of the proposed network. The results highlight the potential of spiking neural architectures as efficient and high-performance solutions for large-scale ETD.

Human Activity Recognition (HAR) using radar signals has gained significant attention due to its non-intrusive nature and robustness in various environments. However, the impact of radar signal preprocessing techniques on the performance of deep learning (DL) models remains an active area of research. This study investigates how different radar domain representations affect HAR accuracy by evaluating four preprocessing methods: Time-Range (TR) maps generated via Range-Fast Fourier Transform (FFT), Range-Doppler (RD) maps obtained through sequential FFTs, and Time-Doppler (TD) features extracted using Short Time Fourier Transform (STFT) and Smoothed Pseudo Wigner Ville Distribution (SPWVD). We employ a baseline Convolutional Neural Network (CNN) and state-of-the-art Transfer Learning (TL) models to assess whether advanced preprocessing or increased model complexity yields greater performance gains. The results reveal that high-resolution TD analysis using SPWVD does not significantly enhance classification performance and incurs substantial computational overhead, limiting its real-time applicability. Conversely, the TR representation offers computational efficiency but struggles to classify complex activities with the baseline CNN accurately. RD and STFT methods provide a favorable balance between classification accuracy and computational efficiency. Notably, transitioning from the baseline CNN to TL models leads to substantial improvements in recognition accuracy: up to 29.36% for TR, 21.42% for RD, 16.66% for STFT, and 11.11% for SPWVD representations. Overall, our findings demonstrate that TL models, when combined with computationally efficient radar preprocessing techniques like RD or STFT, significantly improve recognition accuracy and generalize well across datasets, as confirmed by evaluation on two publicly available radar-based HAR datasets. Among these, the RD representation combined with VGG-19 yielded the best trade-off between accuracy and latency, achieving a total processing time of 0.91 s per sample for a 10 s activity duration, making it highly suitable for latency-sensitive HAR applications.

Millimetre-wave (mmWave) communication systems use large antenna arrays and narrow beams to achieve strong signal power. However, this approach requires extensive beam training, which leads to high overhead. Recently proposed vision-aided beam prediction methods show promising results, reducing this overhead. However, these techniques have considerable computational complexity, hindering practical deployment. To address this issue, we propose a Semantic-Aware Quantised Network (SemQNet) framework that leverages image compression and a lightweight computer vision model to extract semantic information used for training a fully connected neural network (FCNN). Additionally, the proposed SemQNet also uses quantisation-aware training (QAT), which enables low-precision arithmetic operation, reducing the model size in the training process. Our tests on the DeepSense 6G dataset show that SemQNet achieves almost the same top-1 accuracy as existing vision-based methods while reducing the model size by 74.21%. This smaller model size reduces the communication overhead, making SemQNet a practical and efficient solution for energy-constrained mmWave communication systems.

Predicting signal blockages in millimetre-wave and terahertz networks is essential for enabling proactive handover (PHO) and ensuring seamless connectivity. Existing approaches utilising deep learning, multi-modal vision and wireless sensing data primarily depend on centralised model training. Although these techniques are effective, they come with high communication costs, inefficient bandwidth usage, and latency issues, which restrict their real-time applicability. This paper proposes a Semantic-Aware Federated Blockage Prediction (SFBP) framework, leveraging the lightweight computer vision technique MobileNetV3 for edge-based semantic extraction, lowering communication and computation costs. Furthermore, we introduce a Similarity-Driven Federated Averaging (SD-FedAVG) mechanism to enhance the robustness of the model aggregation process, effectively mitigating the impact of noisy updates and adversarial attacks. Our proposed SFBP framework achieves 97.1% blockage prediction accuracy, closely matching centralised learning methods, while reducing communication costs by 88.75% compared to centralised learning and by 57.87% compared to FL without semantic extraction. Moreover, on-device inference reduces the latency by 23% compared to centralised learning and 18% compared to FL without semantic extraction, improving real-time decision-making for PHO. Additionally, the SD-FedAVG mechanism improves prediction accuracy under noisy conditions, directly impacting the PHO by reducing the handover failure rate by 7%.

Human Activity Recognition (HAR) using Channel State Information (CSI) enables energy-efficient and non-invasive healthcare monitoring. However, conventional HAR systems rely on centralised model training, which requires the sharing of raw data, leading to privacy risks, excessive bandwidth usage, and high communication latency that limit scalability. This paper proposesFedFusionQuant (FFQ), a federated learning (FL) framework that jointly performs feature fusion, adaptive aggregation, and quantisation-aware compression during training. A novelfederated distance (FedDist)mechanism dynamically adjusts parameter updates using neuron dissimilarity metrics, enhancing generalisation across heterogeneous clients. Meanwhile,quantisation-aware training (QAT)reduces model size and transmission cost while preserving accuracy. Extensive experiments on real CSI data from 30 participants demonstrate that FFQ improves multi-class HAR accuracy by4.29%and binary fall detection by5.55%compared to raw fusion models. Furthermore, model compression with QAT achieves a47% reduction in communication overheadwhile maintaining accuracy comparable to state-of-the-art methods.

The rapid evolution of drone communication systems necessitates the development of novel approaches for optimal beam management in millimetre wave (mmWave) 6G networks. Beamforming is used to improve signal quality and enhance the signal-to-noise ratio (SNR); however, the existing beam management performs an exhaustive search over the pre-defined codebook, resulting in higher latency due to training overhead that makes it impractical for high-mobility applications. Therefore, this paper introduces an innovative technique for mmWave beam prediction, considering practical visual and communication scenarios. The approach proposed in this study utilizes computer vision (CV) and ensemble learning via stacking, combining multi-modal vision sensing and positional data to achieve accurate estimations of drone positions and orientations. The developed framework first fine-tunes "you look only once" version 5 (YOLO-v5), a CV model to obtain the bounding box (location) of the drone from RGB images. This filtered vision sensing information and position data are used to train two different sets of neural networks, and the output of each model is stacked to train a meta-learner, used for the prediction of K-beams from a pre-defined codebook. The proposed method outperforms with the top-1 accuracy of approximately 90% compared to 86% and 60% for vision and position models, respectively. Furthermore, top-3 and top-5 accuracies are approximately 100%, resulting in a significant receive signal strength.

Federated Learning (FL) is susceptible to backdoor adversarial attacks during the training process, which poses a significant threat to the model's performance. Existing adversarial mitigation solutions mainly rely on the neural network (NN) model statistics and discard an entire client model if attacked. This approach is not feasible as it results in suboptimal performance. Hence, it is crucial to develop lightweight backdoor attack mitigation solutions that efficiently utilize clients' model statistics. To address this issue, we propose (Layer Based Anomaly Aware) LBAA-FedAVG, a modified version of the common aggregation mechanism FedAVG. Our proposed framework employs a clustering-based technique and addresses each NN layer individually. Depending on the type of adversarial attack, this method selectively eliminates one or multiple layers of the NN during the aggregation process. Furthermore, we focused on the model inversion attack and varied the percentage of compromised clients from 10% to 50%. Our experimental findings demonstrate that LBAA-FedAVG outperforms Federated Averaging (FedAVG) in reducing the negative effects of backdoor adversarial attacks. The complexity analysis suggests that the extra training time is the only additional resource limitation in LBAA-FedAVG, which is 19% greater than that of FedAVG. Additionally, we conducted experiments on short-term load forecasting using grid-level datasets to show the effectiveness of LBAA-FedAVG in lightweight backdoor attack mitigation in FL settings, offering a trade-off between time efficiency and enhanced defense.

Smart energy performance monitoring and optimisation at the supplier and consumer levels is essential to realising smart cities. In order to implement a more sustainable energy management plan, it is crucial to conduct a better energy forecast. The next-generation smart meters can also be used to measure, record, and report energy consumption data, which can be used to train machine learning (ML) models for predicting energy needs. However, sharing energy consumption information to perform centralised learning may compromise data privacy and make it vulnerable to misuse, in addition to incurring high transmission overhead on communication resources. This study addresses these issues by utilising federated learning (FL), an emerging technique that performs ML model training at the user/substation level, where data resides. We introduce FedraTrees, a new, lightweight FL framework that benefits from the outstanding features of ensemble learning. Furthermore, we developed a delta-based FL stopping algorithm to monitor FL training and stop it when it does not need to continue. The simulation results demonstrate that FedraTrees outperforms the most popular federated averaging (FedAvg) framework and the baseline Persistence model for providing accurate energy forecasting patterns while taking only 2% of the computation time and 13% of the communication rounds compared to FedAvg, saving considerable amounts of computation and communication resources.

The upsurge in wireless devices and real-time service demands force the move to a higher frequency spectrum. Millimetre-wave (mmWave) and terahertz (THz) bands combined with the beamforming technology offer significant performance enhancements for future wireless networks. Unfortunately, shrinking cell coverage and severe penetration loss experienced at higher spectrum render mobility management a critical issue in high-frequency wireless networks, especially optimizing beam blockages and frequent handover (HO). Mobility management challenges have become prevalent in city centres and urban areas. To address this, we propose a novel mechanism driven by exploiting wireless signals and on-road surveillance systems to intelligently predict possible blockages in advance and perform timely HO. This paper employs computer vision (CV) to determine obstacles and users’ location and speed. In addition,this study introduces a new HO event, called block event (BLK), defined by the presence of a blocking object and a user moving towards the blocked area. Moreover, the multivariate regression technique predicts the remaining time until the user reaches the blocked area, hence determining best HO decision. Compared to conventional wireless networks without blockage prediction, simulation results show that our BLK detection and proactive HO algorithm achieves 40% improvement in maintaining user connectivity and the required quality of experience (QoE).

Education for sustainable development (ESD) is one of the key UN strategies, which aims to encourage the development of competencies such as critical thinking and decision making. To ensure that students benefit from this education, there is general consensus that teaching methods need to focus on experiential learning. This may be challenging during a pandemic, or when experimental resources are limited. Therefore, the purpose of this paper is to showcase our approach in teaching solar energy systems design to higher education students using a game-based virtual reality approach. Our approach enables students to immerse themselves in a virtual environment, which is safe for both students and their teachers. The game consists of two levels, where students are invited to solve an energy-related task in two different homes. Based on user feedback, our interactive learning tool improved student awareness of solar energy systems and how they can be used to satisfy domestic energy demands.

The outbreak of the COVID-19 pandemic has caused unprecedented disruption to education and progressed remote teaching as a predominant model for delivering educational content. However, the online teaching and learning model has its challenges, such as the lack of technological tools to quantity the student attention and engagement with the learning content. This paper focuses on developing an e-learning framework for capturing and analysing the students’ attention during remote teaching sessions and subsequently profiling their learning behaviour leveraging eye-tracking data. Our proposed eye-tracking solution deploys a webcam to capture and track raw gaze points that grant the user the freedom of natural head movement and scalability compared to conventional eye-tracking approaches. We derived various gaze metrics in conjunction with state-of the-art machine learning (ML) models like logistic regression, support vector machine and polynomial regression to classify the student attention with an accuracy above 91%. Furthermore, our findings can help in the early detection and diagnosis of attention deficit hyperactivity disorder (ADHD) among students, thus supporting their learning journeys by creating an adaptive learning environment tailored to their needs.

The COVID-19 pandemic has accelerated our transition to an online and self-directed learning environment. In an effort to design better e-learning materials, we investigated the effectiveness of collecting psychophysiological eye-tracking data from participants in response to visual stimuli. In particular, we focused on collecting fixation data since this is closely related to human attention. Current wearable devices allow the measurement of visual data unobtrusively and in real-time, leading to new applications in wearable technology. Despite their accuracy, head-mounted eye trackers are too expensive for deployment on large-scale deployment. Therefore, we developed a low-cost, webcam-based eye tracking solution and compared its performance with a commercial head-mounted eye tracker. Four-minute lecture slides on the 3rdyear electronic engineering course were presented as stimuli to eight learners for data collection. Their eye movement was collected within the pre-defined area of interest (AOI). Our results demonstrate that a low-cost webcam-based eye-tracking solution, combined with machine learning algorithms, can achieve similar accuracy to the head-worn tracker. Based on these results, learners can use the eye tracker for attention guidance. Our work also demonstrates that these webcam-based eye trackers can be scaled up and used in large classrooms to provide real-time information to instructors regarding student attention and behaviour.