One of the most relevant gains of pure 5 G mobile communication, i.e., Stand Alone (SA) architecture is throughput, as it enabled the addition of new features, including Voice over New Radio (VoNR) that is one of the IP Multimedia Subsystem (IMS) features, which also has other multimedia services, such as Voice over Long Term Evolution (VoLTE) and Voice Over WiFi (Vo-WiFi). Analyzing the energy consumption in mobile devices that support the Fifth Generation (5G) in the SA architecture plays an essential role in developing these devices, which commonly have limited hardware resources. This research aimed to evaluate the impacts of battery consumption during the attachment procedure leading up to VoNR call sessions across different 5 G frequency bands. The frequency band that consumed the least energy was n20.
Tiny Machine Learning (TinyML) enables on-device inference on low-power microcontrollers, but a single optimised model often cannot meet real-time and energy constraints under variable workloads. This paper presents a simulation-based approach to dynamic model or ensemble selection, where a device switches between pre-profiled models at run time to balance latency, accuracy and energy consumption. We introduce a configurable simulator that evaluates selection strategies without hardware experiments. The simulator models a real TinyML system. It supports synchronous and asynchronous samples patterns, tracks energy consumption using a simplified current-based model, and reports metrics including operating time, number of inferences, dropped samples, and weighted accuracy. The tool provides a repeatable and customizable environment for testing resource-aware inference strategies in TinyML systems. The simulator is open source and intended to aid the development of adaptive TinyML applications where responsiveness and energy efficiency must be carefully balanced.
With the widespread adoption of IoT devices, the need for enhanced security has become increasingly important. In Japan, the JC-STAR (Japanese Cyber-Security Technical Assessment Requirements) has been introduced, but the evaluation process remains labor-intensive and highly depends on individual expertise. This paper proposes a method for automating JC-STAR document assessment using Large Language Models (LLMs). Within a Retrieval-Augmented Generation (RAG) framework, the approach incorporates PDF-toMarkdown conversion and two-stage weighted Reciprocal Rank Fusion (RRF), enabling accurate and comprehensive document chunk retrieval. Additionally, a text classification LLM is used to quantify the conformance level of LLM responses, automating the scoring process. Experiments show that the proposed method achieves a lower Mean Absolute Error (MAE) of 0.259 compared to other approaches, demonstrating its effectiveness.
The monitoring of the route and track environment plays an important role in automated driving. For example, it can be used as an assistance system for route monitoring in automation level Grade of Automation (GoA) 2, where the train driver is still on board. In fully automated, driverless driving at automation level GoA4, these systems finally take over environment monitoring completely independently. With the help of artificial intelligence (AI), they react automatically to risks and dangerous events on the route. To train such AI algorithms, large amounts of training data are required, which must meet high-quality standards due to their safety relevance. In this publication we present an automatic method for assuring the quality of training data, significantly reducing the manual workload and accelerating the development of these systems. We propose an open-source tool designed to detect nine common errors found in multi-sensor datasets for railway vehicles. To evaluate the performance of the framework, all detected errors were manually validated. Six issue detection methods achieved 100
This paper investigates the enhancement of energy detection (ED) techniques in cooperative wideband spectrum sensing with subbands for IoT and 5G networks, focusing on both AWGN and Rayleigh fading environments. We propose a comprehensive framework that integrates adaptive thresholding, adaptive k-out-of-N fusion rule, and diversity combining via maximal ratio combining (MRC) to improve detection reliability while maintaining energy efficiency in dynamic spectrum access (DSA) scenarios. Cooperative sensing across subbands is leveraged to address the inherent variability in spectrum occupancy and to enable efficient spectrum sharing. The framework’s performance is evaluated using analytical derivations and extensive MATLAB simulations, considering key metrics such as the probability of detection (PD), probability of false alarm (PFA). Our findings reveal that the proposed adaptive scheme consistently outperforms conventional static approaches, particularly in low-SNR Rayleigh fading environments, thus offering a robust solution for smart city applications that rely on dynamic spectrum access and reliable spectrum sharing across subbands. Finally, we highlight potential extensions of this work to include machine learningbased optimization for adaptive parameter selection in highly dynamic environments.
Distance education has traditionally relied on screen-based methods such as audio/video lectures and online forums. Extended Reality (XR) provides immersive learning experiences that can improve retention, support experiential learning, and enable metaverse applications. This study surveyed 92 faculty and 268 students at a large distance-learning university to assess XR familiarity, adoption potential, and barriers. Faculty reported limited XR experience, while students reported greater familiarity. Key obstacles to integration included high costs, insufficient training, and technical support challenges. While faculty were cautiously optimistic, students expressed strong enthusiasm for XR’s educational value. Participants identified priority applications such as training, collaborative learning spaces, scientific visualization, virtual field trips, and personalized learning tools. Based on these insights, a phased roadmap is proposed for sustainable, institution-wide XR integration toward a global “Eduverse."
Next-generation industrial automation in Industry 5.0 demands sub-nanosecond time synchronization to enable advanced human-machine collaboration and real-time control systems. Quantum Time Synchronization (QTS), enabled by entangled photon-based quantum networks, offers ultra-precise clock alignment across distributed nodes. However, practical deployment requires an authenticated classical communication channel to exchange timestamped data and extract synchronization offsets. This paper investigates the core tradeoffs between classical channel latency and achievable QTS precision in hybrid quantum-classical networks. Using a discrete-event simulator, we evaluate the impact of various classical network protocols, each with distinct latency characteristics, on QTS performance. Our results demonstrate that even modest latency increases result in degraded synchronization accuracy and increased synchronization jitter, particularly in time-critical industrial contexts. To address these challenges, we propose a combined approach that integrates low-overhead authentication, low-latency communication protocols, and predictive clock drift compensation. Our findings provide actionable guidance for balancing latency, security, and synchronization accuracy in QTS services for future networks.
This paper presents a lightweight convolutional neural network designed for Few-Shot Learning (FSL) in OCT image classification. Building on our previous FSL training strategy, we adopt the Half-Append Half-Add (HAHA) Block as the backbone and introduce slight modifications to improve performance. To further enhance feature extraction, we integrate a Squeeze-and-Excitation Block (SE Block) into the architecture. Eventually, with optimized knowledge distillation settings, the proposed model achieves a peak accuracy of $97.87 \%$, while maintaining an ultra-compact size of only 6.6 K parameters and a computational cost of 0.028 GFLOPs.
This paper proposes a user-configurable approach to energy-efficient machine learning inference in TinyML devices. TinyML is the deployment of machine learning models on lowpower, resource-constrained devices at the edge. Rather than relying on a single monolithic model, our method allows switching between classification models of varying complexity based on user preferences or system conditions. Inspired by power-saving modes in consumer electronics, we introduce a method where users can select between a high-accuracy or energy-saving mode. Using a Random Forest classifier on the HAR, WISDM and PAMAP2 datasets and deploying models to STM32MP257F MPU, we demonstrate that varying the number of decision trees allows a balance between prediction accuracy and energy consumption. Experimental results show that energy savings (depending on the mode) can be achieved with minimal loss of classification quality. This flexible model-switching approach offers a promising direction for future adaptive TinyML systems, especially in battery-constrained environments.
This article introduces a novel interdisciplinary framework linking extended tunnelling in photosynthetic pigment-protein complexes with false memory formation in neural systems. We analyze the quantum dynamics of excitonic transport, highlighting how vibronically assisted, non-local pathways can generate transient “false” energy channel deviations from canonical routes that emerge under specific system-bath interactions. While typically suppressed by decoherence, these channels may persist, reflecting intrinsic inefficiencies akin to tunnelling leakage. Drawing a conceptual parallel, we examine how maladaptive synaptogenesis and neural redundancy in cognitive systems may analogously produce false memories not from sensory errors, but from failures in distinguishing overlapping spatiotemporal patterns. These failures mirror quantum tunnelling anomalies and can be quantified as network-level susceptibility rates. We propose a unified model grounded in probabilistic network dynamics, suggesting that quantum-inspired false channel formation offers a metaphor and possibly a computational framework for understanding memory distortion. Extending this to cybersecurity, we argue that such cognitive vulnerabilities can be exploited through strategic information priming, forming the basis for advanced social engineering tactics. By bridging quantum biology, neuroscience, and cybersecurity, this work provides a theoretical synthesis with practical implications, offering new pathways for detecting deception, mitigating cognitive attacks, and informing resilient, bio-inspired AI design.
The rapid evolution of Large Language Models (LLMS) and their increasing adoption by enterprises for diverse response generation, combined with breakthroughs in quantum computing, challenge the viability of using traditional cryptographic methods to secure communication with LLMs. This work introduces a post-quantum-secure communication framework designed to safeguard interactions between LLMs and users. The proposed security framework employs CRYSTALS-Kyber, a Key-Encapsulation Mechanism (KEM), to provide security against quantum adversaries. Application Programming Interface (API)-based communication is secured through a protocol that balances security and efficiency in client-server exchanges. Additionally, the framework also employs CRYSTALS-Dilithium, a post-quantum-secure digital signature scheme, to strengthen authentication and guarantee data integrity in client-server communications. The proposed protocol is implemented using LLaMA 3.1 deployed through Ollama, with Python-based serverside and client-side applications. By mitigating the risks introduced by advances in quantum technologies, this work aims to provide a scalable and efficient solution that strengthens data privacy and integrity for applications leveraging LLMs.
The modern data-centric organizations of today have made the security of databases a necessity, requiring a robust Database Intrusion Detection System (DIDS) to defend against external attackers and prevent internal privilege misuse. In this paper, we propose a DIDS that combines a Convolutional Autoencoder (CAE) integrated Self-Attention Long Short-Term Memory (SA-LSTM) network, optimized using a novel Adaptive hybrid Archimedes Optimization-Flamingo Search Algorithm (Ah(AO-FSA)) for role classification, and a modified Repetitive Nonoverlapping Pattern Miner (RNP-Miner) for data dependency rule mining from a sequence database. We also propose the Amiability Index ($\mathbf{A I}_{\mathbf{d}}$), a novel multi-phase similarity score metric that classifies the transaction as normal or malicious. Experiments on a synthetic dataset based on the TPC-C benchmark show that our proposed approach achieves a 99.33% detection accuracy.
The increasing consumption of streaming content on mobile devices and its transmission to secondary screens, such as televisions, reinforces the need for reliable methods to assess the quality of the video consumed. The central challenge of this work lies in the ability to perform a reliable analysis, capable of detecting subtle variations between the original and received content, which can be introduced by factors such as compression and losses in the wireless network. The methodology uses the Android Debug Bridge (ADB) to capture frames from the mobile device, OpenCV for image processing, Peak Signal-toNoise Ratio (PSNR) to indicate the noise level of the wireless network, K-Means to partition the data into K clusters, where each data point belongs to the cluster with the closest centroid and Structural Similarity Index Measure (SSIM) to assess the similarity between the transmitted and received data. The results demonstrate that the quality of the transmitted video reaches levels above $98 \%$ similarity with the original content, indicating that the main degradation factor is the compression intrinsic to the transmission process via the Wi-Fi network. This finding gives high reliability to the proposed method, even in potentially noisy scenarios, consolidating it as an effective tool for evaluating the quality of video transmission between mobile devices and secondary screens.
In recent years, IoT (Internet of Things) devices have attracted much attention, and IC products have become widespread in our daily lives. With the increase in demand for ICs, third-party companies have intervened in the design and manufacturing phase of ICs, increasing the risk of malicious circuits called hardware Trojans (HTs) being inserted during these phases. As one of the methods to detect HT in the design phase, an HT detection method using graph learning for circuit design information has been proposed, and relatively high HT detection accuracy has been reported. In this paper, we propose a correction method using multiple trained graphlearning models to improve the accuracy of graph-learning based HT detection results. After applying the proposed correction method, the average F -score improve to 0.8861, while TPR and the precision improve to $89.24 \%$ and $93.31 \%$, respectively.
In this paper, we propose Mini-ResNet, an ultralightweight deep learning model designed for the early diagnosis of attention-deficit/hyperactivity disorder (ADHD). This model processes raw electroencephalogram (EEG) data, input through stride-based slicing and augmentation, which circumvents the need for complex preprocessing. We successfully reduced the size of the conventional ResNet18 model from 42 MB to approximately 333 KB through architectural optimization. Experiments were performed on a publicly available ADHD dataset, employing 5-fold crossvalidation and exploring various optimizer configurations. Our Mini-ResNet exhibited superior performance over MobileNetV2, achieving an average accuracy of up to $84.3 \%$ and a higher F1-score. The model’s real-time applicability in mobile environments underscores its potential as an economical ADHD screening solution for both domestic and public health settings.
The influx of smart consumer devices equipped with sensors and high computational capabilities has enabled various important health monitoring applications to be placed in the hands of the consumer. This evolution is potentially transforming the healthcare delivery practice, which is still mainly located in clinical environments. This paper investigates comprehensively electrocardiogram (ECG) vital sign monitoring technologies in consumer medical devices with the goal of assessing and analyzing the maturity of the hardware. The analysis is focused on the evaluation of 12 consumer ECG devices from four manufacturers. We define five maturity dimensions (Technical, Validation, Regulatory, Market, and Clinical Integration) with corresponding scoring scale of 1-5. Based on information gathered from an extensive literature review, we evaluate the scores for each dimension for all considered devices. Based on these scores, we perform a detailed statistical analysis and identify development pathways: Technical-Led, Regulatory-Led, and Market-Driven approaches.
The rapid advancements in Generative AI (GenAI), particularly the Large Language Models (LLMs) present transformative opportunities for enhancing smart home environments. This research investigates the integration of LLM agents as reasoning entities into smart homes, with a focus on optimizing energy efficiency, improving indoor living conditions, and addressing safety and security challenges. By leveraging their natural language processing and reasoning capabilities, LLM agents are used to analyze the data from sensors deployed across the smart home to identify usage patterns, detect anomalies, and generate actionable recommendations. The proposed framework demonstrates how LLM agents can optimize appliance settings, enable predictive maintenance, and mitigate vendor lock-in by providing a versatile and non-proprietary solution for intelligent home management. Key contributions of this research include the development of a platform for monitoring and maintaining energy efficiency and living comfort, as well as issuing structured recommendations aligned with standard living guidelines. The findings highlight the potential of LLMs in transforming smart home ecosystems, paving the way for more sustainable, efficient, and user-centric living environments.
Quantum Key Distribution (QKD) provides a provably secure method for key exchange based on quantum mechanics. However, the deployment of QKD in Internet-ofThings (IoT) environments remains challenging due to the resource constraints of typical edge devices. In this work, we explore the use of the Guessing Random Additive Noise Decoding (GRAND) algorithm as a lightweight alternative to conventional error correction techniques in QKD postprocessing. GRAND decodes by directly guessing noise patterns, eliminating the need for iterative belief propagation or large matrix storage. We evaluate soft-output (SO) GRAND on short block-length linear codes under varying channel conditions. Results show that SOGRAND achieves lower frame error rates (FER) and average decoding times than LDPC in typical QKD scenarios, demonstrating its potential for resource-constrained quantum-secure IoT systems.
Magnetic field coupling wireless power transfer (WPT) methods have become mainstream for near-field wireless charging, and in the future, its application is expected to expand beyond consumer electronics products to electric vehicles, industrial products, and other areas. The inductive coupling method generates an AC magnetic field, which raises concerns about its impact on the human body. In this paper, we propose a lightweight and low-cost aluminum multi-ring magnetic shielding structure. Furthermore, we use deep learning to find the optimal structure that satisfies the electromagnetic field (EMF) emission limit. The analysis results show that the proposed structure can reduce weight while suppressing leakage magnetic flux, and that the proposed method utilizing AI enables high-speed optimal design.
Inertial MEMS sensors have been integral to consumer electronics for over two decades, initially enabling functionalities like screen rotation and motion-based gaming. With the pervasive use of smartphones, earbuds, and smartwatches, new challenges arise regarding user experience. Specifically, many people suffer from motion sickness from reading or working on mobile screens (such as mobile phones, laptops, etc.) while travelling in a car or other vehicles. Displaying motion cues is a promising approach but needs more than accelerometer data to indicate the orientation to alleviate or even eliminate the symptoms of motion sickness. Another common user experience shortcoming is wind noise outdoors or while riding on a bike, which makes telephone communication almost impossible. Bone conduction microphones are known for harsh environments. The challenge for consumer electronics is the integration in miniaturized earables and with good audio capturing quality. This paper examines how advanced inertial MEMS sensors and sensor fusion techniques can address these challenges, presenting practical implementation examples.
