Unmanned Aerial Vehicle (UAV) swarms equipped with cameras are of growing interest for applications such as surveillance and object detection, transmitting real-time videos to a Ground Control Station (GCS) over the Internet. The quality of the videos received depends on both the available network bandwidth and the altitude of the drones. In this paper, a swarm of UAVs is given the task of scanning an area of interest while guaranteeing an overlap between the videos captured by each UAV for post-processing reasons, e.g. video stitching. Drones are enabled to adjust their altitude based on the available bandwidth to enhance visual quality. This work proposes a bandwidth-aware online path planner to maximize the coverage of a given area by extending the classic lawnmower path planner. The UAVs use Nonlinear Model Predictive Control (NMPC) to maximize both the total coverage area and the quality of the videos sent to the GCS. Simulations demonstrate the effectiveness of the proposed approach with different swarm configurations.
Integrating non-terrestrial networks (NTN) with terrestrial networks (TN) extends NB-IoT coverage but challenges random access (RA) with long delays, Doppler, and intermittent visibility. We introduce the Adaptive Hybrid Scheduling and Access (AHSA) framework, blending semi-persistent scheduling and contention control across TN/NTN using device classification, ephemeris-aligned windows, and dynamic Access Class Barring (ACB) with adaptive backoff. Monte Carlo (MC) evaluation averages collision and throughput trends over channel/SNR realizations, while battery lifetime is estimated via charge accounting. Results show AHSA maintains low collision rates, scales throughput, enhances energy efficiency, and ensures fairness across device classes, offering a practical MAC-layer solution for large-scale, energy-aware NB-IoT.
Ransomware poses a significant and evolving threat to modern information systems, often causing data loss and financial and reputation damage. Traditional detection mechanisms tend to focus either on known indicators of compromise or on overly specific behavioral patterns, which can limit their ability to detect novel or stealthy variants. In this paper, we propose, test, and evaluate a novel method for early-stage ransomware infection detection and prediction using sequence alignment techniques and deep learning-based sequence modeling. Specifically, we present an improved version of the well-known NeedlemanWunsch global sequence alignment algorithm tailored to detect partially matching suspicious behavioral patterns within network traffic flows. In parallel, we explore the use of transformer models to predict the continuation of these event sequences, enabling earlier detection and response. Our approach is inspired by bioinformatics methodologies, treating event sequences analogous to DNA analysis. Experiments conducted on a real-world ransomware network traffic dataset demonstrate the promising results of our method, both in aligning noisy, interleaved sequences and in accurately predicting the progression of ransomware behavior.
Battery technologies are widely used in various applications, including smartphones, wearable devices, laptops, and mobile phones. In this study, we propose a methodology specifically designed for segmenting defects in 2D CT slices of battery cells. This methodology involves the creation of datasets aimed at defect detection, along with benchmarking several U-Net-based models on these datasets. Our results indicate that the U-Net+Xception model achieves the highest binary accuracy of 0.9982 and an Intersection over Union (IoU) score of 0.8066, demonstrating its strong capability to differentiate between background and foreground regions. In summary, this work establishes a valuable baseline for CT-based defect detection in battery cells, showcasing the benefits of combining U-Net architectures with advanced pre-trained encoders. This research contributes to the development of scalable and automated inspection tools that can be integrated into battery manufacturing and diagnostic processes.
The paper presents the design of a fractionalorder frequency filter with a tunable Fractional-Order Element (FOE). The FOE is implemented using a distributed Resis-tive-Capacitive-Resistive with N-times resistance (R-C-NR) layer structure and provides the required phase values with minimal error across a wide frequency band. The structure enables tuning of the parameter alpha in the range from 0.6 to 0.71. The parameters and optimized interconnections of the FOE structure were obtained using genetic algorithms in MATLAB according to the specified requirements. The proposed filter exhibits an lowpass transfer function with an order ranging from 1.6 to 1.71. Furthermore, the cut-off frequency can be tuned within the range of 5 to 100 kHz by adjusting the values of the transconductances. The functionality of the designed filter was verified by simulations in OrCAD PSpice for selected values of the order and cut-off frequency.
Untethered extended reality (XR) applications increasingly rely on short-range millimeter-wave (mmWave) communication, where directional beam management is essential for maintaining high-throughput, low-latency links. In this paper, we investigate the performance discrepancy between evaluations based on idealized geometric channel models and those using channel state information derived from a shooting-and-bouncing ray (SBR) simulation. We propose a lightweight framework for producing realistic channel state information (CSI) that captures multipath effects such as scattering and diffraction, which are often neglected in simplified geometrical models. We focus on a 60 GHz wearable scenario, where near-body propagation effects may significantly alter link behavior. We outline the end-to-end workflow and present empirical evidence of a non-negligible performance gap between idealized and SBR-based channel assumptions. We observe the gap to be up to 2-4 dB in the average array gain and 15 dB in signal variance. With our dataset and pipeline, we offer an insight into the importance of high-fidelity channel models for modern beam management methods, including compressed sensing and data-driven beam prediction.
The analysis of security logs remains a major challenge for modern Security Information and Event Management (SIEM) systems due to insufficient standardization and diversity of log formats. While Artificial Intelligence (AI) offers great potential for automating monitoring, its use is limited by data sensitivity and a lack of annotated datasets. Augmentation can help generate realistic synthetic logs, providing broader opportunities for AI deployment. This article presents a framework for training language models to generate structured log variants, focusing on key metadata fields while maintaining syntactic consistency and semantic relevance. This framework increases data diversity, reduces the need for manual labeling, and facilitates the integration of AI into Security Operations Centers (SOCs), thereby enhancing operational efficiency. A heterogeneous corpus from 49 sources was cleaned, deduplicated, and transformed into semantically distinct entities. Two augmentation strategies were evaluated: Masked Language Modeling (MLM) and Next Word Prediction (NWP). Eight transformer-based models were finetuned and tested on simulated attack scenarios generated using the Atomic Red Team framework and compared with largescale models to assess accuracy and computational efficiency. The results demonstrate the potential of domain-specific language models for context-aware protocol augmentation, contributing to more efficient and automated security systems.
This work experimentally evaluates the measurement accuracy of picofarad-range capacitors using the Agilent 4294A impedance analyzer. The influence of measurement bandwidth, excitation voltage, connection method (fixtures vs. probe), and averaging techniques was systematically investigated. Results show that while magnitude measurements remain stable, phase accuracy strongly depends on bandwidth and excitation level. Standard fixtures (16047E, 16034G) significantly outperform the 42941A probe, and moving averaging further improves phase stability. Overall, impedance values of capacitors in the tens-of-picofarads range can be measured with higher accuracy than predicted by datasheet error formulas, enabling reliable element characterization.
This work investigates the feasibility of Wi-Fi RTT-based fingerprinting using Android smartphones under different conditions. We analyze key factors affecting positioning accuracy based on k-Nearest Neighbors, including the distance metric, value of k, centroid computation strategy and data representation. Moreover, we also focused on fingerprint aggregation strategies. We show that common distance functions like Euclidean are suboptimal under device diversity. Our proposed differences to the closest AP (DtC) representation significantly reduces positioning errors, especially in cross-device scenarios, by mitigating device-specific measurement offsets. Additionally, aggregating RTT measurements into 1 s non-overlapping windows at both training and testing stages improves robustness and reduces computational load. Experiments across four datasets demonstrate sub-meter accuracy with the proposed approach, outperforming traditional methods. These results highlight the importance of distance metric, data representation, and aggregation in RTT-based positioning systems.
Scanning electron microscopy (SEM) combined with energy dispersive spectroscopy (EDS) is widely used in geosciences for mineral phase classification. However, the lack of large, labeled datasets-especially for unpolished samples common in forensic pedology-limits the direct application of traditional supervised machine learning methods. In this study, we investigate the use of synthetic EDS data, generated using DTSA-II, to train advanced neural network architectures and evaluate their performance on real SEM-EDS measurements. We compare a U-Net baseline with a proposed 3D ResNet and a Transformer model. All models were trained on synthetic data and tested on real measurements of ten selected mineral phases. U-Net achieved accuracy of 63.7%, 3D ResNet reached 92.3%, and the Transformer model achieved the highest accuracy of 97.0%. These findings demonstrate that Transformer architectures can effectively generalize from synthetic to real EDS data, offering a promising way for accurate mineral phase classification in forensic and geological applications without the need for extensive labeled real datasets.
This paper describes the innovation of Rhodes tines and tone bars. It follows on experiments dealing with modelling of new oscillators in the ANSYS program, production of pilot samples and comparison of how well the spectra calculated by simulation correspond to the spectra of mechanically produced pilot samples. Our article is a continuation of this research and it analyses the timbre of the new bars in more depth, classifies them and considers how the new timbre can be processed subsequently in the electrical circuit or digitally so that they retain their character and at the same time meet the parameters of an electronic keyboard instrument.
The integration of Terrestrial and Non-Terrestrial Networks (T/NTNs) is expected to be a native capability of the Sixth-Generation (6 G) of mobile networks. Driven by the growing demand for reliable and ubiquitous connectivity, recent Fifth-Generation (5 G) releases have already begun addressing some of the challenges in developing a three-dimensional (3D) architecture combining T/NTN. In this context, this work presents a preliminary experimental evaluation of integrated 5 G T/NTN scenarios, enabled by a laboratory testbed currently under development within the RESTART ITA-NTN project. The testbed allows for the emulation of ground-, aerial-, and satellite-based links interfacing with 5G Radio Units (RUs) through advanced Radio Frequency (RF) channel modeling and emulation. After an overview of the 5 G Next Generation Radio Access Network (NG-RAN) and the specific adaptations required to support NTN scenarios, the main integrated T/NTN architectures are outlined. To assess these architectures and emulate T/NTN scenarios under controlled conditions, the testbed integrates commercial Software Defined Radios (SDRs) platforms with the Keysight F8820A PROPSIM FS16 channel emulator to replicate several customisable channel conditions, using the OpenAirInterface (OAI) stack for full 5G NG-RAN and core functionality. Preliminary evaluations of the considered T/NTN scenarios are carried out, highlighting key operational characteristics and selected performance metrics, offering insights into challenges and opportunities of extending 5G capabilities to NTN platforms.
Traditional handwriting analysis for neurological assessment captures motor output but largely misses the guiding cognitive processes like visuospatial planning and attention. This study introduces a multimodal approach, combining online handwriting kinematics with concurrent eye-tracking data from 48 older adults performing the Pentagon Copy Test (PCT). We extracted novel feature sets, including Hand-Eye Coupling (HEC) and Fractional Derivative (FD) biomarkers, and used an XGBoost classifier with Recursive Feature Elimination (RFE) to predict a binarized PCT performance score. Our final model, integrating all features, achieved a Balanced Accuracy (BACC) of 90%, significantly outperforming a model trained on baseline features alone (79% BACC). The findings demonstrate that integrating eye-tracking data with advanced handwriting analysis provides a powerful and holistic tool for objectively assessing cognitive-motor performance, highlighting its potential as a sensitive digital biomarker.
The increasing prevalence of deepfake videos underscores the need for effective and reliable detection methods. In this study, we propose a hybrid deepfake detection framework that integrates a static image forgery detector with a recurrent neural network (RNN) to exploit both spatial and temporal features. Specifically, we utilize an existing frame-level detector that identifies common forgery artifacts within individual frames. This is followed by a Long Short-Term Memory (LSTM) network that models temporal dependencies across frames, enabling detection of inconsistencies that are overlooked in frame-by-frame analysis. Experimental results demonstrate that temporal modeling significantly improves accuracy over frame-level baselines. Our contributions are twofold: (i) we provide empirical evidence that deepfake videos exhibit detectable temporal signatures, and (ii) we construct a compact, real-world evaluation set of deepfake videos. Notably, detection performance on this dataset is lower than on standard benchmarks, suggesting a domain gap between commonly used training data and real-world deepfakes.
Cyber deception offers a promising defense framework against reconnaissance in IoT environments, yet existing approaches often lack explicit risk-awareness or resource constraints. In this work, we propose a risk-informed deception framework that integrates quantitative risk scoring with adaptive deployment of deceptive resources. Building on multifactor scoring models such as NIST CRS, our system computes reconnaissance-phase risk scores using source, frequency, asset criticality, and tactic profiles, and dynamically maps these scores to deception levels under budget constraints. Unlike prior dynamic deception or game-theoretic approaches, our framework prioritizes aligning defensive actions with evolving threat levels while minimizing operational overhead. Experimental evaluation on reconnaissance and benign traffic from CICIoT2023 demonstrates that our framework achieves effective containment of early-stage threats with reduced false positives and improved cost-effectiveness compared to static or indiscriminate deception techniques.
Conventional cellular systems, including 5G operating in millimeter wave (mmWave, 30-100 GHz), serve users located in the far-field of the base station antenna. One of the distinctive properties of future beyond 6G cellular terahertz band (0.3-3 THz) systems is the need to support mobile users located in both near- and far-fields. In the near-field, the received signal strength depends not only on the distance between BS and user equipment (UE) but also on the concrete coordinates of the UE and may fluctuate drastically even for displacements on the order of a wavelength. As a result, near-field propagation models are significantly more complex than far-field models, which makes system-level analysis of such prospective systems challenging. The aim of this paper is to provide a comprehensive review of near-field propagation modeling principles and to identify a model that offers a practical compromise between accuracy and analytical tractability. Our results indicate that the best candidate near-field propagation models for system-level research is the spherical model. Even though such a model is provided in terms of sums of components for phased antenna arrays, it allows for simple approximations for a specific set of carrier frequencies. From the accuracy point of view, it preserves the qualitative behavior of the Hertzian model, which is crucial for near-field communications.
Writer identification of handwritten text is a task in forensic document analysis, traditionally relying on visual comparison by experts. However, this process is time-consuming and subjective. Although the amount of crime involving hand-writing has remained constant, the overall volume of handwritten material has decreased. This paper presents an approach based on a Siamese Neural Network (SNN) to writer identification by analyzing a limited information source - just a single character A - from samples of Czech handwriting. The main contributions are: (i) the design of an SNN with a five-block convolutional branch combined with voting strategies incorporating an uncertainty zone; and (ii) a detailed experimental comparison of over 400 architecture and hyperparameter configurations in terms of accuracy, F1-score, and decision efficiency. The best model achieved relatively high accuracy - 96.1% accuracy and a F1-score of 0.932 while abstaining from classification in approximately 57% of ambiguous cases. The trade-off between classification confidence and coverage, the limitations of single-character analysis, and the potential for generalization to openset scenarios and multimodal inputs are discussed. The proposed approach offers an objective and reproducible method suitable for forensic handwriting analysis.
In recent years, there has been increased interest in using advanced technologies such as artificial intelligence, particularly in public safety and rescue operations. This paper focuses on an innovative approach to monitoring and analysing the movement of firefighters during rescue operations using artificial intelligence. In our research, we implemented a system that uses data obtained from sensors placed on the protective suits of firefighters. This data is analysed using deep-learning neural networks after advanced data preprocessing. The goal is to provide a more accurate real-time interpretation of firefighter movement, improving rescue teams’ coordination and increasing firefighters’ safety in their work. This paper presents the results of initial experiments that demonstrate the effectiveness of the proposed system in different rescue operation scenarios. At the end of the paper, we also discuss possible challenges and directions for further research in this area. Our work represents an important step towards integrating artificial intelligence into critical public safety operations. It offers new opportunities for improving rescue operations and protecting lives.
The robustness of biometric signature recognition systems is critical to their adoption in secure authentication frameworks. However, such systems are vulnerable to geometric transformations, such as rotation, that can be maliciously applied to degrade recognition performance. In this work, we explore a realistic adversarial scenario in which a hacker attempts to conceal the identity of a signer by rotating online handwritten signatures prior to biometric recognition. We propose a defense mechanism based on blind watermarking that embeds orientation metadata directly into the signature signal. During recognition, this watermark is extracted to estimate the applied rotation, enabling restoration of the original signal orientation prior to matching. We evaluate the impact of varying rotation angles on biometric performance and demonstrate that our watermark-assisted correction method significantly improves accuracy, restoring it to near-original levels. Specifically, accuracy improved from as low as 2.8% at +/- 90. to over 99% after applying the correction. Our results highlight the importance of signal-level integrity in biometric systems and show that watermarking can be a viable strategy to counteract geometric spoofing attacks.
This study explores the design of fractional-order capacitors (FOCs) implemented using native n-channel MOS devices in TSMC 65 nm technology. Configurations consisting of four and five interconnected MOS segments are investigated, both considered in two variants, so-called shunted and non-shunted gates. A previously designed genetic algorithm-based program is applied to identify the best structures that approximate constant-phase admittance responses providing input admittance phase within a range of 5 to 85 degrees. The search for suitable solutions is conducted across four distinct frequency bands, enabling a detailed evaluation of the influence of both the topology type and frequency range for each phase value. The results provide insight into the capabilities and limitations of MOS-based Fractional-Order Element (FOE) approximants, offering guidance for their use in analog signal processing applications where fractional-order behavior is desired.
