This study demonstrates how fake news can be detected and analyzed by diffusion trends, mainly by treating it either as a classification problem using ML (Machine Learning) or by examining how it spreads on $\operatorname{Twitter}(\mathrm{X})$ temporal network. Classification involves supervised learning, where labeled training data is used to build predictive models. In this comparative analysis, we evaluate the effectiveness of four wellknown ML algorithms-K-Nearest Neighbors (K-NN), Naive Bayes, Random Forest and BigBird-in detecting fake news. The goal is to compare their accuracy in classifying textual tweet text into two categories: real and fake news. The WELFake dataset, which contains a variety of Twitter-based text, is used for experimentation. After identifying BigBird as the most effective model for our dataset, we apply it to a new tweet text related to COVID-19 news to determine its classification. The models were then tested on a tweet related to COVID-19, and each one identified the content as likely being fake news. In the final phase of the study, we applied two epidemiological models: SIR (Susceptible, Infected, Recovered) and SEIZ (Susceptible, Exposed, Infected, Skeptics) to compare the spread dynamics of contrasting viewpoints. It was observed that the SEIZ model fits the data more accurately and with less error than the SIR model. Furthermore, the $\mathbf{2 4}$-hour prediction analysis of infected nodes using the SEIZ model demonstrates that the diffusion of fake news follows a typical viral spread pattern, characterized by a rapid initial increase in infected nodes, followed by a plateau as the network reaches saturation.
Energy efficiency is becoming increasingly important for sustainability. In a cloud-native environment, orchestration platforms such as Kubernetes can significantly impact the overall energy performance of applications as they manage their lifecycle, including scheduling. Kubernetes, specifically as a de facto cloud platform, has a default scheduler that does not account for energy consumption in its decision-making process, highlighting the need to explore mechanisms that can improve the energy performance of deployed applications. In this paper, we present an energy-aware scheduling mechanism within Kubernetes to optimize the energy usage of cloud-native applications while meeting other scheduling requirements and constraints. This approach extends the scheduling logic in Kubernetes through plugins that integrate energy metrics and employ Machine Learning-based power estimation for energy-optimized decision-making. The evaluation of the proposed approach shows up to 12 % energy savings with minimal performance degradation. When complemented by other consolidation techniques, it could potentially lead to substantial savings ranging from 21 % to 65 %, depending on the system load in the scenario analyzed. Importantly, this method does not introduce significant overhead.
Research in the Network-based Intrusion Detection Systems (NIDS) field is a major focus for academics and the emergence of reliable and applicable detection models quickly became imperative for the integrity of critical infrastructures. Thus, Machine Learning (ML) models-specifically Deep Learning (DL) models- have been developed to detect network attacks and multiple training datasets are available, which allows to reproduce experiments. Nevertheless, the problem of intrusion detection is challenging since generating a representative set of training data is difficult due to the large variety of applications, and the fact that network infrastructures evolve over time. Also, a key concern often ignored is the different performances of models between experiments on known and often curated datasets, and in reality when deployed. In this paper, we propose a methodology to highlight the shortcomings of Anomaly-based NIDS, showcasing significant performance drops using realistic scenarios of evolution in the datasets. We implement a Denoising AutoEncoder (DAE) as a NIDS model. Then, we present a method to test and interpret the NIDS results based on analyzing the DAE's Mean Squared Error (MSE) to provide insights into the model's flow classifications. We advocate for considering the natural evolution of infrastructures when designing NIDS and highlight further research to address this challenge.
Optical turbulence is a phenomenon caused by atmospheric temperature variations and wind shear, leading to changes in air refractive index. This affects light propagation through the atmosphere, resulting in wavefront distortion, scintillation, and image degradation, posing a significant challenge to the reliability of free-space optical (FSO) communication systems. This paper presents a novel ensemble machine learning framework for predicting the refractive index structure parameter ($C_{n}^{2}$) in operational FSO networks. By integrating multiple machine learning algorithms (Random Forest (RF), Neural Network (NN), and Support Vector Machine (SVM), the proposed approach is designed to perform effectively under dynamic environmental conditions. The ensemble model achieves superior accuracy (97.82%) as shown in Fig. 3 and operational robustness compared to conventional methods. Through comprehensive uncertainty quantification, we identify key atmospheric variables influencing prediction confidence, with wind speed as the strongest predictor ($r=0.770$), followed by temperature gradient ($r=0.306$), as seen in Fig. 5. Our method reduces FSO link outages by approximately 20 % relative to traditional techniques through condition-specific uncertainty modeling that enables preemptive adjustments before link degradation. This framework provides a practical strategy for real-time optimization of optical systems across varying atmospheric conditions, particularly benefiting deployments in high-wind scenarios and complex terrain environments.
Large-scale learning management systems (LMS) generate rich interaction logs that offer valuable insights into study behavior over time. This paper analyses more than 11.3 million activity records collected from $\mathbf{5, 3 0 0}$ students enrolled in 1,166 Moodle courses during four consecutive semesters between winter $2022 / 23$ and summer $2023 / 24$. We replicate and extend our 2023 baseline study by (i) contrasting activity volumes, (ii) comparing hourly access patterns using two-sample KolmogorovSmirnov tests, and (iii) testing grade and time-to-deadline distributions with Mann-Whitney and Wilcoxon statistics. While the temporal shape of weekly access curves remained remarkably stable across cohorts ($p > 0.16$ for all KS tests), the absolute number of interactions grew by 60 %. At the same time, students in the most recent summer term submitted assignments significantly earlier (Wilcoxon $\mathrm{W}=5.76^{*} 10^{\wedge} 10, \mathrm{p}<0.001$) and achieved higher final marks ($\mathrm{W}=5.60^{*} 10^{\wedge} 10, \mathrm{p}<0.001$). The results indicate that behavioral analytics must go beyond simple traffic curves and combine volume, timing, and performance metrics to capture meaningful longitudinal change.
The article describes the issue of assessing the impact of packet loss in the network on the degradation of video quality received by users. The case of live streaming service was considered here, where potential error correction by retransmitting lost packets is not implemented. The authors described the results of research conducted in a test laboratory environment under full control of network traffic. The prepared set of reference video samples in three resolutions was encoded using H.264 and H.265 codecs. The video samples were transmitted over the network with different levels of packet loss and recorded at the receiving end. The degraded video was assessed using objective and subjective methods. The obtained results were used to build a machine learning model that can be used to predict the video quality experienced by users based on the designated objective metrics. The authors thus indicate that the machine learning model allows for obtaining better video quality prediction results than the previous modeling based on single objective metrics.
As sixth-generation (6G) mobile networks are being developed to address the limitations of 5 G, such as network congestion, high latency, limited throughput, and poor scalability, evaluating key performance indicators (KPIs) becomes increasingly important. Among these, user throughput is a critical metric that directly affects quality of service for data-intensive applications like real-time streaming, augmented reality, and mission-critical Internet of Things (IoT). Artificial intelligence (AI) is expected to play a central role in optimizing throughput by enabling intelligent resource management and real-time adaptation to network conditions. In this paper, we analyze and simulate user throughput in 6G networks through various analytical modeling approaches, with a particular emphasis on its relationship with signal-to-noise ratio (SNR). By evaluating throughput behavior across multiple modeling frameworks, we aim to offer a more comprehensive understanding of throughput dynamics in 6G environments and provide actionable insights that inform the design of future high-performance mobile networks.
The first part of the two-part paper introduces a new setup for optimization of higher-order $(\text{HO})$ controllers using a base of controller families derived by the multiple real dominant pole (MRDP) method for ultralocal process models with increasing degree of the pure integrator. It is illustrated by the design of a family of HO PID (proportional-integral-derivative) controllers derived for integrator plus dead-time (IPDT) models applied to control an unstable process. The specification of the optimal family term is accomplished by comparing all the available options by the performance portrait method, which shows the possibility of a huge increase in loop performance compared to earlier published papers. The reduction of the IAE value for disturbance responses below 4 % of the previously reported values obviously introduces a completely new generation of controller design. The second part of the paper then continues with the development of another family of HO-PIDs using the double-time-delayed integrator model.
The Internet of Things (IoT), particularly in constrained environments like Wireless Sensor Networks (WSNs), demands scalable and secure networking. Content-Centric Networking (CCN) addresses key IoT challenges through content-based communication, enabling caching, request aggregation, and resilience. However, CCN is vulnerable to Interest Flooding Attacks (IFA), which overload the Pending Interest Table (PIT) with excessive Interest packets. IfNot, a lightweight mitigation approach, adjusts forwarding based on PIT timeouts but suffers from high false positives and limited resilience to adaptive attackers. To improve this, we propose IfNot-R, which penalizes malicious behavior using historical data, and IfNot-FPGuard, which reduces false positives via longterm Interest satisfaction tracking. These enhancements strengthen IFA defense in dynamic IoT contexts.
For cooperative driving both environmental perception and communication with other vehicles and the infrastructure are essential to avoid safety critical situations. The 5G Sidelink, as part of the Cellular Vehicle to Everything framework, enables both communication and, since 3GPP Release 18, positioning for distance estimation between vehicles. Due to the expected bandwidth limitations of 5G Sidelink in the Intelligent Transportation System band, combined with signal distortion caused by multi-path superposition, achieving submeter positioning accuracy is a significant challenge. Polarization diversity is investigated as a potential approach to mitigate errors resulting from signal distortion and subsequent peak shifts due to multi-path superposition. This work investigates the potential for reduced peak shift errors through the use of dual-polarized receive antennas. The received signals are processed using a polarization scan algorithm and compared to the accuracy achieved with single-antenna systems. The analysis is conducted using the Quasi Deterministic Radio Channel Generator framework with statistical Vehicle-to-Everything channel models. Simulation results show that polarization scan-based postprocessing reduces peak shifts and improves robustness in scenarios with polarization-mismatched antenna configurations.
Affordable positioning technologies make it possible for more stakeholders to access and utilize advanced navigation and tracking systems, crucial for economic development; in particular, in African region there is a significant potential of emerging positioning-based services to support the developments of essential infrastructure such as transportation networks, water access, or agricultural systems. Thus, the study of novel low-cost positioning solutions is a vital component of economic growth and progress. Recently, Low Earth Orbit - Positioning, Navigation, and Timing (LEO-PNT) solutions have gained significant interest in the research community, with several studies dedicated to LEO-constellation optimization for standalone LEO-PNT systems. Until now, most of the proposed LEO-PNT constellations, with the exception of Iridium Next, have typically hundreds of satellites in their constellation, thus requiring significant costs in deployment and maintenance of such constellations. In this paper, we propose three new LEO-PNT constellations with 90 to 102 satellites in orbit, which are optimized in the African region according to a multi-objective optimization problem, with PNT targets in mind, and which can reach positioning performances - in terms of coverage, carrier-to-noise ratios, and dilution of precision - comparable with those of much larger LEO constellations.
This study presents a deep learning approach to predict attention drops in human attention by analyzing specific features from electroencephalogram (EEG) signals, with an emphasis on frequency bands that reflect higher attention states. The study develops and evaluates a Bidirectional Long-Short-Term Memory (BiLSTM) model for predicting attention drop in the next 5-20 seconds based on the prior 5-20 seconds of data. The model is trained on EEG recordings collected during various fire detection surveillance tasks, and achieves promising accuracy and recall using a minimal set of EEG features. Results demonstrate its potential for real-time cognitive monitoring in high-risk environments, especially in domains where early warning of attention drop can prevent serious errors and enhance human performance.
This study presents a digital twin-enabled framework integrated with a binary classification Machine Learning (ML) model for forecasting failures in Erbium-Doped Fiber Amplifiers (EDFAs). The framework utilizes GNPy, an opensource optical network planning tool, to construct a digital twin that serves as a virtual replica of the physical EDFA system. This digital twin facilitates the estimation of Quality of Transmission (QoT), along with the collection and analysis of key operational parameters. A binary classification model, based on Long Short-Term Memory (LSTM) networks, is trained on the data generated by the digital twin to predict potential EDFA failures, achieving a high prediction accuracy of 98 %. This predictive capability enables early fault detection and proactive maintenance, thereby minimizing unplanned downtime and service disruptions. By incorporating real-time analytics and predictive insights, the proposed approach significantly enhances the reliability, availability, and intelligence of optical network management.
The paper deals with a simple double disk solenoid plasma thruster. Some fundamental principles of rocket thruster operation and snaliytical modeling are given in the paper. An illustrative computational example that may serve as a kind of an opener to the subject regarding future engineering and experimental activities in this area is presented in the paper.
This work presents an innovative Multi-Parameter Artificial Intelligence (AI)-based decision algorithm, designed to optimize handovers in Wi-Fi networks supporting Multi-Link Operation (MLO). The algorithm leverages a combination of Neural Networks (NN), integrating a Long Short-Term Memory (LSTM) NN with a Classification NN. The former predicts key performance indicators, which are then used by the latter to determine the optimal connection between a Wi-Fi Station and available Access Points. Our approach, compared to traditional distance-based handovers, effectively improves connectivity in more complex and realistic curvilinear trajectories, especially in scenarios with interference.
Understanding the actual energy consumption of edge devices, such as Raspberry Pi boards, is becoming increasingly important as these platforms play a central role in real-world IoT applications. Power consumption information is typically sourced from manufacturer datasheets or coarse approximations, which often fail to reflect operational dynamics under realistic workloads. In this paper, we present an open tool for reproducible empirical characterization of edge device power consumption across varying computational workloads. The tool exploits affordable, off-the-shelf hardware for power measurement in combination with system-level metrics and automated workload generation through the stress-ng framework. Power data are collected at high temporal resolution and analyzed using regression models to derive load-dependent energy profiles suitable for integration into system-level energy modeling. Alongside this, we provide a toolchain and a set of Jupyter notebooks to facilitate reproducibility, cross-platform comparison, and future expansion by the community.
This paper deals with a simple statistical analysis of measured values of $\mathbf{5 G}$ New Radio electric field carried out in the vicinity of typical base station antenna systems installed in the Republic of Croatia. The measured 5 G signals operate in the frequency range of 3.5 GHz to 3.8 GHz and consist of three channels, each 100 MHz wide. The Channel Power measurement method, combined with forced data traffic, are used to determine the electric field levels. A total of $\mathbf{3, 0 0 0}$ measurement results have been collected, and a simple statistical analysis has been performed. Furthermore, common issues associated with the Channel Power measurement method are highlighted. Some illustrative results are presented.
Large Language Models (LLMs) are increasingly integrated into educational contexts due to their ability to process and generate text through prompt-based instructions. Their adoption enables personalized learning and automated assessment, but also raises concerns about accuracy, bias, and responsible use. In this work, we present CLUE LMS, a Learning Management System that combines standard functionalities, such as access to teaching materials and tests, with LLM-based features. Specifically, the system includes a conversational agent for topic clarification and automatic generation of self-assessment quizzes, both grounded in the teacher-provided materials. The architecture integrates Retrieval-Augmented Generation (RAG) to ensure that generated content remains consistent with course content. To assess its potential, we conducted a user study with four participants. Based on the preliminary results, CLUE LMS is simple to use and effectively supports learning, with the quiz generation feature being the most appreciated. These findings suggest that LLM-based tools, when properly constrained and aligned with educational material, can enhance learning while mitigating common risks associated with generative AI in academic settings.
The growing demand for adaptive video streaming poses challenges in maintaining high Quality of Experience (QoE) and ensuring fairness among users under heterogeneous network conditions. Traditional Adaptive Bitrate (ABR) algorithms struggle to allocate resources efficiently, often disadvantaging users with limited bandwidth. This paper presents a Reinforcement Learning (RL)-based approach deployed at the edge, designed to jointly optimize QoE and fairness in adaptive streaming. The proposed framework runs on a Docker-based infrastructure and leverages real-world network traces to simulate dynamic congestion patterns. We compare our RL-based strategy against heuristic ABR algorithms using key QoE metrics—average throughput, buffering duration, and quality switching—and assess fairness through Jain's Fairness Index. Results show that our RL model outperforms traditional heuristics, offering improved fairness and QoE, particularly in degraded and fluctuating conditions. These outcomes underline the potential of edge-assisted RL schemes to overcome limitations of current ABR approaches in real-world streaming scenarios.
This study examines evolving digital perception of Albania as a tourist destination using Reddit posts between 2013 and October 2024. Using sentiment analysis via VADER and dynamic topic modeling via BERTopic on a 4,276-post English-language dataset, the paper indicates that over 75 % of end-user posts are positive in sentiment—especially during peak travel months (May to September). Prominent themes addressed included accommodation logistics, public transportation, beach resorts, cultural identity, and general perceptions of Albania. Temporal observations show a surge in international interest and talk about Albania's tourism in 2023-2024, aligning with the post-pandemic recovery in international tourism. These findings confirm the validity of using Reddit as an in-real-time source of tourism information and affirm the place of Albania as a rising, affordable, and culturally rich Mediterranean destination. The study offers policy and tourism marketing insights to policymakers and marketers interested in knowing the priorities of travelers and guiding evidence-based development strategies.
