Large Language Models (LLMs) increasingly serve as autonomous reasoning agents in decision support, scientific problem-solving, and multi-agent coordination systems. However, deploying LLM agents in consequential applications requires assurance that their reasoning remains stable under semantically equivalent input variations, a property we term semantic invariance.Standard benchmark evaluations, which assess accuracy on fixed, canonical problem formulations, fail to capture this critical reliability dimension. To address this shortcoming, in this paper we present a metamorphic testing framework for systematically assessing the robustness of LLM reasoning agents, applying eight semantic-preserving transformations (identity, paraphrase, fact reordering, expansion, contraction, academic context, business context, and contrastive formulation) across seven foundation models spanning four distinct architectural families: Hermes (70B, 405B), Qwen3 (30B-A3B, 235B-A22B), DeepSeek-R1, and gpt-oss (20B, 120B). Our evaluation encompasses 19 multi-step reasoning problems across eight scientific domains. The results reveal that model scale does not predict robustness: the smaller Qwen3-30B-A3B achieves the highest stability (79.6

The rapid expansion of the Internet of Things (IoT) demands effective communication protocols that accommodate mobile and static end devices (EDs). Long range wide area network (LoRaWAN), a pioneering low-power wide-area network (LPWAN) technology, uses adaptive data rate (ADR) approaches to optimize resource allocation, particularly for static EDs. However, traditional ADR approaches are ineffective in mobile contexts as they struggle to adapt to changing network conditions, resulting in significant packet loss and higher retransmission rates. Although innovative technologies, such as the blind ADR (BADR), have been devised to improve the performance of mobile EDs, they still fall short of dealing with the unpredictable nature of mobile EDs. To address these challenges, this article presents a novel HybridQ-ADR mechanism suitable for static and mobile EDs. This approach addresses the constraints of BADR and related methods in mobile scenarios. In particular, it provides a more efficient solution to reduce packet loss and collisions in dense and dynamic long range (LoRa)-based IoT networks. This is achieved by allocating spreading factors (SFs) using signal orthogonality to minimize interference and provide reliable communication. Furthermore, the proposed HybridQ-ADR mechanism provides a new clustering technique based on estimated path loss. Specifically, it divides EDs into clusters and assigns different channels to each cluster, improving SF allocation and data transmission speeds. The proposed HybridQ-ADR mechanism includes a mobility-aware, Doppler-constrained SF allocation strategy, limiting each ED's maximum SF based on its Doppler/mobility load to maintain reliable performance at high speeds. Performance evaluations using simulations and testbed implementations show that HybridQ-ADR improves latency, packet success rate, power consumption, and throughput for both static and mobile EDs.

The Internet of Drones (IoD) is a dynamic network architecture in which multiple drones, equipped with communication, sensing, and computation capabilities, are interconnected through Internet of Things (IoT) technologies to perform coordinated tasks autonomously. This infrastructure enables seamless real-time data exchange and collaborative operations across diverse applications, ranging from surveillance to delivery services, while ensuring adaptability, scalability, and security in dynamic aerial environments. However, the IoD introduces new security challenges, as drones are highly vulnerable to various cyberthreats and cyberattacks. Existing Intrusion Detection Systems (IDS) for IoD face several limitations, including high false positive rates, resource constraints of drones, limited adaptability to evolving attack patterns, and a lack of standardized datasets for benchmarking, despite ongoing research efforts. Moreover, there is a lack of a comprehensive study that systematically consolidates existing research. In this paper, we present a systematic literature review to examine the current research area of intrusion detection systems for IoD, focusing on the effectiveness of implemented machine learning models, employed datasets, existing challenges and limitations, as well as emerging trends and future research directions. This review follows PRISMA guidelines, with peer-reviewed journal articles and conference papers selected as the inclusion criteria. Publications relevant to the topic are sourced from a range of databases, including Scopus, IEEE Xplore, ScienceDirect, SpringerLink, ACM Digital Library, and MDPI, covering a 10-year period from 2014 to 2024. From an initial pool of 1,909 records, 62 relevant reports are selected to address the identified research questions. The selected studies are categorized according to publication year, venue, journal, drone domain, IDS type, utilized algorithms, datasets, attack classifications, and software environments. Additionally, a comparative analysis across various factors is presented.

LoRa has proven to be an ideal solution for Internet of Things networks and applications that require long-distance communications, such as those related to smart cities or precision agriculture. Its low cost combined with the wide availability of LoRa-compatible devices make it possible to easily deploy a large number of sensors capable of collecting and transmitting key information for new services and applications. However, the process of adding new devices into a Long Range Wide Area Network (LoRaWAN) network represents a significant challenge on a large scale, as each device must be individually configured and manually registered to join the network. This manual approach is costly and impractical when it comes to deploying a very large number of devices. To address this problem, this paper proposes two deployment strategies (semi-automatic and automatic) to simplify and streamline the process of activating and registering LoRaWAN devices. These strategies facilitate the deployment of large-scale devices in smart cities, and their adoption can significantly enhance the deployment of LoRaWAN devices. Experimental results clearly demonstrate the benefits of our solution. Specifically, for 500 devices, the semi-automatic deployment is 3.75 times more efficient, and the automatic deployment is an impressive 394.87 times faster than the manual deployment.

Evapotranspiration (ET0)-the sum of evaporation and plant transpiration-is a key variable for optimizing water use in precision agriculture. With increasing challenges due to climate change and water scarcity, accurate ET0 forecasting is essential for designing efficient irrigation systems that enhance productivity while conserving resources. This study evaluates advanced time-series models for ET0 forecasting-Nixtla TimeGPT1, Long Short-Term Memory Networks (LSTM), and Kolmogorov-Arnold Networks (KAN)-using IoT data from Campo de Cartagena (Murcia, Spain). Results show that KAN achieves superior performance for multi-step forecasting (MSE: 0.045), while Nixtla Linear excels in one-step predictions (MSE: 0.009). These findings provide practical insights into model selection for adaptive irrigation strategies under diverse climatic conditions.

The rise of AI has positioned edge computing as a pivotal domain for deploying machine learning technologies, fostering agile processing, and enhancing network robustness and decision‐making capabilities. This paper addresses the underexplored aspects of DDoS and phishing attacks, and precise decision‐making at network edge devices within blockchain‐based frameworks. The contribution lies in proposing an incentive‐based security mechanism to divert intruders from genuine routes. Legitimate devices conducting accurate decision‐making are rewarded, enticing their participation in identifying false devices. A honeypot intrusion detection system attracts false devices, and real‐time trust computation monitors communication devices. This approach is analyzed under security threats and network delays, demonstrating its efficacy compared to existing methods in safeguarding edge computing environments.

Research and development on task offloading over the Internet of Drones (IoD) has expanded rapidly in the last few years. Task offloading in a fog IoD environment is very challenging due to the high dynamics of the IoD topology, which cause intermittent connections, as well as the stringent requirements of task offloading, such as reduced delay. To overcome these challenges, in this paper, we propose a task-offloading optimization strategy using a heuristic genetic algorithm (GA) with hybrid fog computing technology for the Internet of Drones, named GA Hybrid-Fog. The proposed solution employs a GA for task offloading from edge Unmanned Aerial Vehicles (UAVs) to both fog base stations (FBSs) and fog UAVs (FUAVs) in order to optimize offloading delays (transmission and fog computing delays) and guarantee higher storage and processing capacity. Experimental results show that GA Hybrid-Fog achieves greater improvements in task-offloading delays compared to other IoD technologies (GA BS-Fog, GA UAV-Fog, and GA UAV-Edge).

The increasing use of UAV swarms for collaborative autonomous missions presents significant challenges in coordination, safety, and scalability, especially during dynamic formation reconfigurations. This study introduces the Magnetic Swarm Reconfiguration (MSR) protocol, a fully distributed navigation method that enables UAV swarms to transition smoothly and safely between geometric formations. MSR achieves this by combining two main components: first, it employs the Hungarian algorithm to compute an optimal assignment of UAVs to target positions within the new formation, thereby minimizing trajectory overlap and interference; second, it utilizes virtual magnetic attraction and repulsion forces for real-time navigation, drawing each UAV toward its assigned destination while dynamically repelling nearby agents to avoid collisions. To evaluate the performance of the MSR protocol, six representative formation transitions were simulated across swarm sizes of up to 100 UAVs. Results show that MSR reduces reconfiguration time significantly compared to existing methods, maintains strict safety standards by achieving minimal to zero collisions, and supports fully decentralized and simultaneous maneuvering. The scalability and robustness of the MSR protocol make it suitable for complex, large-scale swarm operations requiring rapid and reliable formation changes.

This work tackles the problem of target navigation in Global Navigation Satellite Systems (GNSS)-denied scenarios by adapting two Deep Learning (DL)-based approaches: the Temporal Fusion Transformer (TFT) and the Neural Hierarchical Interpolation for Time Series (NHITS). These methods are trained on custom created datasets from which the methods learn, after which they are able to make their own predictions. Obtained numerical results via simulations and real testbed reveal that, on the one hand, the proposed methods improve navigation accuracy and are less vulnerable to noise when compared to existing Machine Learning (ML) approaches. On the other hand, the results also exhibit a reduction in training time. The superior performance of the proposed solutions is primarily attributable to the enhanced network architecture they are based on, which facilitates more efficient learning, which consequently leads to more accurate predictions, when compared to Long-Short Term Memory (LSTM) approach. In this way, the proposed methods allow navigation in GNSS-denied scenarios without relying on expensive hardware (e.g., LiDARs or cameras).

Data exchange in the Internet of Vehicles (IoV) is challenging due to high interference, hidden terminal problems, and frequent transmission collisions. To overcome these issues and enhance overall network performance, we propose a Space Division and theory of Graph Coloration-based protocol for Multi-Channel Multi-Transceiver operation of the MAC layer (SDGC-MCMT) in the IoV. The protocol dynamically assigns service channels to vehicles based on their positions and schedules the transmission of packets across multiple transceivers to enhance efficiency and enable synchronous communication. Furthermore, it uses an adaptive channel busyness-based scheme when the number of available transceivers is lower than the number of service channels to improve data delivery. Simulation experiments, targeting packet delivery ratio, multi-hop channel access latency, and collision probability reveal the effectiveness of SDGC-MCMT in improving network performance compared to the IEEE 1609.4 protocol.

With the rapid deployment of 5G technology, Cellular Vehicle-to-Everything (C-V2X) communication is becoming increasingly critical for enabling real-time data exchange in connected and autonomous vehicles. This demands ultra-low latency and high reliability, especially in high-mobility environments. In this study, we evaluate the performance of two prominent IoT application layer protocols—MQTT and ZeroMQ—for V2X communications over 5G networks. A comprehensive simulation environment was developed using OMNeT++ integrated with INET for networking, Veins for vehicle mobility, and Simu5G to emulate the 5G infrastructure. Custom modules for both protocols were implemented to assess their behavior under varying data loads. Performance metrics, particularly end-to-end latency, were analyzed across diverse traffic scenarios. Results demonstrate that ZeroMQ, due to its lightweight and brokerless architecture, consistently outperforms MQTT in terms of latency, making it a strong candidate for latency-sensitive V2X applications.

While traditional ensemble methods have dominated tabular intrusion detection systems (IDSs), recent advances in foundation models present new opportunities for enhanced cybersecurity applications. This paper presents a comprehensive multi-modal evaluation of foundation models—specifically TabPFN (Tabular Prior-Data Fitted Network), TabICL (Tabular In-Context Learning), and large language models—against traditional machine learning approaches across three cybersecurity datasets: CIC-IDS2017, N-BaIoT, and CIC-UNSW. Our rigorous experimental framework addresses critical methodological challenges through model-appropriate evaluation protocols and comprehensive assessment across multiple data variants. Results demonstrate that foundation models achieve superior and more consistent performance compared with traditional approaches, with TabPFN and TabICL establishing new state-of-the-art results across all datasets. Most significantly, these models uniquely achieve non-zero recall across all classes, including rare threats like Heartbleed and Infiltration, while traditional ensemble methods—despite achieving >99% overall accuracy—completely fail on several minority classes. TabICL demonstrates particularly strong performance on CIC-IDS2017 (99.59% accuracy), while TabPFN maintains consistent performance across all datasets, suggesting robust generalization capabilities. Both foundation models achieve these results using only fractions of the available training data and requiring no hyperparameter tuning, representing a paradigm shift toward training-light, hyperparameter-free adaptive IDS architectures, where TabPFN requires no task-specific fitting and TabICL leverages efficient in-context adaptation without retraining. Cross-dataset validation reveals that foundation models maintain performance advantages across diverse threat landscapes, while traditional methods exhibit significant dataset-specific variations. These findings challenge the cybersecurity community’s reliance on tree-based ensembles and demonstrate that foundation models offer superior capabilities for next-generation intrusion detection systems in IoT environments.

Recently, task offloading in the Internet of Drones (IoD) is considered one of the most important challenges because of the high transmission delay due to the high mobility and limited capacity of drones. This particularity makes it difficult to apply the conventional task offloading technologies, such as cloud computing and edge computing, in IoD environments. To address these limits, and to ensure a low task offloading delay, in this paper we propose PSO BS-Fog, a task offloading optimization that combines a particle swarm optimization (PSO) heuristic with fog computing technology for the IoD. The proposed solution applies the PSO for task offloading from unmanned aerial vehicles (UAVs) to fog base stations (FBSs) in order to optimize the offloading delay (transmission delay and fog computing delay) and to guarantee higher storage and processing capacity. The performance of PSO BS-Fog was evaluated through simulations conducted in the MATLAB environment and compared against PSO UAV-Fog and PSO UAV-Edge IoD technologies. Experimental results demonstrate that PSO BS-Fog reduces task offloading delay by up to 88% compared to PSO UAV-Fog and by up to 97% compared to PSO UAV-Edge.

Particle Swarm Optimization (PSO) has been widely employed to optimize the deployment of Unmanned Aerial Vehicles (UAVs) in various scenarios, particularly because of its efficiency in handling both single and multi-objective optimization problems. In this paper, a framework for optimizing the deployment of edge-enabled UAVs using Pareto-PSO is proposed for data collection scenarios in which UAVs operate autonomously and execute onboard distributed multi-objective PSO to maximize the total non-overlapping coverage area while minimizing latency and energy consumption. Performance evaluation is conducted using key indicators, including convergence time, throughput, and total non-overlapping coverage area across bandwidth and swarm-size sweeps. Simulation results demonstrate that the Pareto-PSO consistently attains the highest throughput and the largest coverage envelope, while exhibiting moderate and scalable convergence times. These results highlight the advantage of treating the objectives as a vector-valued objective in Pareto-PSO for real-time, scalable, and energy-aware edge-UAV deployment in dynamic Internet of Flying Things environments.

Traffic congestion and carbon emissions remain pressing challenges in urban mobility. This study explores the integration of UAV (drone)-based monitoring systems and IoT sensors, modeled as induction loops, with Large Language Models (LLMs) to optimize traffic flow. Using the SUMO simulator, we conducted experiments in three urban scenarios: Pacific Beach and Coronado in San Diego, and Argüelles in Madrid. A Gemini-2.0-Flash experimental LLM was interfaced with the simulation to dynamically adjust vehicle speeds based on real-time traffic conditions. Comparative results indicate that the AI-assisted approach significantly reduces congestion and CO2 emissions compared to a baseline simulation without AI intervention. This research highlights the potential of UAV-enhanced IoT frameworks for adaptive, scalable traffic management, aligning with the future of drone-assisted urban mobility solutions.

Autonomous driving technology has achieved remarkable advancements, offering substantial potential to revolutionize traffic safety and smart mobility. However, when faced with rare scenarios (weather, accident scenes, and lighting), autonomous driving systems can still only play a limited role due to insufficient learning in these rare situations. To address this challenge, we propose a novel approach that leverages low-rank adaptation (LoRA) and Mixture of Experts (MoEs) technologies to enhance the performance of pretrained autonomous driving models in handling rare situations. Specifically, we first use LoRA to fine tune the pretrained model of autonomous driving to focus on capturing knowledge related to rare scenarios and enhance the model's ability to handle rare situations. Furthermore, we introduce MoEs and propose local, global, and hybrid adaptive solutions to overcome LoRA's fixed intrinsic rank limitation. These approaches enable adaptive adjustment of LoRA's rank, and improve the model's performance from both local and global perspectives. Finally, we design detailed algorithms for different adaptation schemes. Extensive experiments demonstrate that our proposed solutions not only effectively improve the performance of the autonomous driving perception model in rare scenarios but also maintain lower inference latency compared to baseline methods.

The steady rise in the use of unmanned aerial vehicles (UAVs) is leading to the development of an ever-growing number of applications. In urban settings, efforts like the U-Space initiative in Europe are striving to standardize and regulate the operations of UAVs. To support these applications and further UAV research, it is essential to thoroughly understand UAV communication, both among and between UAVs. Nonetheless, we have identified a lack of studies on communication models, especially in urban areas where obstacles like tall buildings can disrupt communication. This study offers a comprehensive review of current measurement campaigns on channel models for aerial communication. In addition, we conducted experiments on (i) the separation distance between two UAVs, (ii) Multi-UAV communication and (iii) Multi-UAV to ground communication using three different city profiles in Spain (Valencia, Barcelona, and Madrid). To accomplish this, we utilized an advanced co-simulation framework that accurately models both UAV mobility (Ardusim) and communication (OMNeT++). Our results regarding UAV-to-UAV communication in a city environment indicate that: (i) the communication range, in our specific experiments, is limited to around 400 meters. Afterward, the Packet Delivery Ratio (PDR) declines significantly. (ii) Different communication models yield similar results. (iii) UAV-to-UAV communication becomes feasible at higher altitudes (e.g., 120 m), particularly in the presence of tall buildings. With respect to the Multi-UAV to ground communications, we can conclude that again, the altitude of the UAVs is paramount. Furthermore, increasing the number of UAVs providing service to the ground does increase the PDR, but only ever so slightly.

The integration of digital twins (DTs) with intelligent traffic systems (ITSs) holds strong potential for improving real-time management in smart cities. However, securing digital twins remains a significant challenge due to the dynamic and adversarial nature of cyber–physical environments. In this work, we propose TwinFedPot, an innovative digital twin-based security architecture that combines honeypot-driven data collection with Zero-Shot Learning (ZSL) for robust and adaptive cyber threat detection without requiring prior sampling. The framework leverages Inverse Federated Distillation (IFD) to train the DT server, where edge-deployed honeypots generate semantic predictions of anomalous behavior and upload soft logits instead of raw data. Unlike conventional federated approaches, TwinFedPot reverses the typical knowledge flow by distilling collective intelligence from the honeypots into a central teacher model hosted on the DT. This inversion allows the system to learn generalized attack patterns using only limited data, while preserving privacy and enhancing robustness. Experimental results demonstrate significant improvements in accuracy and F1-score, establishing TwinFedPot as a scalable and effective defense solution for smart traffic infrastructures.

Nowadays, Unnamed Aerial Vehicles (UAVs) are used in many different fields, ranging from agriculture and entertainment to parcel delivery, among others. For several of these tasks, the UAVs are programmed to follow a specific path, as defined in their flight missions. In addition, as more sophisticated solutions begin to be adopted, new protocols should be developed to handle possible collisions among UAVs, as well as to create UAV swarms. However, directly testing new protocols on UAVs can be hazardous and time consuming. Therefore, many investigators first perform simulations. Although different UAV simulators exists, not all of them offer a 3D rendering of the UAVs in the target flight environment. Hence, in this work, we present a real-time 3D visualization interface that can be easily coupled to any simulator. In this way, we offer developers a powerful way to validate their solutions and make in-depth analysis.

The utilization of sentiment analysis as a method for predicting stock market trends has gained significant attention recently, especially during economic crises. This research aims to assess the predictive accuracy of sentiment analysis in the stock market by constructing a reinforced model that integrates both sentiment and technical analysis. While prior studies have concentrated on social media sentiment for stock price prediction, this research introduces an enhanced model that combines sentiment analysis with technical indicators to improve the precision of stock market prediction. The study creates and evaluates predictive models for stock prices and trends using a substantial dataset of tweets from twenty prominent companies. Finally the re-enforced model has been developed and tested on the stock prices of: Apple, General Electric, Ford Motors and Amazon. The deliberate selection of these companies, each representing distinct industry sectors, serves a dual purpose. It not only facilitates a practical evaluation of our model across diverse market conditions but also ensures computational feasibility, allowing for a focused and detailed analysis of the model’s predictive accuracy and reliability in various economic landscapes. The study’s outcomes offer valuable insights into the effectiveness of the reinforced model, which combines sentiment and technical analysis to predict stock market movements, providing a more comprehensive approach to understanding market sentiment’s influence on stock prices. Furthermore, these findings contribute to the existing knowledge on stock market prediction techniques and emphasize the importance of considering multiple factors in decision-making.