Large language models (LLMs) have significantly advanced natural language processing (NLP) by breaking through a variety of tasks like summarization, text generation and sentiment analysis. This study addresses aspect-based sentiment analysis (ABSA) for Arabic, a low-resource language. ABSA, which identifies aspect categories and their sentiment polarity, comprises subtasks like aspect category detection (ACD) and aspect category polarity (ACP). Focusing on the aspect of category polarity, we investigate zero-shot, few-shot, and chain-of-thought (CoT) prompting strategies using the Gaza Reviews and HAAD Book Reviews datasets. Our experiments demonstrate that CoT prompting outperforms other methods, achieving F1-scores of 70 % on the Gaza Reviews dataset and 67% on the HAAD Book Reviews dataset.
The rapid expansion of the Internet of Things (IoT) and the growing need for real-time data processing have accelerated the development of decentralized computing paradigms. This paper presents a detailed comparative analysis of edge computing and fog computing architectures, exploring their distinct roles in modern decentralized networks. By evaluating architectural differences and performance metrics, we assess the advantages and limitations of each approach. The proposed methodology combines theoretical analysis with comparative frameworks. Our findings indicate that edge computing excels in ultra-low-latency applications with minimal infrastructure, while fog computing offers greater scalability and efficient resource management for complex distributed systems. We conclude that the choice between edge and fog computing depends on application-specific requirements, with hybrid models emerging as a promising solution for future decentralized networks. This study enhances the understanding of decentralized computing architectures and provides actionable insights for system designers and network architects.
The growing scale and complexity of IoT (Internet of Things) networks demand robust security models that provide both secure identity authentication and reliable trust evaluation. While Self-Sovereign Identity (SSI) enables decentralized and user-centric authentication, ensuring the continuous trustworthiness of participating nodes remains a challenge. This paper presents a unified framework that combines SSI-based decentralized authentication with IOTA-ledger-assisted trust evaluation to provide an integrated security architecture for IoT systems. The proposed architecture guarantees that only authenticated nodes can participate in the network while continuously assessing their behavior and trustworthiness through immutable and tamper-resistant trust scores. Our approach is, to the best of our knowledge, the first to unify SSI-based authentication with IOTA-assisted trust evaluation, offering a holistic IoT security framework.
Cyberbullying is an increasing threat on social media, with serious consequences for mental health, particularly among young people. Despite growing global efforts to address this problem, developing accurate detection systems remains challenging for low-resource languages in both text and audio modalities, such as Arabic and its dialects. This paper presents a binary-labeled dataset for cyberbullying detection in Moroccan Darija. The dataset merges over 4,000 newly collected YouTube comments with the OMCD (Offensive Moroccan Comments Dataset) corpus of 8,024 comments, originally labeled for offensive language. To better reflect the nature of cyberbullying, the entire corpus was reannotated from scratch, clearly distinguishing between bullying and non-bullying content, including subtle forms like sarcasm, shaming, and indirect aggression. Several Arabic transformer models were fine-tuned and evaluated using standard classification metrics. The results show that dialectspecific models, particularly DarijaBERT, achieve the best performance, underlining the importance of context-aware annotation and in-domain pre-training for cyberbullying detection in lowresource settings.
This study introduces a novel centrality metric designed to identify influential actors in social media networks, based on two key parameters: relative degree superiority ($\mathbf{x}$ over y) and the minimization of shared neighbors. Our measure examines direct influence and also takes into account neighborhood non-redundancy. Both of these elements shape network dynamics. Centrality measures used in social media networks are different. They generally only consider direct connections. They often miss part of the network dynamics. We conducted theoretical analyses and empirical validations across various datasets to compare our metric against existing measures, highlighting its unique advantages in understanding network influence. According to our research, this measure improves the accuracy of influencing social media networks contributing to our understanding of the dynamic and structural characteristics of real-world systems.
The rapid expansion of Internet of Things (IoT) networks has made them increasingly vulnerable to diverse cyber threats, necessitating the development of efficient Intrusion Detection Systems (IDS). Traditional models for IDS often face challenges such as data imbalance, scarcity of labeled samples, and suboptimal performance due to manual hyperparameter tuning. To address these issues, we propose a comprehensive IDS framework comprising three key components. First, we mitigate data imbalance using the proximity weighted synthetic oversampling technique, which enhances class distribution, followed by the use of Pointer Network (PtrNet) for classification due to its ability to model variable-length sequential data. Second, to handle the scarcity of labeled data, we introduce an entropy-based active learning strategy on PtrNet, termed Entropy-based Active Learning Pointer Network (EAL-PNet). Finally, we optimize model performance through harris hawk optimization applied to PtrNet, resulting in Hawk-Pointer Attention Network (HPA-Net). Experimental results demonstrate that the proposed models significantly outperform traditional approaches. EAL-PNet achieves a performance improvement of 9.30% in accuracy, 8.14% in F1-score, 8.14% in precision, 9.30% in recall, 3.16% in Receiver Operating Characteristic - Area Under the Curve (ROC-AUC), 13.92% in Matthews Correlation Coefficient (MCC) and Cohen's Kappa, and 45.71% reduction in log loss. Similarly, HPA-Net shows a 10.47% gain in accuracy, 10.47% in F1-score, 9.30% in precision, 10.47% in recall, 2.11% in ROC-AUC, 15.19% in MCC and Cohen's Kappa, and 51.43% decrease in log loss. These findings validate the effectiveness of the proposed framework in enhancing intrusion detection for IoT environments.
This study introduces an innovative Low-Rank Autoencoder (LRAE) for anomaly detection in automotive Controller Area Network (CAN) systems, demonstrating significant advancements over traditional Standard Autoencoders (SAEs). By factorizing weight matrices into compact subspaces (rank $\leq 16$), the LRAE achieves a 91.3 % reduction in parameters (2,010 vs. 23,198) and a 9 times lower memory footprint (0.01 MB vs. 0.09 MB), enabling efficient deployment on resourceconstrained Electronic Control Units. Training dynamics reveal LRAE converges 50 % faster (5 vs. 10 epochs) with a validation loss stabilizing at 0.35, compared to SAE's at 0.05. Evaluated on the SynCAN dataset, LRAE outperforms SAE with a $1,233 \%$ higher recall (0.04 vs. 0.003) and $1,440 \%$ higher F1-score (0.077 vs. 0.005) in continuous attacks, a 148.4 % recall gain (0.226 vs. 0.091) in suppress attacks, and a 48.1 % higher PrecisionRecall Area Under the Curve (PR-AUC) (0.228 vs. 0.154) in plateau scenarios. Despite minor trade-offs in flooding (2.7 % lower PR-AUC, 0.741 vs. 0.761) and playback (22.2 % lower F1score, 0.112 vs. 0.144), LRAE's accelerated convergence and enhanced specificity position it as a pioneering advancement. These results highlight LRAE's potential to revolutionize automotive cybersecurity, blending efficiency with state-of-the-art detection performance.
In this paper, we investigate the secure consensus problem for multiple-input-multiple-output (MIMO) linear multi-agent systems (MASs) under false data injection (FDI) attacks, where an adversary corrupts the output measurements exchanged among agents. To mitigate the effect of such attacks, we design an unknown input observer (UIO) that relies only on relative output information and reconstructs the consensus error while filtering out the injected malicious signals. On the basis of the proposed observer, we design a robust consensus control law. By employing a common Lyapunov function (CLF) together with linear matrix inequality (LMI) techniques, we establish sufficient stability conditions ensuring that consensus among agents is preserved even when false data injection (FDI) attacks occur, provided certain observability and robustness conditions are met.
Federated Learning (FL) allows multiple clients to collaboratively train a machine learning model without directly sharing their raw data. While federated learning provides a degree of data privacy, model updates are still susceptible to various inference attacks. This paper presents FLASS, a lightweight and decentralized federated learning framework that leverages additive secret sharing within a multi-server architecture. Each client encodes its local model update into additive shares and shares them with several non-colluding servers, which execute secure aggregation without gaining knowledge of individual contributions. FLASS does not rely on heavy-weight cryptography or a trusted authority. The security analysis and experiments demonstrate the effectiveness and efficiency of the suggested scheme. The analysis indicates that, while achieving the same accuracy as conventional FL schemes, FLASS ensures robust privacy protection with acceptable computational and communication overhead.
The growing integration of photovoltaic (PV) energy into smart electrical grids presents significant challenges related to intermittency, fault detection, system sizing, and real-time control. This review provides a comprehensive analysis of recent advancements in Artificial Intelligence (AI), particularly Machine Learning (ML) and Deep Learning (DL), for optimizing the operation and management of PV systems within smart grids. The study categorizes and compares AI-based methodologies in energy forecasting, fault detection and diagnosis, intelligent PV system sizing, and optimization of Maximum Power Point Tracking (MPPT) techniques. Through a critical evaluation of recent literature (2020-2025), the review highlights the complementary roles of classical ML models and advanced DL architectures, including Long ShortTerm Memory (LSTM) networks, Transformer models, and hybrid approaches combining physics-based knowledge with datadriven learning. Moreover, heuristic optimization algorithms such as Particle Swarm Optimization (PSO) are discussed for their contribution to enhancing convergence speed and accuracy in MPPT and diagnostic applications. This review not only synthesizes state-of-the-art AI strategies but also identifies key technical gaps related to data availability, computational complexity, and model interpretability. Emerging solutions such as federated learning, edge computing, and Explainable AI (XAI) are discussed as promising pathways toward scalable, trustworthy, and autonomous PV energy systems. The study aims to support researchers and practitioners in designing robust, real-time AI-based frameworks for next-generation renewable energy integration.
Cybersecurity threats are becoming more frequent, complex, and harmful, constantly putting organizational infrastructures at risk, especially when critical services are hosted on Linux-based systems. When an attack occurs, having a solid incident response process is essential not only for analyzing the breach afterward but also for detecting threats early and helping contain them in real time. Incident response plays a key role before, during, and after an attack, from preparation and detection to containment and post-incident analysis. Traditionally, forensic investigations rely on tools like FTK Imager and Autopsy, which create and analyze full disk images. While powerful, these tools are often slow, resource-intensive, and not ideal for live systems, requiring significant CPU, memory, and storage resources. This makes them difficult to use during urgent or ongoing incidents. To overcome these limitations, our research introduces a lightweight and modular forensic toolkit tailored specifically for Linux systems. Unlike traditional approaches, our toolkit is designed to work in scenarios where speed, minimal overhead, and targeted analysis are essential. It directly addresses the challenges of heavy resource usage, delayed execution, and the need for automation in incident response. The toolkit works through two main approaches. First, it sends system logs to a SIEM (Security Information and Event Management) platform, where visual dashboards, queries, and detection rules help quickly identify attack techniques. Second, it includes a custom script with two components: a collector that gathers key Linux system artifacts, and an analyzer that analyzes them to reconstruct a detailed attack timeline. A major advantage of this toolkit is its ability to deliver fast, structured, and automated forensic analysis, significantly reducing response time while preserving investigation quality. It allows analysts, IT personnel, and system administrators to act quickly and confidently even in high-pressure scenarios without relying on bulky traditional tools.
There is an increased adoption of Google Remote Procedure Call (gRPC) as one of the key communication protocols in microservice architecture (MSA) applications. This adoption leads to extending the need for runtime adaptive access control policies through the service mesh to reduce the risks of compromised microservices reaching other microservices. A service mesh, as a layer within the orchestration solution, provides features, such as secure communication, within a deployed MSA application. However, it lacks the mechanisms for adaptive gRPC access control policies to meet the zero-trust principle of always verifying and never trusting. This paper introduces a novel mechanism that extends our previous work on risk-adaptive access control to the use of gRPC through the service mesh. We demonstrate our approach and show how gRPC adaptive access control policies can be applied dynamically at runtime for deployed microservices. We use Kubernetes as the orchestration solution and Istio as the service mesh platform in a model MSA application called Online Boutique hosted on the CloudLab scientific computing platform.
The main objective of this contribution is to provide an in-depth analysis of the vulnerabilities present in the various layers of the SDN environment as well as to propose a novel solution to detect distributed denial-of-service (DDoS) attacks, as they pose serious threat to the stability and availability of softwaredefined networks. In addition, a federated learning framework is introduced to identify DDoS attacks in SDN environments, which simultaneously protects privacy while maintaining high detection accuracy. Our proposed solution reduces the risk of data breaches and protects the confidentiality of sensitive data by training models locally. We have used FL to train three classifiers: Deep neural networks (DNN), convolutional neural networks (CNN) and (LSTM) Long Short-term memory to classify two categories of DDoS attacks, namely: UDP Flood, TCP SYN. Achieving 99.99% accuracy and a 99.99% F1-score on TCP SYN floods, alongside 99.94% and 99.97% on UDP floods, our federated CNN not only exceeds the most robust centralized benchmarks but also outperforms our own federated DNN and LSTM models, establishing a new benchmark for SDN DDoS detection while ensuring complete privacy of raw traffic within each domain.
Research on underwater wireless sensor networks (UWSNs) has been significant for applications such as forecasting adversity and disaster, hydrological and military surveillance, seepage monitoring, and underwater triangulation. These networks however, face challenges like significant delay spread, soaring interference, noise, jarring environments, poor connectivity, and restricted battery life. They also cause significant problems in terms of energy efficiency and network longevity. Nodes in UWSNs are subject to additional limitations, including fluctuating ambient conditions, large propagation delays, and limited energy supplies. Designing routing protocols for UWSNs is a promising solution to overcome these issues. The Adaptive Fuzzy Energyefficient Clustering and Energy Optimization (AFECEO) protocol, proposed especially for UWSNs, is thoroughly evaluated in this study in comparison to six popular clustering protocols: GEC, LEACH, PEGASIS, DCHS, DEEC, and LGCA. The fuzzy logic-based adaptive clustering process used by AFECEO dynamically chooses cluster heads by taking into account variables including distance to the sink, node residual energy, and underwater communication difficulties such as acoustic signal attenuation. According to simulation data, under various underwater settings, AFECEO performs better than its competitors in a number of critical performance parameters, such as average remaining energy, dead node count, and network longevity. Notably, AFECEO outperforms conventional protocols in terms of residual energy by up to 45% and dead node reduction by 60 %, guaranteeing improved energy optimization and dependable data transfer in UWSNs. This study demonstrates how well AFECEO works as a reliable option for energy-efficient communication in submerged settings, opening the door for more advanced monitoring and exploration uses.
In this work, a compact broadband patch antenna integrated with an Artificial Magnetic Conductor (AMC) structure is proposed for WiMAX $(3.6 / 4.8 \text{GHz})$ and WLAN (5.1/5.825 GHz) 5G applications. The antenna, fed by a 50Ω microstrip line, consists of a partial ground $(0.035 ~\text{mm})$, an FR4 substrate $\left(\varepsilon_{r}=4.4\right.$, thickness $=1.6 ~\text{mm}$, loss tangent = 0.025), and a rectangular patch with parasitic elements and an elliptical slot. It achieves a wide impedance bandwidth of 3.03 GHz within a compact size of $26.5 \times 29 \times 1.6 ~\text{mm}^{3}$. A $4 \times 4 \text{AMC}$ array is placed on the back to suppress back radiation lobes and enhance gain. The overall antenna size is $40 \times 40 \times 10.4 ~\text{mm}^{3}$. Simulated and measured results confirm a bandwidth from 3.62 to 6.65 GHz, with a peak gain of 10 dB at 6 GHz and 8.5 dB at 5.8 GHz, making it a promising candidate for WiMAX and WLAN $\mathbf{5 G}$ systems.
The shared and open-access nature of wireless technologies makes wireless networks susceptible to interference, such as jamming, that could lead to delays in communication and result in delays with corresponding timely information updates. Such delays in information updates can be crucial when an operator controls mobile objects since their untimely updates might lead to drastic consequences. Specifically, we consider an operator communicating with drones performing a joint mission in a protected zone. The protector (jammer) employs a spherical jamming signal to jam the communication of the operator with the drones to fail their joint mission. To design an anti-jamming strategy in such a scenario involving agents (the operator and the jammer), with different objectives, game theory is employed. Traditionally, in such a background, the anti-jamming strategy aims to minimize the total information freshness reflected by the total delay in information update. A con of such a traditional approach is that, due to limited resources, the operator might have to sacrifice the delay in information update for some drones to minimize the total delay. This could fail the joint drone mission since some drones might follow outdated commands. In this paper, motivated by this observation, we suggest using the guaranteed information freshness metric as a cost function for the operator. Under this metric, we specify the maximal time passed since the last data update for each of the drones. An anti-jamming strategy is derived, and its advantage in comparison with the minimizing total information freshness strategy is illustrated.
Object tracking in real-time applications often relies on the tracking-by-detection paradigm, where object localization and temporal association are decoupled. While methods such as DeepSORT integrate a detector like YOLO with additional CNN-based ReID module for appearance matching, the ReID network imposes significant computational and memory burdens, especially on edge or CPU-only systems. This work explores the use of zero-training appearance descriptors extracted directly from YOLOv8's internal multi-scale feature maps (P3, P4, P5), used individually as lightweight alternatives to conventional ReID embeddings. Our method selects a feature scale based on object size and retrieves a fixed-length feature vector via a simple center-cell lookup. On the MOT20 train benchmark, our fastest configuration, (i.e. based on the P3 feature map) achieves more than 90% of the HOTA score of the standard DeepSORT with ReID, while reducing tracking time per frame from 0.279 s to 0.035 s on a CPU. These results demonstrate that YOLOv8's native features hold sufficient discriminative power for training-free, real-time object association, eliminating the need for costly ReID networks in resource-constrained deployments.
Connected and Automated Mobility (CCAM) is undergoing a paradigm shift, with safety and efficiency increasingly dependent on connectivity. Cooperative Intelligent Transport Systems (C-ITS) support this transformation by enabling the exchange of Cooperative Awareness Messages (CAMs) between vehicles and roadside infrastructure. These messages, transmitted periodically at $\text{1 - 1 0 ~ H z}$, must be digitally signed in compliance with ETSI standards using Pseudonym Certificates (PCs). However, this security process introduces a significant overhead, as the size of the security data can be up to three times larger than the CAM payload, thereby consuming a considerable portion of the communication channel bandwidth. In this paper we propose a new authentication scheme based on deep learning. Instead of exchanging signed CAMs every time, the vehicles will authenticate each other once to establish cluster-based trust relationships, and then they will exchange only unsigned CAMs during the cluster lifetime. To ensure security within the cluster, an unsupervised variational autoencoder analyzes vehicle behavior to detect anomalies and confirm that each vehicle remains the same entity originally authenticated. Through simulations using OMNeT++, SUMO, and Artery, our method achieved a $\text{48.9 \%}$ reduction in the volume of messages exchanged between vehicles, significantly decreasing communication channel overhead.
In the need for high-security mechanisms, Trust Management Systems (TMSs) are implemented in vehicular networks such as Vehicular Ad Hoc Networks (VANETs) and the Internet of Vehicles (IoV) to ensure reliable interactions between vehicles. These systems nowadays are a key factor in building up security by evaluating and managing trust relationships among network participants. However, TMSs are inherently very vulnerable to Trust Manipulation Attacks (TMA), where we find malicious nodes attempting to deceive trust models by exploiting their evaluation mechanisms. One critical variant of this attack involves malicious nodes creating multiple fake identities, known as the Sybil attack, to falsely reinforce their trustworthiness. This deception will totally mislead legitimate vehicles, manipulate the decision-making processes, and at the end compromise the overall security and reliability of the network. To address this challenge, we propose a Zero-Knowledge Proof (ZKP)-based trust authentication scheme that ensures each vehicle can prove its legitimacy without exposing sensitive information. Our approach can and will prevent attackers from fabricating multiple identities to manipulate trust values. By integrating cryptographic authentication with trust management, our method significantly strengthens security and ensures that only legitimate vehicles can participate in trust-based evaluations. Through simulations, we demonstrate the effectiveness of our proposed solution in reducing the risk of Sybil-based. The results indicate that our approach not only enhances security but also maintains efficient trust computation, making it a viable solution for real-world vehicular networks.
Programmable matter in the context of microrobots is like mobile microcomputers with size of a millimeter, which can move around each other, communicate to form different shapes. Current challenges are as much theoretical as technological. The main research challenges in computer science focus on clustering, synchronization time, decision-making, analysis of detected data, security and the autoconfiguration process, which today remains the most fundamental. When the number of modules is very high, the self-reconfiguration challenge becomes more crucial for applications requiring rapid transition between two forms. Clustering can enable parallel transitions by allowing geographically close modules to make intra-cluster transitions, thus reducing reconfiguration complexity, the number of transitions and the time required. Clustering-based autoconfiguration solutions are available, but they do not include the security aspect. In this paper, we propose a security protocol for clustering-based self-reconfiguration solutions. This solution relies on resource-efficient cryptographic mechanisms to implement a robust authentication solution that underpins a flexible key management mechanism coupled with a confidential data exchange of the resulting structure
