Real-time monitoring systems for Industry 5.0 manufacturing settings need to accommodate a balance of data freshness, safety-critical compliance, and resource optimalization. In this work, we provide a workable UAV coordination framework for industrial applications, combining three aspects: a multi-factor dynamic priority scheme that integrates data freshness metrics with thermal anomaly identification, a drone selection strategy that aims for a balance of space comprehensiveness and task assignments, and a zone-coordination scheme for efficient scalability of conflict resolution. Using a rigorous digital twin simulation model for 23 industrial facilities over a total of seven stages of product manufacturing, we show that our holistic framework outperforms other approaches with an improvement of 34.7
Automatic modulation classification (AMC) is a critical task for cognitive radio and dynamic spectrum access. Its deployment on mobile, power-constrained platforms, such as unmanned aerial vehicles (UAVs), is severely limited by on-device energy consumption. This paper proposes a federated transfer learning (FTL) framework where models are pre-trained on a power-unlimited “ground station” and then fine-tuned with minimal energy on the “UAVs.” We compare the performance, on-device energy cost, and overall efficiency of several transformer-based architectures integrated in the proposed FTL framework against a baseline Deep CNN. We analyze two fine-tuning strategies: a full fine-tuning of the classifier head (FTL-3L) and a lightweight “linear probe” (FTL-1L) that tunes only the final layer. Our results on the RML2016.10b dataset show that while the large transformer model with FTL-3L achieves the highest peak accuracy (93.9 2× that of the Deep CNN. We demonstrate that the small transformer model with FTL-1L provides the optimal trade-off, achieving 93
Achieving energy efficiency is a critical challenge in 5G and Beyond (5GB) cellular networks, particularly in complex architectures such as Heterogeneous Cloud Radio Access Networks (H-CRAN). The dynamic nature of traffic demand, coupled with stringent Quality of Service (QoS) requirements, makes joint resource allocation and energy management a highly challenging optimization problem. In this paper, we investigate the joint control of Advanced Sleep Modes (ASM) and Resource Block (RB) allocation using Deep Reinforcement Learning (DRL). Specifically, we implement and compare two representative DRL algorithms: Deep Q-Network (DQN) and Proximal Policy Optimization (PPO), within a unified H-CRAN system model. Simulation results demonstrate that PPO consistently outperforms DQN in terms of convergence stability and energy efficiency, while satisfying QoS constraints. These results highlight the suitability of policy-gradient-based DRL methods for dynamic energy-aware resource management in future 5G and beyond networks.
The National Institute of Informatics (NII) annually assesses whether Identity Providers (IdPs) participating in the Japanese academic federation GakuNin comply with federation policies. The survey of about 50 items assigns institutions “A” (compliant) or “B” (needs improvement) grades, yet its weighted scoring method and partial manual evaluation limit consistency and trend detection. To address these issues, we applied unsupervised learning methods—K-Means, Gaussian Mixture Model (GMM), and hierarchical clustering—to survey data from FY2022–2024. The approach achieved a Precision of 0.984 or higher and a True Negative Rate (TNR) above 0.77, confirming its effectiveness in distinguishing institutional reliability. Analysis also identified one known evaluation error in the 2022 dataset. The study demonstrates how clustering supports objective trust assessment, reduces human bias, and enhances transparency and reproducibility in evaluating academic authentication federations.
Compression and deduplication are routinely combined in datacenter storage and streaming-service backends, yet the benefits and trade-offs of such pipelines remain unclear. This paper presents a reproducible single-node benchmark of four compression tools and three deduplication tools, evaluated individually and in combination on two contrasting workloads: a large structured CSV dataset and the Linux Kernel source tree. Using containerized experiments, we measure compression ratio, execution time, CPU and RAM usage, and I/O activity to study storage savings, execution order, and workload sensitivity. Results show that combining tools rarely improves space savings over the best standalone method, while execution order matters: deduplication before compression systematically reduces time and resource usage. Deduplication is most effective on redundancy-rich, multi-file data, whereas compressors dominate on flat structured datasets, and high-ratio compressors incur higher CPU/RAM costs. Overall, there is no universal “compress+dedupe” recipe; effective optimization is workload-aware and order-conscious.
User micro-emotion detection is a critical component of affective computing with significant applications in human-computer interaction, particularly in digital gaming. This study investigates the recognition of two distinct emotional states, boredom and stress, in 2D game players through micro-facial expression analysis using Extreme Learning Machines (ELMs). We propose a systematic approach that collects facial data from 237 participants during gameplay with three custom 2D games, resulting in 28,440 samples representing genuine emotional reactions. Our primary contribution focuses on a rigorous comparative evaluation of eleven different kernels employed in ELM classifiers, with particular emphasis on the statistical validity of comparisons. We categorize these kernels into two distinct groups: simple/linear kernels (with limited non-linear mapping capacity) and flexible non-linear kernels (with advanced non-linear mapping capacity). The RBF kernel achieved the best performance with 82.89
Fifth-generation (5G) and emerging sixth-generation (6G) networks pose new cybersecurity challenges due to their complexity and dynamic traffic. Intrusion Detection Systems (IDS) are essential, yet traditional methods often fall short. We introduce CBRI (CNN–BiLSTM–ResNet–Inception), a hybrid deep learning framework that integrates complementary architectures, hyperparameter optimization, and prediction fusion to enhance anomaly detection. Tested on the 5G-NIDD dataset, CBRI achieved of 98.06
In the big data era, massive data security faces threats due to traditional encryption’s high overhead, ciphertext deep learning’s accuracy loss, and poor coordination between privacy protection and encryption. This paper proposes a hybrid encryption optimization method and a ciphertext-privacy dual-layer framework. Based on the CKKS homomorphic scheme, it designs adaptive precision control and quadratic polynomial activation approximation, and realizes lightweight YOLO deployment on encrypted images via batch normalization fusion and knowledge distillation. Experiments on PyTorch/NVIDIA A100 show: CIFAR-10 ciphertext ResNet18 achieves 87.2
Advanced cybersecurity systems are often vulnerable for attacks supported by AI technologies. This paper will describe selected computer security protocols that are resistant for attacks carried out by computer bots and generative AI. Such protocols will be supported by cognitive systems that allow the creation of user-oriented protocols that use semantic features and perceptual parameters. Presented solutions allow to improve security standards for advanced cybersecurity systems.
The new Battery Regulation from the EU requires the use of Digital Battery Passports (DBPs) in order to improve transparency, sustainability, and circularity in the battery lifespan. In order to enable dynamic, impenetrable updates, this study introduces a novel blockchain-based architecture specifically designed for DBPs. It integrates smart contracts with IoT-enabled battery monitoring. Our solution integrates real-time IoT sensor data, which enables continuous verification of battery health, consumption, and environmental conditions, in contrast to conventional static record systems. The Proposed solution is built on top of Hyper ledger Fabric, a permission block chain technology that works well in business and government settings. Data sharing between authorized stakeholders is General Data Protection Regulation (GDPR)-aligned to private data channels, while smart contracts manage lifecycle events like recycling and compliance checks. Secure traceability, improved data integrity, and automatic compliance validation are all demonstrated by the prototype’s testing on a simulated battery supply chain. As required by EU law (EU 2023/1542), our results provide a scalable basis for the industrial deployment of DBPs.
This work develops a web-based spatio-temporal mapping system for London that integrates open crime data, tourist and accommodation points, and hotel review sentiment analysis. Monthly crime statistics appear as interactive clusters, with overlays for hotels, attractions, and Underground stations. Users can compare crime trends month-to-month, filter displays, and view safety scores from hotel reviews mapped by sentiment analysis. Findings show stable high-intensity crime areas near central tourist sites and stations serving multiple lines, with fluctuations during holiday months. Review data reveal crime is often underreported but correlates with safety mentions, suggesting this tool helps users assess local risks. The report also covers limitations, ethics, and future directions.
Modular robots adapt to complex tasks through self-reconfiguration but often lack global geometric knowledge for efficient shape analysis. We propose a fully distributed message-passing algorithm for border tracing in lattice-based modular robots that requires no centralized memory. The algorithm has two phases: first, it identifies all internal and external borders and builds communication bridges via shortest-path spanning trees; second, it enables any module to traverse the robot’s boundaries using these bridges. Simulations on different configurations demonstrate that the algorithm is scalable, with linear time complexity, and supports efficient border traversal for distributed geometric reasoning.
Experimental validation of IoT solutions typically requires setting up message brokers, orchestrating containers and integrating different devices for each new experiment. Current Digital Twin frameworks have limitations. Commercial platforms introduce vendor lock-in and privacy risks, open-source tools demand infrastructure expertise, and academic solutions tend to be domain-specific. We present a metadata-oriented framework that addresses these issues. Users define message structures and data flows declaratively, allowing cross-domain experiments without changes to the infrastructure. The architecture enables physical and virtual devices to coexist transparently, so researchers can start with virtual prototypes and incrementaly replace them with real hardware. We validated the approach with synthetic and physical IoT devices, showing seamless substitution between them, schema compliance, and reproducibility through portable configurations.
Large Language Models (LLMs) have become integral components in modern applications, from customer service chatbots to code completion tools. However, their widespread adoption has introduced significant security vulnerabilities. This paper investigates two critical attack vectors against LLM-based applications: prompt injection attacks and token drain attacks via false positive API responses. Prompt injection attacks exploit the model’s instruction-following capabilities to manipulate system behavior, bypass security constraints, and extract sensitive information. Token drain attacks represent a novel form of resource exhaustion where adversaries exploit cross-API interactions by generating false positive responses that trigger excessive token consumption in downstream LLM applications without the application’s knowledge. We analyze the mechanisms, attack vectors, and potential mitigations for both attack classes, providing a comprehensive security assessment based on recent research findings. Our analysis reveals that existing defenses are insufficient, with prompt injection attacks achieving success rates exceeding 84
Arabic is one of the most widely spoken languages globally, with over 400 million speakers across more than 20 countries. It holds a central role in religion, culture, and communication throughout the Arab world. This paper presents a comprehensive overview of existing approaches to Arabic diacritization, categorizing them into rule-based, statistical, hybrid, and deep learning methods. Each category reflects a different paradigm in how linguistic knowledge and computational models are applied to address this problem. In addition, we examine the main resources, corpora, and tools used in Arabic Natural Language Processing, as well as the current challenges that hinder research progress in this domain.
The widespread deployment of Large Language Models (LLMs), characterized by their vast parameter counts and sophisticated statistical pattern recognition, creates new attack surfaces such as jailbreaks. This work evaluates available vulnerability detection tool within the scope of LLM Red-Teaming by comparing state-of-the-art algorithms against emerging proposals exploiting LLMs for adversarial purposes. The evaluation centers on an performance analysis and reproducibility of four specific red-teaming tools: GCG, DeGCG, AutoDAN, and ReNeLLM.
As large language models (LLMs) increasingly integrate into mobile devices, there is a growing need to deploy lightweight models that can operate efficiently on resource-constrained edge devices. However, smaller models often suffer from reduced accuracy compared to their larger counterparts. In this paper, we investigate the use of Retrieval-Augmented Generation (RAG) to enhance the performance of weak LLMs on edge devices. We present a comprehensive benchmark consisting of three device-specific datasets totaling 1,500 questions (500 questions per dataset) generated from iPhone 8.4, Samsung GS25, and OnePlus 12 user manuals, covering hardware, phone settings, connectivity, privacy and security, troubleshooting, customization, apps and features, and accessibility. Our evaluation demonstrates that RAG can dramatically improve the accuracy of smaller models: gemma:2b and deepseek-r1:1.5b both achieve 100
Alzheimer’s disease is emerging as a significant public health concern. The elderly are typically affected by this neurodegenerative pathology. The signs are memory loss, accompanied by a harsher capacity for speech and different disabilities over the years. Thus, in recent years, early detection of Alzheimer’s disease has become an active field of study. In this article, we suggest a technique founded on the application of a convolutional neural network model known as Mask R-CNN for detecting the hippocampus, which is the region most impacted by this illness. Mask R-CNN includes four key steps: feature extraction using a residual network architecture, the use of the region proposal network, alignment of the region of interest, and finally, mask generation. We evaluated the proposed method in the Open Access Series of Imaging Studies (OASIS). The effectiveness of the Mask R-CNN method is demonstrated by its ability to accurately detect the hippocampal region. Subsequently, we integrated the U-Net architecture to segment this region. Finally, we performed the classification task using CNN, with an accuracy of 98.51
The issue of social isolation among the elderly is a critical one for public health. However, conventional detection methods are subject to dependency on labeled data, privacy concerns and lack of interpretability. This paper proposes a novel multi-agent framework for detecting social isolation via behavioral digital biomarkers in mobile communication data. The proposed approach addresses the limitations of supervised learning, model opacity and privacy concerns. It introduces three key innovations: (i) personalized anomaly detection without labeled clinical data is enabled by an unsupervised deep auto-encoder, (ii) Shap-based explainability for identifying behavioral changes that lead to alerts, and (iii) a federated multi-agent architecture that uses Personal Monitoring Agents for edge analysis and Coordinator Agents for intervention at a community level. The novelty lies in the coordinated multi-agent system that combines unsupervised learning, explainable AI and privacy-by-design. An evaluation was conducted using the Reality Mining dataset. The proposed framework outperforms four baseline methods with an F1-score of 0.85. It also reduces data exposure risk by 92
As cyberthreats against OT systems continue to rise, it is essential to proactively remove vulnerabilities before products are shipped from the producers to the consumers. This paper proposes an OT fuzzing framework that makes fuzzing for OT devices more efficient through partial collaboration between product vendors and verification providers. The architecture absorbs different interfaces of fuzzers and OT devices through their adapters. In addition, we design the coverage bitmap transmission encoding that makes the transmission of fuzzing feedback between the devices and fuzzers more efficient. A preliminary evaluation experiment shows this approach improves the fuzzing coverage by 10–30
