Pancreatic pathologies, including pancreatitis and adenocarcinoma, frequently disrupt exocrine-endocrine synergy, leading to pancreatogenic (type 3c) diabetes. Existing computational models often treat these systems in isolation, failing to capture the mechanistic transition to endocrine failure. This study proposes a hybrid modeling framework integrating physiologically based pharmacokinetics (PBPK) with Neural Ordinary Differential Equations (Neural ODEs). The PBPK backbone defines anatomical compartments, including acinar tissue and the islet microenvironment, to govern concentration-balance kinetics. At the same time, Neural ODE residuals are introduced on two specified coupling channels to represent latent nonlinearities such as inflammation-mediated β-cell decline. Designed to synthesize transcriptomic, enzymatic, and clinical datasets, this architecture maintains mechanistic interpretability while capturing complex dynamical interactions. In addition, this work establishes a foundation for predictive simulation and systematic analysis of the pathways linking exocrine injury to endocrine dysfunction. By merging physiological modeling with machine learning, the proposed framework enables robust investigation into the progression of multi-system pancreatic diseases.
This paper proposes ImageTrust, a detection system which addresses some of the shortcomings usually associated with AI media detection, namely, poor performance after recompression and the missing calibrated uncertainty estimates along with predictions. ImageTrust concatenates embeddings from ResNet-50 (2,048-d), EfficientNet-B0 (1,280-d) and ViT-B/16 (768-d) into a single 4,096-dimensional embedding, which is classified by both an XGBoost and MLP meta-classifier. We employ temperature scaling to decrease calibration error (ECE = 0.016 for XGBoost) and use conformal prediction to provide 95.2
Tourism is a valuable financial support for Governments worldwide. Therefore, Governments promote it by undertaking new facility construction projects for tourist spots. These projects are to be selected from a project pool that satisfies a budget constraint. To make this selection process transparent, the Government often involves the public using technological platforms available in existing and upcoming smart cities. Several mechanisms have been proposed to date to either optimize or extend the process’s functionality. This article discusses the project selection process in the smart city context, then formalizes the problem, and provides a detailed analysis and performance evaluation of recent works that either optimize or extend its functionality. This article concludes by summarizing the challenges, including over-tourism, open issues, and directions for future work in this domain.
Thanks to its fully customizable stack and its alignment with European initiatives toward technological autonomy, RISC-V is emerging as a viable alternative to other more mature architectures. Yet despite this momentum, deploying cloud environments and running cloud-native workloads on RISC-V remains an open challenge. This paper reports on our experience deploying experimental Kubernetes cluster on RISC-V hardware and evaluating two representative workloads: Apache Spark for large-scale data analytics and PyTorch for convolutional neural network training, both running inside containerized environments. To enable a meaningful cross-architecture comparison against an x86 baseline, we adopt two throughput-normalized metrics, Queries Per Cycle (QPC) for Spark and Frames Per Cycle (FPC) for PyTorch. Our results show that Kubernetes operates correctly on RISC-V with no fundamental orchestration-level barriers, while a significant performance gap with x86 persists, driven primarily by memory hierarchy limitations and the relative immaturity of the software stack. These results provide a reproducible baseline for evaluating future, higher-performance RISC-V platforms as the ecosystem matures.
With the rapid growth of the logistics and delivery industry, there has been a growing demand in recent years for delivery planning that balances safety and efficiency. In this paper, the authors propose a safe and efficient delivery planning method based on road characteristics, taking into account the risk level and distance of each road segment calculated using telematics data. The proposed method models the actual road network as an undirected graph, assigns distance and risk levels to each link, and constructs a mathematical model that optimizes the actual routes traveled by delivery vehicles. To reduce computational load, the authors introduce delivery area segmentation via clustering and perform route optimization using a hybrid genetic algorithm (GA). Simulation results on a large-scale road network covering central Fukui City showed that the proposed method can formulate delivery plans appropriately considering both driving distance and risk levels by adjusting the weighting coefficients. Furthermore, through comparison with the previous method, the authors confirmed that the proposed method can reduce a computation time even for large-scale networks.
State-of-the-art Deep Neural Networks (DNNs), such as ResNet, VGG, and large Transformers, incorporate extensive parameterisation, limiting deployment on resource-constrained edge computing and IoT platforms. Vector quantisation (VQ) addresses storage and memory constraints. Existing K-means-based approaches encounter the codebook overhead dilemma, wherein sufficiently large codebooks to maintain reconstruction accuracy may result in substantial memory consumption, offsetting the intended efficiency. To address this, we propose Group-based Permutation Quantisation (GPQ), which employs Permutation Group theory to construct a structured codebook derived from a single prototype vector. To preserve accuracy amid quantisation-induced distributional shifts, Channel-wise Scaling and a Bias Correction mechanism are integrated, targeting inter-channel weight distribution variability. Empirical evaluation using ResNet-18 on the CIFAR-10 dataset demonstrates that GPQ achieves Top-1 accuracy of 92.49 × compression ratio and reduces codebook storage by over 500 × relative to traditional unstructured VQ, thoroughly mitigating codebook storage overhead.
Mask-wearing detection plays a crucial role in controlling and preventing the spread of infectious diseases. However, this task presents several challenges due to the occlusion of facial features, the diversity of mask designs, and environmental factors such as lighting and camera angles. In this study, we propose an intelligent automated system capable of classifying three distinct mask-wearing states: correctly worn, not worn, and improperly worn. A significant contribution of this work is the enhanced identification of improper mask usage, which is often a major challenge for manual monitoring and existing automated systems. Additionally, the system integrates a real-time distance measurement algorithm to identify and alert when social distancing violations occur. The model was trained and evaluated on a dataset of 1,655 manually annotated images collected under diverse real-world conditions. Experimental results demonstrate that the proposed system achieves a classification accuracy of 97.22
Crossover operations in Differential Evolution generate new candidate solutions through parent recombination. Standard approaches typically employ constant crossover, where genes are inherited from parents based on fixed probabilities. This paper extends Maximum Crossover - a gene-wise refinement mechanism previously proposed for single-objective optimization - to the multi-objective setting within a Non-Dominated Sorting Differential Evolution (NSDE) framework. In the extended formulation, an offspring is first generated using constant crossover, after which gene sources are iteratively flipped and retained only when multi-objective acceptance criteria are satisfied. The approach is first validated on 3 standard multi-objective benchmark functions, where it is found to yield improved Pareto-front quality compared to constant crossover under identical NSDE selection. The same framework is then applied to a Neural Architecture Search problem on an autoencoder for 13-dimensional noisy industrial sensor data from the steel manufacturing domain, where reconstruction accuracy (MAPE), number of model parameters, and latent layer compression are optimized concurrently using a problem-specific two-stage acceptance criterion. Overall, the results indicate that extending Maximum Crossover to multi-objective optimization leads to fair and consistent improvements over constant crossover in both benchmark and application-driven settings.
Microsurgery requires extremely precise manipulation, but physiological hand tremor degrades surgical accuracy. To enable real-time tremor suppression without using inertial sensors, an image-based motion estimation method suitable for FPGA implementations was proposed. The method employs a pattern-matching optical flow algorithm based on rotation-invariant histogram of oriented gradients (RIHOG) features. However, its design process lacks quantitative analysis on accuracy and thus it has plenty of room for improvement. In this paper, we introduce three modifications: a spiral search strategy, a Top2 differential reliability measure, and an elimination of a distance-based penalty term. Then we show a quantitative analysis of the introduced modifications using synthetic motion sequences with known ground truth. Based on the Euclidean distance between an average optical flow and the ground truth, it is revealed that the spiral search without the penalty term improves accuracy without increasing computational complexity.
The EEMVTO (Energy-Efficient Multi-Version Timestamp Ordering) algorithm is proposed in our previous studies to increase the throughput of servers as well as reducing the electric energy (EE) consumption. However, the EEMVTO algorithm does not control the number of transactions which are concurrently performed. When a large number of transactions are performed concurrently, the execution time (ET) of methods performed in each server increases since read and write rates allocated to each method decreases. Then, the ET of transactions and the EE consumed by servers increases. In this paper, the EEMVTO-CC (EEMVTO with Controlling the number of Concurrent transactions) is newly proposed to reduce the ET of transactions and the EE consumed by servers when the number of concurrent transactions increases. We show the EEMVTO-CC algorithm can reduce the ET of each transaction and the EE consumption of servers compared with the EEMVTO algorithm when the number of concurrent transactions increases by simulation.
Shift-left security has improved early vulnerability detection in the Software Development Lifecycle (SDLC), but post-deployment security remains under-governed. This paper examines governance, compliance, and monitoring challenges in cloud-native production environments and proposes an integrated framework that combines Cloud Security Posture Management (CSPM), runtime threat intelligence, and automated compliance orchestration. By integrating Azure Sentinel, Palo Alto Prisma Cloud, and SentinelOne with infrastructure-as-code (IaC) scanning tools such as Trivy, tfsec, and Checkov, the framework enables continuous security assurance across multi-cloud deployments. Evaluation in three enterprise contexts shows a 60
Unmanned Aerial Vehicles (UAVs) are increasingly adopted to sense, map, and monitor complex environments, such as urban areas, and require more and more reliable scene understanding to support monitoring tasks. Semantic segmentation provides dense pixel-level labeling, yet deploying and adapting segmentation models in UAV swarms remains challenging due to strict on-board resource constraints, heterogeneous data distributions, and the opportunity offered by multimodal sensing. In this paper, federated learning is applied to train a lightweight RGB-D+ altitude semantic segmentation model across distributed UAV clients without centralizing raw data. We combine an efficiency-oriented architecture based on MobileNetV3+LR-ASPP, multimodal early fusion, and a FedAvg-based training protocol. We evaluate the proposed approach on the SynDrone dataset under both IID and altitude-specialized non-IID client partitions, while varying the swarm size and the client data distribution. The results indicate that increasing the federation scale improves performance on mixed-altitude testing and that the accuracy degradation under altitude-based non-IID partitions can be reduced with sufficient rounds and client participation. In the best large-swarm configuration, we achieve 46.1
SmartCityHub is a public platform for accelerating safe, sustainable, and scalable smart city innovation in Luxembourg by enabling cities and companies to experiment and validate AI solutions in real-life urban environments. It provides a secure environment where cities and companies can test AI before full deployment. SmartCityHub also connects municipalities from the Greater Region with European technology providers and researchers, fostering collaboration and innovation. This paper proposes a methodological framework for the co-design of Local Digital Twins (LDTs) based on an iterative and data-driven approach. It relies on the real-world deployment of an LDT with the city of Differdange, in the heart of the industrial basin in South Luxembourg, supporting its transition towards carbon neutrality. The paper identifies key challenges and derives practical lessons for developing scalable and interoperable LDTs in smart cities.
In high-mix, low-volume production, prototyping, maintenance, repair and small-scale environments, fastening quality and operational reproducibility are important for screwdriving tasks. However, fastening results are affected by multiple factors, including tool operation, posture and the spatial relationship between the operator and workpiece. For this reason, the final fastening state alone is insufficient for fully evaluating the work process. In this study, we propose a motion analysis system for screw-tightening tasks based on a depth camera. The proposed system acquires the operator’s three-dimensional skeletal information from RGB-D images and evaluates work posture while taking into consideration the operator’s physique. Also, the proposed system recognizes the screwdriver in RGB images by segmentation and analyzes of the tangential and radial components of screwdriver operation by applying optical flow to the recognized region. The final state is also evaluated through recognition of the assembled object. In the experiments, correct screw-tightening conditions were compared with incorrect ones. The evaluation results confirmed that differences between correct and incorrect postures can be represented using three-dimensional skeletal information. Also, correct screw-tightening operations showed a sustained increase in the tangential component, whereas incorrect operations showed only transient fluctuations together with a radial component. The experimental results indicate that as the assembly approaches completion, correct tightening is recognized more reliably than incorrect tightening. These results suggest that the influence of hand occlusion is significant during the early to middle stages of the task, whereas differences in the tightening state are reflected in the recognition results toward the end of the task.
In Vietnam, durian is one of the most important fruit crops for the economy, especially in Binh Phuoc, Tien Giang, and Vinh Long provinces. However, durian orchards are often affected by leaf diseases that reduce yield and fruit quality, and traditional visual assessment remains subjective, requiring considerable experience and time. To address this problem, this study proposes a deep learning-based method to automatically identify and classify durian leaf diseases using a real-world dataset from Vietnam. This dataset, collected from durian orchards in Binh Phuoc and Tien Giang in January 2025, consists of 2,595 high-resolution images collected under real-world conditions and classified into six classes. Transfer learning is applied with modern convolutional neural networks, specifically DenseNet169 and Xception. Our findings reveal that DenseNet169 achieved superior performance, with 95.68
Data spaces transform data quality from an internal engineering concern into a market, trust, and certification property. The current landscape is fragmented: enterprise frameworks such as DAMA and ISO/IEC SQuaRE define quality characteristics; the evaluation of the UNE 0079 and UNE 0081 standards are a practical extension of the ISO standards; observability platforms such as Monte Carlo focus on freshness, quality, volume, schema and lineage; validation tools such as Great Expectations (GX), Deequ, dbt, and Databricks implement concrete checks; and IEEE P2510 as part of the Libelium iris360 solution extends it with IoT and sensor-centric quality parameters such as Digital Logical Unit (DLU) values, accuracy/error evidence, security, encryption, tamper. This paper proposes a standard-first and implementation-oriented synthesis. First, it analyses the major standards and tools as complementary layers of a data-quality stack. Second, it defines a practical catalogue of data-quality parameters, explaining why each parameter is needed, where it is used, and how it can be calculated. Third, it organizes representative use cases for data at rest and data in motion in smart cities, industry, marketplaces, and AI pipelines. Finally, it proposes an evolution of Smart Data Models and the ETSI NGSI-LD standards for data modelling and ISO-IEC 11179 metadata definition, to make quality evidence machine-actionable through attribute sub-properties, linked DataQualityAssessment entities, sensor-quality profiles, quality rules, incidents, labels, and AI-output quality records. We require quality evidence to travel with the data: not as a single opaque score, but as auditable, multidimensional, standards-aligned metadata that enables the capacity to bring data quality as feature of any dataset, to guaranty that its integration into value-chains with AI models, digital twins, and more advanced capabilities as the emerging Physical AI, guaranty that agents can evaluate what is the level of trust, reliability, and guaranty that they can rely a decision or action.
The deployment of Local Digital Twins (LDTs) across European cities and communities is advancing rapidly under the Digital Europe programme, notably with the project Local Digital Twins for Sustainable and Smart Communities (LDT4SSC) supporting cross-border pilot initiatives toward an interoperable, federated ecosystem. A key challenge to this effort is the parallel coexistence of two technology families for structuring and exchanging data in local digital twins: one RDF-based; another NGSI-LD based. These families differ not only in their serialisation formats, but more fundamentally in their architectural assumptions, information models, approaches to vocabulary governance, and relationship to geospatial standards. This paper briefly analyses the structural tensions between the two paradigms, drawing on the requirements of the LDT4SSC Calls for Pilots and the emerging common architecture for the CitiVERSE European Digital Infrastructure Consortium (EDIC). Rather than asking whether interoperability between the two families is achievable, this paper argues why a blueprint for European LDTs should make this effort, inviting the LDT4SSC consortium, its pilot projects, the Minimal Interoperability Mechanisms (MIMs) Plus, and the general stakeholder community to develop how to technically achieve it.
Building high-quality datasets for fake news detection presents a significant challenge, especially for under-resourced languages with rich morphology, such as Albanian. We use an Albanian fact-checking portal to create a baseline dataset with binary classification of fake news. This dataset, containing 648 records of fake and true claims, is later augmented in two strategies: (i) using Large Language Models (Llama-3.1-8B, Mistral-7B, Llama-3.2-3B) using adversarial training, (ii) merging with existing datasets with the only existing Albanian fake news benchmark. For the nine training variants, we have used a fine-tuned XLM-RoBERTa-base. Across the three LLMs used, we detect no statistically significant benefit and high variance, concluding that the generated Albanian text introduces noise rather than being reliable. On the other hand, our findings show that domain-aware corpus merging, which up-samples the target corpus by a factor of 2, substantially outperforms the baseline and LLM-based methods. Also, domain shift analysis is a necessary complement to cross-validation when combining heterogeneous corpora.
Artificial intelligence (AI) is rapidly transforming work by automating routine cognitive and information-processing tasks. This paper develops a task-based mathematical model to estimate occupational exposure to AI by 2036 using data from the occupational database O*NET. Occupations are represented as vectors of task shares in six categories relevant to AI: cognitive, writing, analysis, technical, social, and creative tasks. The weights of the AI capabilities are applied to these task categories to construct an AI Exposure Index (AEI). Combining exposure with assumptions about AI adoption and task substitution yields a Replacement Pressure Index (RPI). The results suggest that occupations dominated by routine cognitive, writing-intensive, and analytical tasks face the highest potential disruption, while occupations that require substantial social interaction, contextual judgment, or creative reasoning remain comparatively less exposed. At the same time, the empirical results show that professional occupations, such as physicians, lawyers, engineers, and postsecondary teachers, may still exhibit meaningful exposure when large parts of their work involve documentation, analysis, and structured information processing.
Due to the transparent nature of public blockchains, several studies have shown that adversaries can trace transactions and link public addresses to real-world identities, raising serious privacy concerns. To mitigate these risks, several mechanisms based on relays and proxies have been proposed to introduce intermediaries that decouple users from their on-chain activity. In this paper, we present a literature review of these privacy-preserving mechanisms, covering network-layer protocols, on-chain approaches, and off-chain payment channel networks. For each approach, we outline the main proposals and discuss their limitations. Our comparison shows that each mechanism protects a single privacy dimension, whether the originator’s IP address, the on-chain sender or recipient address, or the off-chain payment path, and that no existing solution covers all three at the same time. Comprehensive privacy therefore requires combining mechanisms from different layers, which to our knowledge has not yet been investigated.
