Federated learning (FL) makes it possible to train models across distributed data sources without collecting raw data in one place. However, even in federated settings, trained models may still leak sensitive information at inference time. This problem is particularly evident for Gaussian Process regression (GPR), where predictive uncertainty is explicitly returned and can differ between training and non-training samples. Such differences can be exploited for membership inference. In this work, we examine inference-time privacy and robustness in federated GPR by focusing on the behavior of predictive variance. To enable scalable training, we employ a Random Fourier Feature approximation together with an Alternating Direction Method of Multipliers (ADMM) based distributed optimization scheme. On top of this learning framework, we apply key-dependent orthogonal feature transformations that enable multi-key inference time access control. When inference is performed using the correct key, prediction accuracy and uncertainty behavior remain close to those of plaintext federated GPR. When incorrect or mismatched keys are used, prediction errors increase sharply and predictive variance becomes uniformly large. Experimental results show that this variance inflation removes the usual gap between training and unseen samples, reducing the effectiveness of variance-based membership inference. Importantly, this effect arises without adding noise or relying on cryptographic operations. These findings suggest that predictive uncertainty can play a practical role in enforcing inference-time access control and improving privacy robustness in federated Gaussian Process models.

Industrial maintenance increasingly relies on extended reality (XR) technologies, yet comprehensive analysis comparing AR, VR, and MR implementations with human-centric evaluation remains limited. This systematic literature review employs the PRISMA methodology to analyze 95 primary studies (2014–2025) that investigate implementation patterns, benefits, challenges, and evaluation criteria across XR modalities. AR dominates operational guidance (71.6%), VR prevails in training (38.9%), and MR enables collaborative maintenance (6.3%). Temporal analysis reveals three evolutionary phases: basic visualization (2014–2017), cognitive enhancement (2018–2021), and AI-integrated adaptive systems (2022–2025), mirroring the transition from Maintenance 4.0 to 5.0. Demonstrated benefits include 38% task completion time reductions and 92.4% error decreases. Human-centric factors appear in 44.2% of studies overall, with temporal analysis showing a progressive increase from approximately 25% in 2014–2017 to 58% in 2022–2025, substantiating a paradigm shift toward prioritizing cognitive support and user experience. Two novel frameworks advance theoretical understanding: an Evaluation Framework that ex-presses effectiveness as a function of technology, task complexity, and human factors; and a Technology-Task-Context Alignment Model that prescribes optimal XR-maintenance pairings. Critical gaps include the need for longitudinal field studies, standardized evaluation protocols, and economic analyses.

Semantic segmentation is a fundamental component of autonomous driving systems, enabling accurate scene understanding and object-level perception. However, achieving precise instance-level delineation while maintaining real-time performance on resource-constrained platforms remains a significant challenge, particularly for edge deployment scenarios. This paper proposes a lightweight dual-YOLOv8 fusion framework for instance-aware semantic segmentation in autonomous driving applications. The proposed approach integrates YOLOv8n-seg and YOLOv8s-seg through a multi-scale fusion strategy that exploits their complementary feature representations to improve the segmentation of road-relevant objects, including cars, buses, trucks, and motorcycles. The framework is evaluated on the Reetiquetado de Vehiculos dataset using standard instance-level segmentation metrics. Experimental results demonstrate strong performance, achieving an overall mAP@0.5 of 92.9% and mAP@0.5:0.95 of 80.8%, while maintaining real-time inference with an average processing time of 7.9 ms per image (126 FPS) on an NVIDIA RTX 3050 GPU. Class-wise and confidence-based analyses confirm consistent segmentation accuracy across vehicle categories, highlighting the robustness of the proposed fusion strategy in handling scale variation, occlusions, and object diversity. In addition, an embedded deployment analysis provides insight into the feasibility and practical constraints of deploying the proposed framework on representative edge platforms. Overall, the proposed dual-YOLOv8 fusion framework achieves an effective balance between segmentation accuracy and computational efficiency, making it suitable for real-time autonomous driving perception on edge ARM/GPU platforms and Advanced Driver Assistance Systems (ADAS).

The study analyses Turkish and English tweets about climate change on the social media platform Twitter and comparatively examines individuals” perceptions, concerns, and emotional reactions to this issue. A total of 2,046 Turkish and 18,000 English tweets were collected; 1,104 Turkish and 6,449 English tweets were analyzed after the cleaning process. Artificial intelligence-based methods such as text mining, sentiment analysis, and topic modelling are used. Topic modelling with Latent Dirichlet Allocation (LDA) identified prominent themes in tweets in both languages. Sentiment analysis is performed using deep learning techniques to categories tweets into positive, negative, and neutral categories. The findings show that English tweets contain stronger emotional reactions, while Turkish tweets contain a higher proportion of neutral expressions. Additionally, it was observed that the perception of climate change can differ in local and global contexts. Based on a multidimensional analysis of social media data, the study provides valuable insights into the development of environmental communication strategies. The comparison of Turkish and English tweets contributes to understanding the effects of cultural contexts on climate change perception. The findings have important implications for policymakers and environmental awareness campaigns, as they highlight the need for tailored communication strategies that consider cultural differences in climate change perception.

Biomedical question answering presents significant challenges due to the complexity of biomedical language and the need for precise information retrieval. This study aims to improve the performance of a biomedical information retrieval system through a hybrid learning-to-rank framework. Specifically, we combine lexical (BM25) and semantic (BioBERT) representations to form hybrid inputs for RankFormer, a transformer-based ranking model. This hybrid representation captures both surface-level term matching and deep contextual understanding. Experiments conducted on the BioASQ dataset show that our approach achieves better ranking performance compared to the standalone lexical or neural baselines, reaching a MAP@10 of 0.9614 and an nDCG@10 of 0.9320. These results highlight the effectiveness of hybrid input representations in enhancing biomedical answer ranking.

Fog computing extends cloud capabilities to the net-work edge, providing low-latency services for mission-critical applications such as healthcare and industrial automation. However, validating security-aware scheduling policies in this distributed paradigm remains a challenge due to the lack of native support for trust modeling in standard simulation tools like CloudSim and iFogSim. Existing simulators focus primarily on resource provisioning, cost, and energy metrics, often neglecting “Trust” and “Data Confidentiality” as first-class simulation parameters. This study presents the design and implementation of a trust-aware extension for WorkflowSim. We detail the software architecture modifications required to support attribute-based trust verification, secure task fragmentation, and encryption overhead modeling. Unlike standard simulators, our extended framework treats Trust as a dynamic entity in the Virtual Machine (VM) and Task class hierarchy. We validate the framework’s correctness through unit tests and scenario-based trace analysis, demonstrating its ability to accurately model security-performance trade-offs in heterogeneous fog environments. Experimental results indicate that enabling trust-aware scheduling increases makespan by approximately 19% and cost by 16%, while achieving a security score of 0.88 compared to 0.65 for the baseline. The simulation engine overhead remains below 7.1% even for 1000-task workflows. The source code and design patterns presented provide researchers with a robust, extensible tool for evaluating secure scheduling algorithms without the cost of physical testbeds.

This study investigates a class-splitting strategy for improving object detection under limited training data using YOLOv11n with transfer learning and data augmentation for agricultural images containing leaves and peppers. The proposed approach evaluates leaf-only, pepper-only, and combined-class configurations using mAP@0.5, mAP@0.5:0.95, precision, recall, and F1-score to examine how class splitting affects detection performance. On the small validation set used in this study, single-class training improved performance relative to the combined-class baseline, but the results should be interpreted as preliminary because the validation set contains only two samples.

Recommender systems have been developed to serve users and provide them with the best suggestions. Despite their success, offering fully identical recommendations to users’ preferences remains a difficult task, where the complexity of human taste results in different challenges. Grey sheep user phenomena continues to be one of the most common, where the user is defined by his unique interactions with the system, making it difficult to associate with similar users, as he rarely agrees with them. In this study, we presented a new approach for identifying grey sheep users. It is based on the taste context and nature of user interaction with the system. We grouped similar users using an enhanced Kmedoids clustering method with a new dissimilarity metric and introduced a novel process to distinguish between users. The differentiation is achieved by assigning weights to each cluster based on how much it reflects the grey sheep user characteristics. We evaluated the efficiency of Grey Threshold Medoids (GTMedoids) using the FilmTrust and MovieLens 100k datasets. The results show the superior performance of our approach in detecting grey sheep users.

The rapid advancement of Industry 4.0 and adoption of cyber-physical production systems (CPPS) demand real-time, adaptive, and privacy-preserving optimization that conventional centralized AI architectures cannot adequately provide. Such approaches remain susceptible to data privacy vulnerabilities, regulatory non-compliance (GDPR, CCPA), communication latency, and insufficient responsiveness to dynamic production variability within edge device constraints. Although federated learning (FL) offers a promising paradigm for distributed privacy-sensitive intelligence, existing implementations fail to address practical security requirements and hardware limitations of shop-floor edge environments, rendering real-world deployment infeasible. This study introduces FedSecure-OPE, a secure autonomous AI framework designed to concurrently optimize production scheduling, quality management, and predictive maintenance across distributed manufacturing cells. The framework integrates homomorphic encryption-based federated aggregation, secure multi-party computation (SMPC) for model updates, and dynamic neural architecture search subject to edge hardware constraints. FedSecure-OPE is evaluated against centralized deep learning (Model A) and unsecured federated learning (Model B) using the Manufacturing Cyber-physical Middleware Testbed (MCMT) and Synthetic Manufacturing Trace (SMT) datasets. All experimental results were obtained through digital twin simulation under hardware emulation and have not been validated on physical edge environments. Within this context, FedSecure-OPE (Model C) achieves 31.2% and 16.8% operational performance improvements over Models A and B respectively, reduces edge energy consumption by 43.7%, attains 99.2% cryptographically protected model-update coverage under defined simulation security assumptions, and maintains average inference latency of 38 ms per control cycle. These findings establish a simulation-based foundation for security-conscious federated AI in smart manufacturing, while underscoring the necessity of future validation in physical environments.

Intelligent fruit grading systems require automated sensing solutions capable of rapid and reliable non-destructive internal quality assessment. While near-infrared spectroscopy has been widely used for soluble solids content (SSC) prediction, most existing studies rely on manually acquired spectra, limiting scalability in smart agricultural environments. Although online Vis–NIR systems based on transmission configurations have been reported, automated interactance-based systems designed for deployment-oriented grading remain limited. This study presents the design and validation of an automated interactance near-infrared spectral acquisition system for mandarin SSC evaluation. The system integrates controlled clamping, rotational positioning, and automated probe actuation to ensure stable optical geometry and repeatable probe–fruit contact during measurement. Spectral consistency was assessed by comparing consecutive scans obtained using manual and automated acquisition modes. The automated system reduced spectral dispersion among consecutive acquisitions within a measurement session by approximately 65%relative to manual measurement, indicating improved acquisition stability. Chemometric models based on partial least squares regression, support vector regression, and extremely randomized trees were developed under multiple preprocessing strategies. Prediction performance under automated acquisition remained within the same range as manual measurement, with several preprocessing–model combinations (particularly PLS and SVR with smoothing-based preprocessing), showing slightly higher Rp values and lower RMSEp values under automated acquisition. The findings demonstrate the feasibility of the proposed auto-mated system for stable interactance spectral acquisition suitable for SSC prediction, supporting its potential future integration into automated fruit quality assessment systems.