Accurate state-of-charge (SOC) prediction on personal mobile devices requires a model that can represent heterogeneous discharge mechanisms, transfer knowledge across users, and adapt from scarce target records while raw telemetry remains privacy isolated. This study proposes a module-decomposition-driven SOC prediction model that separates baseline drain, synchronous state response, and asynchronous multi-scale event memory within one discharge-dynamics formulation. The same three semantic coefficient blocks remain aligned through local source learning and round-wise sample-count-weighted federated aggregation. Target-side personalization then adjusts one locally derived personal discharge-rate scalar while preserving the shared temporal response, and forward SOC propagation produces threshold-based time-to-empty (TTE) estimates. A user-held-out evaluation on a public small-sample smartphone-log dataset uses eight target partitions and 447 final evaluation origins. Personalized fine-tuning reduced TTE MAE by 67.07% and RMSE by 71.88%, yielding an R2 of 0.936. The resulting task-specific architecture connects module-structured data processing, federated synthesis, few-shot adaptation, and local TTE inference while raw usage logs remain within user-specific data boundaries.
Wrist fractures present significant challenges in clinical diagnosis, often leading to treatment delays and compromised recoveries. Manual diagnosis is resource-intensive and error-prone. To address these challenges, we develop DeepWristFNet, a compact convolutional architecture designed for end-to-end wrist fracture classification using a small dataset of 193 wrist X-ray images. The DeepWristFNet architecture integrates multi-scale convolutional operations with Fire and Shuffle modules within a compact network design, followed by fully connected layers for binary classification. We applied data pre-processing techniques such as data augmentation, image enhancement, and image resizing to increase the number of images, improve image quality, and resize images to match the DeepWristFNet input size. The proposed method comprised three phases. In the first phase, we trained, validated, and tested end-to-end and achieved validation and testing accuracies of 99.04% and 87.93%, respectively. Testing was performed on a hold-out subset of image instances that was kept separate from model development. The evaluated hold-out images originated from the same dataset distribution and included the corresponding augmented variants. In the second phase, we further evaluated the learned representation by extracting deep features from the first fully connected layer of DeepWristFNet. ReliefF was then used to select informative features, which were subsequently evaluated using 10 conventional machine learning classifiers. Out of 10 classifiers, 5 classifiers, i.e., Efficient linear SVM, quadratic SVM, Narrow NN, wide NN, and medium NN, achieved 100% testing accuracy on unseen samples. In the third phase, an auxiliary Fuzzy Inference System provides an intensity-based foreground-background representation of the X-ray images. This representation provides complementary visual information for interpretation but is not intended to directly classify or localize fractures. Grad-CAM is additionally used to visualize image regions contributing to the DeepWristFNet predictions, providing a model-specific explanation of the classification decision. Additionally, we evaluated how well the proposed DeepWristFNet approach performed against cutting-edge deep transfer learning models. In the evaluated experiments, DeepWristFNet outperformed the compared pre-trained deep learning architectures on the unseen hold-out subset from the same dataset distribution (test set). This study demonstrates the potential of DeepWristFNet for wrist fracture classification under a small-data setting. However, further evaluation on larger, independently collected clinical datasets is required to establish its robustness, generalizability, and suitability for clinical decision support.
Background/Objectives: Reported performance on public brain tumour MRI benchmarks is hard to interpret because of near-duplicate train/test overlap, ImageNet pretraining bias, and single-seed evaluation. We address this with a leakage-aware evaluation protocol and a compact model trained entirely from scratch. Methods: WICA-Net-M is a 2.47 M-parameter CNN whose gated Haar Discrete Wavelet Transform (DWT) separates low- and high-frequency components and fuses them through a learnable gate, complemented by Coordinate Attention. We evaluate it on the standard and image-level deduplicated splits of the Nickparvar brain tumour MRI dataset under a three-seed, leakage-aware protocol, benchmark it against five ImageNet-pretrained baselines and conduct a near-duplicate overlap audit against BRISC 2025. Results: WICA-Net-M reaches 99.42 ± 0.16% accuracy on the standard V1 split and 95.25 ± 0.32% accuracy/95.17 ± 0.31% macro F1 on the deduplicated V2 split, closely matching five ImageNet-pretrained baselines (95.17–95.67%) with fewer parameters, with sub-half-point differences across three seeds indicating comparable rather than superior accuracy. The audit identifies 861 exact SHA-256 pairs involving 857 of the 1000 BRISC test images against the full Nickparvar collection. Alongside perceptual-hash candidate pairs, this exact overlap shows that BRISC cannot serve as independent external validation. Conclusions: The descriptive 4.17-point V1-to-V2 difference reflects the combined effects of duplicate removal, class rebalancing, and altered sample composition, with none separable from public releases, though it shows that a scratch-trained compact model can approach pretrained performance on the controlled split. Patient-level leakage remains unresolved because patient identifiers are unavailable. Leakage-aware, multi-seed evaluation should be standard before clinical translation.
Big Data Analytics (BDA) has evolved from a predominantly technical batch function into a socio-technical capability integrating cloud-native platforms, stream processing, Lakehouse architecture, machine learning operations (MLOps), visualization, governance, and managerial judgment. This paper proposes an integrated BDA decision-making framework developed through a structured conceptual synthesis of research on data platforms, analytical capabilities, decision processes, organizational readiness, technology adoption, governance, and responsible artificial intelligence. The framework comprises seven interconnected stages: data sources, ingestion and integration, storage and platform, processing, analytics and artificial intelligence, visualization and interpretation, and decision, action, and learning. Governance, human oversight, organizational readiness, task characteristics, and continuous feedback influence all stages. Key implementation requirements include data quality, interoperability, security, privacy, scalability, cost, explainability, bias, skills, and sustainability. The proposed configurable reference architecture links technical integration, task–analytics fit, governance assurance, human judgment, and organizational readiness with decision quality and organizational outcomes. Organizational size and maturity, sectoral risk, decision criticality, technological context, and regulatory environment are defined as boundary conditions for future empirical validation.
Multi-modal anti-spoofing aims to differentiate live users from spoofing attacks using multiple biometric modalities during model training. While existing anti-spoofing methods often incorporate just one biometric modality, the effectiveness of attacking two or more biometric traits remains questionable. In this work, we introduce the multi-modal anti-spoofing approach to detect spoofing attacks across face and fingerprint. Our framework is built around an Angular Margin Loss (ArcFace) that increases interclass separation without disrupting cross-modal alignment, which enables reliable spoof detection across both face and fingerprint biometric characteristics. Moreover, to enhance model generalization against unseen spoof attacks, we include three adversarial attacks (i.e., FGSM, PGD, DeepFool) to evaluate our system. Extensive experiments on multi-modal benchmarks show that the proposed method not only significantly outperforms previous anti-spoofing methods but also uniquely offers the ability to handle potential attack types.
Few-shot image classification suffers from severe data scarcity and unstable generalization. Existing data augmentation strategies still have three major limitations: pixel-level fusion strategies are incompatible with the support–query structure of episodic learning, category selection for cropping-based augmentation is overly simplistic, and most approaches rely on a single augmentation method, limiting robustness. To address these issues, this study proposes a deep mixed data augmentation framework that jointly enhances both the support set and the query set. The method first performs global pixel-level fusion to construct fused support and query sets. A Hopfield network then turns fused-support similarities into a pairing matrix H, which assigns a different-class gallery partner for query-side cropping–mixing. Finally, cropping–mixing produces an enhanced query set for model training. The framework is validated using ResNet18+BDC as the backbone. Experimental results on MiniImageNet demonstrate that the proposed method is competitive in few-shot classification, attaining a five-seed test mean of 73.25%/81.88% under 5-way 1-shot and 5-shot. A single complementary run on FC100 attains 66.63%/77.80% and is not a same-backbone ranking against heterogeneous published protocols.
Most studies on vehicular cybersecurity focus on classification accuracy but fail to evaluate the applicability of the models to the edge. This paper proposes a secure edge intelligence framework for V2X/IoV intrusion and misbehavior detection, called TrustEdge-V2X, which is deployment-aware. The framework combines dataset-role qualification, attack taxonomy, AI model benchmarking, feature-budget analysis, deployment ranking based on EdgeScore, offline risk-aware orchestration, external validation, robustness testing, explainable AI, repeated-run statistical analysis and ablation studies. Three datasets are assigned different experimental roles: VeReMi_NextGen is used for core V2X/VANET misbehavior detection, CICIoT2023 is used for supporting edge/IoT intrusion experiments and HCRL_CarHacking is used for external IoV/CAN validation. LightGBM outperformed all other AI models in EdgeScore (0.9383), F1-score (0.9878), MCC (0.9758), and inference latency (0.009419 ms per sample) across all eight AI models and six scenarios on VeReMi_NextGen for binary detection. In five dataset-task cases, the accuracy-best model was different from the EdgeScore-best model, which is the most important point to note: the best model in terms of accuracy is not necessarily the best model in terms of EdgeScore. Compact feature subsets were competitive, and robustness testing demonstrated an average F1 decrease of 0.1423 when tested under stress. The orchestration layer was found to be beneficial for the tasks, but it did not always perform better than the best fixed policy. As a whole, TrustEdge-V2X offers a systematic approach to the assessment and selection of vehicular cybersecurity models based on the operational and deployment conditions, not only on the classification accuracy.
Tibetan-to-English machine translation (MT) models frequently falter under extreme domain data scarcity, often producing translations that violate the distinctive agglutinative rules of Tibetan and suffer from domain-specific stylistic mismatches. To overcome these limitations, we propose Semantic-Syntax Prealignment (SSPA), an innovative corpus generation framework. SSPA constructs high-quality pseudo-parallel pairs by explicitly minimizing the deviation between the syntactic-semantic profiles of generated samples and professional reference texts. Specifically, source-target structural representations are standardized through length-unified truncation and terminology normalization, followed by a dual-domain alignment process that maximizes syntactic cosine similarity under rigorous structural constraints. We further augment these aligned frames via a cross-length dynamic filling mechanism, which is integrated with an Expectation-over-Transformation (EOT)-based style regularization mechanism specifically adapted for stylistic perturbations, to simulate authentic linguistic variations. Extensive evaluations on our newly constructed Tibetan Medicine-Tibetan English (TM-TE) dataset demonstrate that SSPA significantly outperforms existing competitive baselines. Notably, SSPA achieves a BLEU-4 score of 36.2 and improves long-sentence BLEU-4 by 16.8 points, with a parser-verified grammatical compliance rate of 96.2%. The framework exhibits remarkable cross-domain adaptability and stylistic consistency, offering a robust, versatile solution for low-resource Tibetan professional domain MT.
Serious games have demonstrated significant potential for supporting learning, cognitive stimulation, and skill development. However, most existing solutions rely predominantly on visual and auditory interaction, while the integration of real tactile experiences remains limited despite their potential to support richer multisensory interaction. Frutopia is a hybrid tangible serious game designed to integrate physical interaction with digital gameplay in order to support tangible interaction and tactile exploration. The game was developed following a structured process that combines the iPlus methodology for educational game design with the Scrum agile framework, enabling the systematic definition, implementation, and refinement of gameplay mechanics, tangible interaction, and usability-oriented features. The resulting system incorporates tangible user interaction through conductive physical objects with different textures connected via a Makey Makey interface, enabling players to control in-game actions through real tactile exploration. The game features progressive maze-based challenges inspired by Ecuadorian cultural environments and representative fruits from Ecuadorian regions, integrating multisensory feedback, gamification techniques, and embodied interaction principles to foster engagement and sensory exploration. The system was implemented using the Godot Engine and evaluated through functionality and usability assessments. Functional validation achieved a success rate of 94.74% across the defined test cases, demonstrating the technical stability of the proposed solution. Additionally, a usability evaluation involving 50 participants was conducted using the Serious Games Usability Evaluation Instrument (SGUEI). The assessment produced a final rating of 90.37%, reflecting favorable perceptions of the interaction quality and overall user experience. The results demonstrate the feasibility of integrating tangible interaction and multisensory feedback within serious game environments and suggest that hybrid tangible interfaces can enrich user engagement and interaction quality. This work contributes to the design and development of hybrid tangible serious games by presenting a structured development workflow and providing preliminary evidence of the technical feasibility and usability of tangible interaction in serious game environments. The proposed system establishes a foundation for future studies involving the intended target population and the evaluation of educational and cognitive outcomes.
Graph neural networks remain challenging to apply to traffic forecasting, often failing to consistently outperform simple baselines. We observe that the strongest baseline is regime-dependent: last-value persistence dominates at short horizons on smooth freeway data, while time-of-day climatology prevails at longer horizons and on bursty arterial networks. Rather than treating this as a limitation, we propose a learnable approach that automatically selects the optimal baseline. Our method anchors predictions to a learned, per-horizon convex blend of persistence and climatology, introducing only twelve scalar parameters. This learned anchor recovers whichever baseline is locally most effective, allowing the model to focus on capturing residual variations that neither baseline captures. We evaluate on six public benchmarks (METR-LA, PEMS-BAY, PEMS03/04/07/08) spanning traffic speed and flow data under standard 70/10/20 chronological splits with masked evaluation metrics and holiday-aware climatology. Our anchored temporal models consistently beat both baseline methods on the 12-step average across all datasets, and outperform at every horizon on five of the six benchmarks. When integrated into two strong architectures (STID and Graph WaveNet), the anchor yields substantial gains at long horizons where climatology is most informative. Notably, within our lightweight framework, learned spatial graph components do not improve accuracy and can slightly degrade performance, a finding we analyze and discuss.
Poisoning attacks present a significant challenge for federated learning (FL), particularly when clients operate with Non-IID data and updates are transmitted through multi-tier aggregators. This paper introduces FLVaccin, a hierarchical personalized FL framework structured as an unbalanced tree. In this framework, each non-root node both hosts local clients and aggregates shared MobileNetV2 features (FedPer), whereas the root node does not possess local data. The proposed defense integrates node-level CIFAR-100 vaccination, which calibrates depth- and round-adaptive tolerances, with per-client trend quarantine and root backbone rejection. Experimental results on CIFAR-10 with 100 clients, 25 aggregators, and Dirichlet Non-IID partitioning (α=0.5) demonstrate that the clean baseline achieves 79.9% accuracy. In contrast, unconstrained mixed attacks reduce performance to near-chance levels (20.2% k-fold). When vaccination and quarantine are enabled, 535 attack events still result in a 76.5% ± 0.4% k-fold accuracy (77.3% test), remaining within 2.6 percentage points of the clean model. These findings indicate that tree-aware, vaccination-calibrated monitoring can maintain model utility under persistent multi-vector poisoning without the need to share raw data.
In this research, a new approach is proposed to accurately predict indoor three-dimensional localization based on Received Signal Strength (RSS) values. This work uses WiFi beacons to collect time-series RSS data, preprocess it, and feed it to the proposed model. The proposed model presents a novel architecture based on 2D convolutional neural networks, and this model employs a multitask learning approach. Hence, the model simultaneously has a classifier for floor classification and a regressor for estimating X and Y coordinates, and tries to perform accurate indoor localization even in environments with furniture and other obstacles. The proposed CNN-based model efficiently utilizes RSSI data, achieving 99.0% floor classification accuracy and 6.5 m in terms of Euclidean distance error based on coordinate estimation on the UJIIndoorLoc dataset. On the other hand, the validation results using Tampere datasets were ‘distance error (m) = 3.7868’ form localization and ‘accuracy = 98.99%’ for floor classification. Comprehensive preprocessing significantly enhances localization accuracy.
In the field of 3D reconstruction for photovoltaic scenes, current texture mapping techniques frequently encounter significant texture segmentation and apparent joins in uniform plane regions, such as solar panels, because they lack geometric structural assumptions. To tackle these challenges, we introduce a new texture-mapping strategy for 3D solar panel scene reconstruction that focuses on planar structure awareness. We term the proposed method PSATM. Initially, we suggest a global constraint and a local refinement process to incorporate clear geometric structure details. This process automatically detects and labels planar regions through a region-growing approach. Next, we integrate a planar structure-aware module into the smoothness term of the Markov Random Field (MRF) energy function. This module uses dihedral angles and plane membership to adjust label transition costs, enhancing texture coherence within planar regions and maintaining smooth transitions at genuine geometric breaks. Furthermore, we establish a boundary treatment technique relying on local geometric support. This method combines area-based weighting and normal consistency to modify erroneous labels, successfully removing small remnants and smoothing texture edges. We tested the proposed PSATM with texture patch counts and visual quality measures on actual solar panel scenes. The results indicate that the proposed PSATM considerably reduces texture segmentation errors and improves texture flow and overall visual quality compared to the existing method.
The deployment of deep-learning vulnerability detectors in production remains difficult. Models are large, false-positive rates are high, output is opaque, and a persistent gap separates benchmark performance from real-world utility. The objective of this work is to close part of that gap by combining a learned detector with interpretable rules so that accuracy, efficiency, and actionability are achieved together. A hybrid framework is therefore presented in which a lightweight edge-conditioned GNN of 71,810 parameters, named FastVulnGNN, trained in 96.2 s on a single CPU core, is paired with rule-based heuristic detection for six C/C++ vulnerability classes, namely buffer overflows, format-string defects, null-pointer dereferences, double-free errors, integer overflows, and race conditions. On the MegaVul dataset, an accuracy of 71.1%, an F1 score of 0.70, and an AUC-ROC of 0.77 are obtained by the GNN component. On a production codebase of 499 files and 312,758 lines of code, the full hybrid scan completes in 5.5 s, which corresponds to about 57,000 lines per second, without any GPU hardware. Per-file risk tiers and pattern-level explanations are produced, and these are suitable for continuous-integration use. The significance of this work lies in demonstrating that a deployable, explainable detector can be assembled from compact components, and an edge-type ablation study, a cross-dataset evaluation, and a per-vulnerability analysis are reported to characterize the approach.
In Industrial Internet of Things (IIoT) deployments, mobile edge computing (MEC) offloads computation-intensive tasks, a constrained biobjective problem trading time delay against energy consumption (MTOP). We show that this benchmark is exactly separable across micro base-station (MiBS) regions: its delay and energy objectives are additive over regions, and the only coupling, intra-cell interference, stays within a region. Exploiting this, we propose CR-MTMEMTO-D, a structure-aware decomposition multitasking method that treats each region as an independent subtask, solves it with a feasibility-repaired NSGA-II, and reconstructs the global feasible Pareto front as the non-dominated subset of the Minkowski sum of the regional fronts, an exact composition that adds no global evaluations. Across 12 instances (45–432 variables, 20 seeds), it attains the best hypervolume and IGD on every instance (mean HV 0.9340 vs. 0.8021 for a plain NSGA-II baseline; average rank 1.00), with the margin widening as the problem scales, and it is unchanged under total-evaluation matching because every evaluation is a regional main task. A feasibility-priority acceptance gate keeps the population fully feasible. Under matched budgets, a prior cheap-task pool with bandit-controlled transfer adds no significant gain, which motivates the structural approach.
Mobile-network operators must interpret spatial anomalies in aggregated cell-level telemetry and decide whether, where, and how to respond. This paper presents a privacy-conscious, intrusion detection system (IDS)-oriented triage and response architecture that consumes cell-level anomaly signatures and couples spatial reconstruction, short-horizon forecasting, origin inference, self-resolution and remaining-time estimation, adaptive gating, ETA-aware team selection, conservative redeployment, explanation, and audit logging. It is a downstream spatial-attribution and response-orchestration layer, not a packet- or flow-level attack detector. The evaluated configuration uses transparent deterministic, heuristic, and optimization-based procedures and synthetic aggregated signatures without subscriber identifiers; aggregation is treated as data minimization, not a formal privacy guarantee. Across 20 paired synthetic scenarios, the full policy reduced conditional mean response time from 37.58 to 22.86 min, total travel from 576.0 to 273.5 min, and coverage ETA from 32.28 to 26.76 min, while on-time service increased from 54.0% to 60.0%. These benefits were accompanied by lower persistent-incident coverage (91.1% to 72.1%) and a higher miss rate (8.9% to 27.9%). The inverse-origin configuration showed no repeated localization-error advantage, and conservative redeployment had only a marginal average effect. The results therefore demonstrate a configurable downstream triage trade-off under controlled synthetic conditions, not attack-classification accuracy, adversarial robustness, formal privacy, or deployment readiness.
Patent document–keyword count data are typically high-dimensional, sparse, and dominated by zero entries, which makes it difficult to simultaneously reconstruct keyword frequencies and identify meaningful technological structures. This study proposes a machine learning approach to latent structure learning for zero-inflated patent keyword count data. The proposed zero-gated latent factor model (ZG-LFM) combines nonnegative matrix factorization (NMF) with keyword-specific logistic occurrence models. NMF is used to extract interpretable document–factor and factor–keyword representations, while the occurrence gate estimates the probability that each keyword appears in a given patent document. The method was evaluated in an initial domain-specific case study using a document–keyword matrix constructed from 9434 quantum computing patent documents and 175 keywords, of which 87.60% of the entries were zero. Predictive performance was assessed using root mean squared error, mean absolute error, and the area under the receiver operating characteristic curve across different numbers of latent factors. The experimental results showed that NMF provided more accurate keyword count reconstruction, whereas the proposed model consistently achieved better discrimination between zero and nonzero keyword entries. These findings indicate that latent count reconstruction and keyword occurrence modeling provide complementary information for analyzing sparse patent data. The learned latent factors further revealed coherent quantum computing subdomains, including hybrid quantum–classical execution, quantum machine learning, quantum state measurement and error analysis, quantum cryptography, superconducting chips, quantum circuits, optical control, qubit devices, and optimization algorithms. The proposed framework therefore provides interpretable latent technology structures while improving the identification of keyword occurrence patterns in zero-inflated patent data. These findings demonstrate the feasibility of the framework within the analyzed quantum computing corpus; its generalizability across other technological domains remains to be evaluated.
Lissajous figures are frequently studied and widely used objects in engineering and physics. Although these patterns are usually analysed using computer simulations or oscilloscopes, such tools may limit their educational value by covering the core physical processes that generate the curves. In order to address these problems, the design, kinematic validation, and prototyping of a dual-axis drawing robot were carried out on the LEGO Education SPIKE Prime platform. The hardware implementation centres on a LEGO-based dual Scotch yoke mechanism, which supports precise transformation of uniform circular motion into simple harmonic motion. This setup implements the superposition of two independent simple harmonic oscillations by simultaneously moving the paper tray along the x-axis and the pen along the y-axis. High-fidelity trajectories are achieved through a 40:1 worm gear reduction, which enables precise control of the parameter configuration. The phase shift can be manually set by adjustment levers. The robot’s geometry supports discrete amplitude settings of 8, 16, and 24 mm by adjusting the crankpin position. System control is managed by Python code that synchronises motor speeds and angular displacements according to frequency ratios. The research methodology used the Double Diamond design thinking framework, structuring development into four phases: identifying historical mechanical solutions, defining pedagogical and technical classroom requirements, iteratively developing the LEGO prototype, and testing the system through representative drawing experiments. Results show that the robot can reproduce a broad range of periodic Lissajous curves with high repeatability, and that its physical outputs show strong visual and mathematical correspondence to ideal trajectories simulated in the Desmos graphing calculator. The final prototype satisfies classroom constraints, providing a transparent, low-cost, modular STEAM tool that bridges the distance between abstract parametric equations and complex mechanical implementations.
This study proposes a hybrid computational model for diagnosing such biases using groundwater observation data across Kazakhstan. The analytical dataset included 2402 georeferenced observations, including 492 natural springs from OpenStreetMap (OSM), 109 boreholes from OSM, and 1801 spatially filtered pseudo-absence observations. Springs and boreholes together formed 601 positive groundwater observations, while pseudo-absence samples represented a spatially filtered background level rather than confirmed groundwater absence. Each observation was characterized by 89 environmental predictors extracted from Google Earth Engine. The proposed hybrid observation source bias index (HOSBI) combines a normalized robust effect size based on the median absolute value of the Cliff delta, multivariate distribution divergence quantified using RBF-MMD, and spatially confirmed source distinctiveness. These components were assigned fixed weights of 0.40, 0.35, and 0.25 to emphasize statistical and distributional data while maintaining spatial validation. Spatial cross-validation achieved a balanced accuracy of 0.855 for distinguishing OSM sources from OSM wells and 0.846 for separating positive observations from background pseudo-absences. Climate showed the strongest source-related bias (HOSBI = 0.923), while Sentinel-1 SAR contributed the most to the contrast between positive and background data (HOSBI = 0.923). The proposed framework provides an interpretable and replicable preliminary assessment of source bias in heterogeneous geospatial datasets.
