The first generation of post-quantum cryptographic standards from the U.S. National Institute of Standards and Technology (NIST) rests on one mathematical foundation. Three of the four schemes standardized or in development, ML-KEM, ML-DSA, and FN-DSA, rest on the hardness of finding short vectors in structured lattices, while only SLH-DSA departs from it; the first two use Module Learning With Errors (Module-LWE) and the third the related but distinct NTRU assumption. This paper examines that concentration as a systemic risk. Drawing on the computer-security literature on software monoculture and design diversity, it argues by analogy that algorithmic homogeneity creates conditions for correlated failure of much deployed public-key cryptography. It makes the mechanism concrete: the security of all three schemes is estimated under one lattice-reduction cost model, so a sufficiently general advance in lattice reduction could move their security margins together while the non-lattice options stay untouched, as a cross-family correlation matrix and a sensitivity analysis show. It then separates three hazards often conflated in the debate, transition risk, assumption-monoculture risk, and implementation risk, and shows that the classical-plus-lattice hybrids now carrying most post-quantum Web traffic hedge the first but not the second. The claim is one of risk management, not mathematical certainty: no standardized lattice scheme is broken at its parameters as of mid-2026, and Module-LWE enjoys strong worst-case-to-average-case reductions. The case for diversification rests on the irreducible uncertainty of cryptanalysis, the fragility of implementation security, and the severe, potentially unrecoverable cost of a correlated failure.
Large-scale record linkage in multi-source microdata environments requires infrastructures capable of preserving entity identity, provenance, versioning, and governance over time. This paper presents iDatos, a metadata-driven production infrastructure developed by the Canary Islands Institute of Statistics (ISTAC) for the systematic linkage of heterogeneous administrative, census, and survey microdata under strict privacy and governance constraints.The core of iDatos is a relational linkage layer based on persistent identifiers, identifying forms, provenance relations, and harmonised common variables. Rather than proposing a new record linkage algorithm, the paper formalises how deterministic rules, similarity-based techniques, and complementary linkage strategies are integrated within a metadata-governed institutional workflow. The resulting structure is interpreted retrospectively through the Data Vault paradigm as a conceptual lens for modularity, historisation, separation of concerns, and auditability.The paper documents the operational deployment of iDatos in three master registers—persons, enterprises, and portals. By the evaluation cut-off, 35 distinct sources had been incorporated, with 1685 distinct files and 382.1 million rows associated with those sources. Across the three register-specific linkage environments, 678.6 million linked-record relations connect 5.67 million representative entities. System-level indicators describe register composition, linkage intensity, identification variability, linkage-outcome distributions, internal confidence categories, and processing cycles.These results document deployment at substantial institutional scale, but are not presented as controlled benchmarks of runtime scalability or externally validated linkage accuracy. The contribution is the formalisation and operational documentation of a governed infrastructure for provenance-aware record linkage and master data integration in official statistics.
As cyber and physical anomalies escalate, researchers seek efficient detection and mitigation methods. Traditional physical inspections are costly and inefficient, but advanced metering infrastructure (AMI) with smart meters offers a promising alternative for detecting various anomalies. In this study, a hybrid deep learning framework known as the Convolutional Neural Network -Bidirectional Long Short-Term Memory (CNN-BiLSTM) is used with hyperparameter optimization. Six distinct types of attacks have been artificially generated, and data preprocessing techniques such as outlier removal, data standardization, and interpolation are performed before incorporating SCADA data. To ensure the model’s robustness, five-fold cross-validation, dropout, and L2 regularization have been implemented. Furthermore, a comparative analysis was conducted using three distinct loss functions to evaluate their performance against the proposed model. Performance evaluation metrics, including accuracy, recall, precision and F1-score, demonstrate that the proposed framework outperforms others. Additionally, a sensitivity analysis shows the effectiveness of the model in handling variations of input data.
High-quality image rotation traditionally forces a compromise between efficient but lossy local FIR interpolators and high-fidelity global recursive IIR solvers. While recursive methods achieve superior spectral preservation, their causal recursive data dependency creates a sequential bottleneck that disrupts the parallel execution pipeline of modern SIMD vector units.In this work, we propose Polyphase Hybridization (PH), an engineering framework that bridges this performance-fidelity gap by translating infinite-support B-spline interpolation into a localized, SIMD-friendly gather operation. By aligning the rotation lattice with exact rational coprime geometries (Farey anchors), PH structurally stabilizes the sub-pixel phase density. This enables an offline continuous synthesis stage to execute precise, energy-preserving truncations of high-order cardinal splines. Consequently, the online execution is reduced to a strictly bounded, single-pass non-recursive pipeline operating entirely from L1-resident phase dictionaries.Extensive validations across standard benchmarks confirm that our formulation successfully bypasses the computational bottlenecks of classical B-spline interpolation. By strictly linking execution latency to the bounded spatial kernel size D rather than the mathematical complexity of the underlying interpolation order β, our architecture significantly reduces runtime while fully capturing the high-frequency structural energy of high-order splines. Empirical results demonstrate that our framework successfully emulates a range of generalized interpolators: from the mid-tier PH-HQ matching cubic recursive solvers, to the Super High-Quality (PH-SHQ) mode which captures the structural fidelity of quintic recursive solvers using a folded septic kernel, delivering equivalent reconstruction fidelity in ≈1/30 of the average execution time.
Transportation networks increasingly face disruptions that require rapid interpretation as well as coordinated operational response. This study develops and computationally evaluates an interdisciplinary framework integrating Large Language Models (LLMs), Multi-Agent Systems (MAS), and digital twins in port-centred multimodal container logistics. The framework separates contextual interpretation, operational validation, decentralized execution, and feedback-based adaptation. A controlled simulation compares four architectures: rule-based control, MAS-only coordination, classical machine learning (ML) with MAS, and LLM-supported interpretation with validation and MAS, across five disruption conditions and 30 matched replications. MAS coordination substantially improves delivery performance relative to rule-based control. LLM-supported interpretation achieves higher interpretation accuracy and fewer validation rejections than classical ML and reduces response time, but does not improve downstream delivery or recovery performance relative to classical ML+MAS. The findings show that improved semantic interpretation does not automatically translate into greater system-level agility when downstream coordination and fallback mechanisms compensate for interpretation errors. The study therefore positions LLM value as conditional rather than universal and identifies validation as a boundary mechanism separating probabilistic interpretation from operational execution.
Numerical integration is a fundamental component of scientific computing, yet the relative performance of integration methods depends strongly on the structural characteristics of the integrand and the computational cost associated with each solver. This work proposes an intelligent framework for adaptive numerical integration that formulates method selection as a supervised algorithm-selection problem. The main novelty lies in combining interpretable integrand descriptors with a data-driven selector to choose among heterogeneous classical and learning-based integration methods under a unified accuracy–cost criterion, rather than using machine learning as a direct replacement for established numerical algorithms. The framework evaluates Simpson, Gauss–Legendre, and Romberg quadrature together with neural-network and physics-informed neural-network integrators. Experiments on 600 benchmark instances spanning six function families showed that Romberg and Simpson provided the most favorable accuracy–cost trade-offs, being selected in 70.83% and 29.17% of the instances, respectively, whereas the evaluated learning-based integrators were not selected as optimal under the adopted per-instance training protocol. The Random Forest selector achieved 83.89% accuracy and a weighted F1-score of 0.83 on the held-out test set. Ablation analysis further showed that predictive performance benefits from combining statistical and structural descriptors of the integrand, particularly information related to local variation and slope. The proposed formulation can be extended to broader portfolios of numerical methods and problem classes.
Autonomous robots performing exploration, inspection, and mapping tasks collect large volumes of sensor data for computationally intensive downstream perception tasks, such as object detection, optical character recognition (OCR), and semantic segmentation. In many deployments, these tasks are deferred to offline or cloud-side processing due to limited onboard computational resources. However, existing systems typically continue data acquisition without assessing whether degraded observations caused by blur, noise, motion artifacts, or adverse lighting conditions are likely to lead to offline perception failures. As a result, failures are often detected only after mission completion, causing unnecessary storage, computation, and costly data reacquisition. To address this limitation, we propose a lightweight success prediction model (SPM) that estimates, during data acquisition, the probability that a downstream perception task will successfully process an observation. The predicted success probability enables online edge-side decisions, including selective data retention, adaptive sensing, and resource-aware scheduling of computationally intensive perception modules. Evaluated across OCR and barcode detection tasks using public benchmark datasets and real-world robotic data, the proposed method achieves over 0.94 ROC-AUC and average precision on in-domain benchmarks and strong performance under real-world warehouse conditions, while requiring up to 236× lower inference latency than the downstream perception algorithms. Selective acquisition experiments further demonstrate substantial reductions in storage and offline processing while preserving the majority of successful detections, highlighting success prediction as an effective optimization strategy for resource-constrained Edge AI and robotic perception systems.
Electronic prescription systems support efficient and auditable medication dispensing. During prescription verification, pharmacies need to establish prescriber authorization, legitimate credential possession, and compliance with validity constraints while limiting unnecessary disclosure of sensitive medical information. This creates a fundamental tension between verification requirements and data minimization. Conventional approaches typically rely on full-record presentation and public-key-based signatures bound to persistent identifiers. Although these mechanisms support authenticity and accountability, they can disclose more information than is required and enable cross-transaction linkage of patients and prescribers. Moreover, prescription-validity checks can depend on protected attributes, such as the expiry date, that should remain hidden during verification. This paper presents a privacy-preserving prescription verification framework that supports anonymous-yet-accountable healthcare-professional authorization, selective disclosure and an issuer-authenticated holder public-key reference through BBS+ prescription credentials, and privacy-preserving verification of a hidden expiry attribute. The framework defines a prescription presentation protocol through which a patient submits a BBS+ selective-disclosure presentation proof, an expiry commitment, and two Bulletproof range proofs to a pharmacy. To address the consistency gap that can arise from a naïve composition of credential and validity proofs, the protocol connects these components through a common expiry commitment. The canonical hash of this commitment is authenticated as an issuer-authenticated BBS+ credential attribute and treated as disclosed during presentation, while both range-proof commitments are derived from the same presented expiry commitment. The pharmacy independently verifies the commitment hash before verifying the range proofs. Under the stated assumptions, this construction mitigates cross-component substitution attacks. The framework is analyzed under an explicit threat model and trust assumptions, and its protocol-level message flow is examined through symbolic verification. A proof-of-concept implementation evaluates the computational and communication overhead of the cryptographic operations under synthetic workloads. The results quantify the cryptographic processing and presentation payload overhead across the evaluated configurations within the stated experimental scope.
Ensuring robust security in resource-constrained Internet of Things (IoT) and Cyber-Physical Systems (CPS) requires reliable sources of randomness for cryptographic key generation. This study examines the internal structure of entropy generated from multisensor data collected by a commercial smartwatch in static (still) and dynamic (shake) conditions. A dataset of 9696 binary sequences, each 11,000 bits long, was produced through randomized multisensor vectorization and segmented into local Shannon entropy features. Statistical analysis revealed a significant difference in entropy variability between regimes, with a high Cohen’s effect size (d=3.0). An independent two-sample statistical test additionally confirmed that the entropy distributions between the shake and still regimes differed significantly (p<0.001). The corresponding confidence intervals further supported the observed separation between operational conditions. Furthermore, unsupervised dimensionality reduction (PCA) and supervised machine learning models (Random Forest, Logistic Regression) successfully identified latent patterns within the entropy space, achieving a classification F1-score of 0.99. These findings lead to the concept of an “entropy fingerprint”, a context-dependent signature associated with the physical sensing process. The obtained results suggest that such entropy structures may provide additional contextual information applicable within decentralized identity (DID), wearable-assisted cryptographic workflows, and offline entropy harvesting architectures. This connection between physical-layer dynamics and entropy characterization establishes a foundation for future research into context-aware entropy analysis and decentralized entropy-acquisition methodologies in wearable IoT environments.
The rapid expansion of the Internet of Things (IoT) has revolutionized significant sectors such as energy, healthcare, smart cities, and security by enabling intelligent sensing, data exchange, and real-time automation. However, this rapid growth has brought two interconnected and critical challenges: ensuring system security and maintaining energy sustainability. As IoT devices are becoming increasingly more ubiquitous and resource- constrained, ensuring both low power consumption and strong robust protection against cyber threats is now a critical priority. This survey provides a thorough overview of energy-aware security mechanisms in sustainable IoT environments. It examines the dual challenges of optimizing energy efficiency while maintaining robust security, analyzing current energy-saving techniques such as energy harvesting from renewable sources including solar, mechanical, and radio frequency methods. Moreover, the study investigates how these energy-saving practices impact security by highlighting vulnerabilities arising from power constraints and evaluating potential solutions as lightweight cryptography and privacy-preserving approaches based on cutting edge technologies such as Artificial Intelligence (AI) and Blockchain. Finally, the survey outlines main open challenges and proposes future research directions aimed at developing scalable, integrated, and secure energy-efficient IoT architectures. The insights presented by this paper aim to guide the development of IoT infrastructures that are not only environmentally sustainable but also resilient and secure by design.
This study presents a scalable, data-centric predictive analytics pipeline for modeling healthcare performance across multiple hierarchical levels using large-scale administrative data. The framework addresses key challenges inherent in real-world healthcare datasets, including heterogeneity, feature redundancy, data sparsity and multisource integration. To preserve the administrative structure of the healthcare system, the framework introduces a hierarchy-aware feature representation that integrates district, facility, and workforce-level indicators with cross-level contextual features. To enhance predictive capability, we introduce a stability-based consensus feature-selection strategy that integrates tree-based ensemble methods to identify the most informative predictors. This approach is extended through stability-based consensus selection, in which feature importance and selection frequency are aggregated across XGBoost, Random Forest, and grouped validation folds. These features are subsequently leveraged within a blended machine learning architecture, where multiple base regressors are combined through a meta-learning layer to improve robustness and generalization. The proposed cross-level architecture integrates out-of-fold predictions generated from district, facility, and workforce-level models, while hierarchy-preserving validation prevents administrative units from being shared across training and testing partitions. The proposed framework enables multi-level analysis, supporting predictive modeling at district, facility, and workforce levels. Under five-fold district-grouped cross-validation, the proposed cross-level blended model achieved a mean RMSE of 0.574 ± 0.031, demonstrating stable generalization across previously unseen district groups. After model selection, the final model achieved an RMSE of 0.568 on the separate untouched district-grouped test set. The findings underscore the potential of integrating advanced data-driven methodologies into decision support systems for more effective monitoring and optimization of healthcare services.
Hand pointing is a fundamental nonverbal gesture used to establish joint attention and convey spatial intent. Despite its importance, existing recognition approaches often remain limited to binary detection or rely on specialized hardware, constraining their applicability in natural environments. This paper introduces a vision-language spatiotemporal transformer framework that integrates visual, textual, and temporal cues to estimate hand point direction. The approach leverages zero-shot diffusion segmentation to automatically generate upper-body masks without manual annotation, which are fused with input frames to enhance spatial focus on pointing regions. Spatial and temporal features are then encoded using transformer-based architectures to learn sequence-level visual features. Language-level learning is enabled by combining semantic textual embeddings from a vision–language model with visual features via a cross-attention mechanism. This synergistic design allows the model to jointly predict both pointing occurrence and direction vectors with high accuracy. Extensive experiments on benchmark datasets show that the proposed model consistently outperforms state-of-the-art baselines, achieving precision and recall above 90% and reducing mean angular error to less than 10°. Moreover, the framework enables semantic prompt-guided inference on previously unseen video samples using text-encoded prompts, making it highly valuable for practical applications where labeled data or semantic context is unavailable.
Severe construction accidents continue to impose substantial human and economic costs, yet most existing severity prediction models rely primarily on structured variables and overlook the contextual information embedded in narrative reports. This study addresses this limitation by proposing an explainable multimodal machine learning framework that integrates structured injury attributes with enriched narrative-based features to predict post-accident hospitalization outcomes. The framework was developed using 17,270 construction-related incident reports submitted to the U.S. Occupational Safety and Health Administration (OSHA) between 2015 and 2024. Cases were classified as hospitalized or non-hospitalized, where the latter outcomes that did not involve hospital admission. Accident narratives were processed through a text mining and sentiment analysis pipeline, where sentiment reflects the severity and intensity of descriptive language rather than emotional tone. These text-derived features were combined with structured variables to form a multimodal feature space. Class imbalance was addressed using SMOTETomek, and model hyperparameters were optimized through Bayesian Optimization. Three ensemble classifiers—Random Forest, Gradient Boosting Machine, and eXtreme Gradient Boosting—were evaluated against a logistic regression baseline. Among the models tested, Random Forest achieved the strongest performance, with a weighted F1-score of 0.9738 and an AUC of 0.9769. Model interpretability was ensured using SHapley Additive exPlanations (SHAP), which linked predictions to key injury-related factors, including Nature of Injury, Part of Body, Event Type, and narrative sentiment. Although the framework is trained on post-incident data, the extracted patterns and interpretable feature relationships enable proactive risk reduction by identifying recurring high-risk mechanisms, work conditions, and body regions that can be targeted through preventive controls, training, and safety planning. This study extends prior OSHA-based multimodal research by systematically refining narrative features and embedding explainability to support transparent, data-driven construction safety management.
This study proposes an Adaptive Threshold Multiple Correspondence Analysis (AT-MCA) framework for analyzing data-driven categorical handwriting features in dysgraphia pattern analysis. Handwriting images from the Potential Dysgraphia Handwriting Dataset of School-Age Children were processed using Otsu thresholding, connected-component labeling, and noise removal to extract quantitative features related to letter size, consistency, letter spacing, word spacing, and stroke density. These measurements were converted into categorical variables using data-driven criteria based on their empirical distributions. AT-MCA extends Classical MCA by applying an information-based nonlinear transformation to the matrix of proportions derived from the Burt matrix, followed by empirical thresholding and matrix rescaling before decomposition. Both methods used the same active categorical representation, while the expert-provided classification was retained as an external supplementary variable. Sensitivity analysis was conducted for the transformation parameter over candidate values from 0.40 to 1.00. The first two dimensions accounted for 60.06% to 75.69% of the total inertia across the tested values. Based on the stability assessment, was selected as the final parameter, yielding 73.37% cumulative inertia in the first two dimensions, compared with 56.75% for Classical MCA. The resulting category coordinates were used to examine associations among the extracted handwriting characteristics and their relationship with the expert-provided classification. These findings indicate that the proposed adaptive transformation and thresholding substantially modify the inertia distribution and increase the amount of categorical association represented in the first two dimensions relative to Classical MCA.
Occupational safety monitoring in resource-constrained factories remains dominated by manual and reactive inspection. This study presents a simulation-based proof-of-concept evaluation of a context-aware safety decision architecture that integrates three sensing modalities - environmental measurements, proposed image-based PPE detection, and RFID zone tracking - together with incident-history context, an explicit rule table, and a designed human-in-the-loop (HITL) review layer. The evaluation uses 300 simulated observations. Because the target classes were generated by the same expert-defined rule table that maps sensor risk, PPE status, and zone level to No Alert, Warning, or High Risk, the supervised classifiers approximate that policy rather than predict independently observed accidents. Six classifiers were compared using stratified cross-validation and a held-out test partition. At the original seed-42 split, Random Forest and XGBoost produced identical held-out results (accuracy and weighted F1 = 0.833; macro F1 = 0.829; balanced accuracy = 0.800; Cohen's kappa = 0.718; High-Risk recall = 0.700). A supplementary reconstructed fixed-dataset analysis across seeds 0, 1, 42, 123, and 999 yielded held-out weighted F1 of 0.836 +/- 0.022 for Random Forest and 0.859 +/- 0.027 for XGBoost. Across five seed-level cross-validation means, the Full and Sensor+Zone configurations differed by only 0.003 +/- 0.012 weighted F1, with no significant difference detected in an exploratory paired test (W = 5.0, p = 0.625). The proposed YOLOv8 and HITL components were not empirically validated. The results therefore demonstrate policy-approximation feasibility under simulation and identify sensor and zone information as the dominant contributors, while real-factory validation remains essential.
While speech spoofing detection techniques based on deep learning have performed well in the recent years, many deep learning-based speech spoofing detection methods based on convolutional neural networks (CNN), Recurrent Neural Networks (RNN), and Transformer based architectures have difficulty in effectively representing discriminative local acoustic features and long-range dependencies, which reduces their resistance to the sophisticated attacks of synthesized and voice converted speech. To overcome these disadvantages, we introduce a new detector for automatic speech spoofing detection, which is called Transformer Encoder and Ensemble Learning-based Detector (TEEL-Det). The proposed framework combines the ensemble learning, Mel-Frequency Cepstral Coefficients (MFCCs), and Transformer Encoder to improve the representation of features and contextual modeling. The ensemble learning module extracts the complementary discriminative features and the MFCCs give compact and informative acoustic speech representations from raw speech. The Transformer Encoder uses a self-attention mechanism to learn the characteristics of the speech that are of interest for accurate classification of bonafide and spoofed speech while maintaining long range dependencies. The proposed model was tested on a subset of the Logical Access (LA) category of the ASVspoof 2019 challenge, which includes synthesized speech and voice converted speech. Through experimental results, the effectiveness and robustness of TEEL-Det are clearly shown in speech spoofing detection, and its performance is better than that of other methods, with an Equal Error Rate (EER) of 0.003% for the LA evaluation set.
Accurate identification of sitting postures is essential for ergonomic assessment and preventive healthcare. This study presents a fully non-intrusive posture classification approach based solely on an industrial pressure sensor mat (PSM) placed on the seat. We analyse the effectiveness of well-designed features extracted from pressure images, incorporating spatial distribution, image moments, and regional descriptors, and compare them with raw data representations. A major limitation in the current literature is the absence of standardized benchmarking and public datasets. To address this gap, we collected two datasets in different environments, using different participants and chair types, and we evaluated performance through two complementary scenarios: 5-fold cross-validation within a single dataset and testing on completely unseen data. Classical machine learning models (SVM, K-NN, and Random Forest) demonstrate that the proposed features reduce input dimensionality while preserving or improving classification accuracy (up to 86% accuracy), particularly in cross-dataset generalisation (63% accuracy rate). We proposed a set of features designed to be independent of the sensing hardware, chair type, and participant group used during training, together with a benchmarking protocol for evaluating model generalization.
Objective: Robot task programming in manufacturing remains tightly coupled to vendor-specific languages, hindering portability and reuse. This paper presents a vendor-neutral, schema-validated format for robot task specification that enables large language model (LLM) generation and multi-robot execution without modifying the task description. Methods: The Robot Task Specification Format (RTSF) is a JSON format with nine step types covering all six control-flow constructs, governed by a JSON Schema (draft-07). A gap analysis against ten existing formats positions RTSF. Schema-guided LLM generation is evaluated on 30 task descriptions across four domains, with a prompt ablation study. A runtime architecture (recursive parser, PDDL converter, BehaviorTree.CPP converter) enables multi-backend execution. A multi-robot experiment validates six scenarios on four platforms: UR5e, Franka FR3, Kinova Gen3, and Denso Cobotta (physical robot). Results: No surveyed format combines a task-level abstraction with a formal JSON Schema, vendor-independent execution, complete control-flow coverage, and schema-guided LLM generation. The multi-robot experiment verifies identical action sequences and arguments across over 500 adapter calls on four platforms using unmodified task files. Schema-guided prompting achieves high construct recall (100% corrected validity); an independent benchmark by three external evaluators confirms generalisability. Prompt ablation confirms that disambiguation and field-name guidance rules contribute roughly equally to recall. Scalability benchmark confirms sub-second processing at 5000 steps. Conclusion: RTSF fills the gap between natural-language task descriptions and heterogeneous execution backends, enabling vendor-independent robot task programming with schema-guided LLM generation and multi-platform validation. The format, examples, scripts, and test suite (107 tests) are released under the MIT licence.
Online recruitment platforms are increasingly exploited by fraudulent job postings that harvest applicants’ personal data, extract upfront fees, or facilitate identity theft. Prior machine learning approaches report strong benchmark performance but are typically single-model classifiers that overlook the operational requirements of real deployment: cost-aware inference, calibrated probability output, human-in-the-loop routing, and structured explainability. This paper addresses these gaps with a two-stage, cost-aware fraud-screening framework evaluated on the Employment Scam Aegean Dataset (EMSCAD). A calibrated Hybrid Support Vector Machine (SVM) operating on TF-IDF and structured metadata features resolves the large majority of postings at negligible computational cost, escalating only the most uncertain cases to a DistilBERT-based hybrid classifier with a learned softmax attention gate for deeper semantic analysis. A seven-configuration ablation study shows free-text content is the dominant fraud signal, while structured metadata contributes a smaller but consistent precision gain. On a corrected, deduplicated evaluation split, the two-stage pipeline achieves fraud-class F1 of 0.906 (macro-F1 0.951; three-seed mean fraud-F1 0.899 ± 0.007) while reducing per-prediction inference cost by approximately 98.28% (dollar terms; three further measures – GPU-hours, FLOPs, and energy – are mathematically linked rescalings of the same escalation rate rather than independent evidence, while measured latency provides a genuinely distinct empirical confirmation). This advantage over the Stage-1 model alone is statistically robust across three independent training runs (McNemar’s exact test, p<0.03 in every run), though a threshold-confound ablation shows that only about 40% of this advantage is attributable to Stage 2 specifically, with the remaining 60% attributable to re-thresholding Stage 1’s own probabilities using the same zone boundaries; a McNemar test isolating Stage 2’s marginal contribution specifically is significant in only 1 of 3 seeds (p=0.043, 0.093, 0.210), so this marginal share is numerically positive but not robustly statistically significant on its own – Stage 2 still contributes a genuine, positive, but more modest and less statistically consistent share than the headline comparison alone suggests. An extended architecture search – comparing SVM, LightGBM, CatBoost, and RoBERTa as alternative components, and multi-task learning as an alternative training objective – confirms the adopted design remains the most cost-effective choice, with most architecture-selection decisions validated on held-out data rather than the test set. A manual review of every false negative, cross-dataset validation on the independently curated DIFrauD corpus, and token-level attribution via Integrated Gradients are reported, together with a corrected mechanistic-interpretability finding and a significant, previously unquantified operational limitation: the majority of residual false negatives (61.9%) are never seen by Stage 2 at all, since Stage 1 misclassifies them with enough confidence to auto-approve them directly – the paper’s largest unresolved operational risk, for which our exploratory mitigations remain inconclusive. The framework outputs calibrated risk scores, term-level explanations, and three-zone routing decisions designed for direct integration into human review workflows.
Small object detection in remote sensing imagery remains challenging due to extremely limited object pixels, dense target distributions, and strong structured background clutter (e.g., roads, rooftops, and shadows). Moreover, many practical airborne and spaceborne platforms impose strict constraints on computation and power consumption, making it difficult to improve detection accuracy without increasing model complexity.To address these challenges, we propose WFFM-TA-YOLO, a lightweight one-stage detector built upon a YOLO-style baseline and tailored for remote sensing small-object detection. The core contribution is a Weighted Feature Fusion Module (WFFM) that integrates a P2-aware multi-branch fusion structure with adaptive scale-wise weighting. By preserving high-resolution P2 features and learning data-driven fusion weights from global feature descriptors, WFFM enhances fine-grained spatial cues while adaptively balancing shallow detail and deep semantic context across aerial scenes. A lightweight two-branch Triplet Attention module is incorporated at the shallow fusion output to suppress repetitive and directional background textures with negligible computational overhead.Extensive experiments on the VEDAI and AI-TOD datasets demonstrate that WFFM-TA-YOLO consistently outperforms the FFCA-YOLO baseline in terms of mAP@0.5, mAP@0.5:0.95, and small-object metrics. Ablation studies confirm that the WFFM and shallow attention components contribute complementary performance gains while maintaining a lightweight and deployable architecture.
