Search Results (3,913)

AI-driven credit scoring is supervised as a high-risk application in banking and insurance, yet unfairness is rarely operationalized as a measurable category of model, conduct, legal, and reputational risk. Using 20,000 anonymized applications from a Southern European digital lender (15.2% twelve-month default rate), we estimate three model families—a regularized logistic regression, a gradient-boosting machine, and a multi-layer perceptron—under a fully crossed design in which each family is evaluated without mitigation and under pre-processing (reweighing), in-processing (an exponentiated-gradient reduction, applicable to any base learner, together with adversarial debiasing where gradient-based training permits it), and post-processing (reject-option) interventions, so that the mitigation effect is no longer confounded with the choice of estimator. No sensitive-group field enters any estimated specification; group membership is used exclusively for auditing. Predictive performance (AUC-ROC, Brier score and Brier skill score relative to the base-rate forecast, F1 on the default class, Gini, and the Kolmogorov–Smirnov statistic) is reported jointly with group fairness (demographic-parity and equal-opportunity differences, disparate-impact ratio, Theil index) and with group-conditional calibration, at an explicitly stated and economically justified decision threshold. Every fairness quantity is accompanied by stratified-bootstrap confidence intervals and, for stochastic learners, by seed-level dispersion. The interpretable benchmark attains an AUC of 0.780 and a Brier score of 0.104 against 0.129 for the constant base-rate forecast, and the high-capacity models improve on it by under one AUC point. Disparity is present but is located geographically rather than in the composite group label: the disparate-impact ratio is 0.724 [0.693, 0.754] for the lowest socio-economic neighborhood cluster, excluding the four-fifths screening value, against 0.809 [0.776, 0.840] for the ethno-socioeconomic proxy, whose interval contains it, and no measurable gender disparity. Group membership is recoverable from the neutral feature set at an AUC of 0.654, and 42% of the group gap in predicted risk travels through the bureau credit score alone, so feature deletion cannot close the channel. Feature attributions and an auxiliary group-recoverability test locate the proxy pathways through which disparity arises, and a misclassification-sensitivity analysis bounds the effect of error in the group proxy, which attenuates measured disparity toward parity. We map the results onto Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744, the GDPR as interpreted in SCHUFA Holding, Directive (EU) 2023/2225, EBA loan-origination guidance, and Solvency II, EIOPA, and IAIS expectations, and propose fairness-risk controls organized around impact assessment, independent validation, and three lines of defense governance. Because the evidence comes from credit origination at a single lender, the insurance argument is developed at the level of regulatory and governance architecture rather than as an empirical transfer of estimates. Full article

Automated inspection of prepared meals is important for improving food safety, quality assurance, and production efficiency. However, vision-based inspection remains challenging because boxed meals contain multiple adjacent food items with irregular shapes, varying portion sizes, and visually similar appearances, while oily surfaces may introduce specular reflections that degrade image quality. This study presents a vision-based framework for multi-object recognition and quantitative defect analysis in complex meal images, with Chinese-style lunch boxes used as representative test samples. The framework integrates region-of-interest (ROI) extraction, fixed-grid regional feature representation, deep neural network (DNN) classification, flood-filling post-processing, and empirical food-quantity thresholds. The lunch-box ROI is first extracted using the Hough transform, followed by median filtering to suppress reflection noise. The ROI is partitioned into 6 × 6 regions, from which the mean and standard deviation of RGB, HSV, and CIE Lab* color components are extracted and classified using a DNN. Flood filling is subsequently applied to refine the classification results, and category-specific empirical thresholds are used to identify missing food items and insufficient portions. Foreign objects are detected as an additional abnormal category. Under the evaluated experimental conditions, the proposed framework achieved an overall image classification rate (CR) of 96.45%, a defective-image detection rate (1−β) of 97.76%, a normal-image false alarm rate (α) of 5.34%, and a defective-image misclassification rate (γ) of 0.82%. Sensitivity experiments involving illumination variation, two lunch-box configurations with different food compositions, and conveyor-based image acquisition further demonstrated the feasibility and stability of the framework under the evaluated laboratory and prototype conditions. These findings support the feasibility of the proposed lightweight framework for vision-based quality inspection of representative boxed meals, while broader validation across meal types, contaminants, and industrial production environments remains necessary. Full article

Zn-Al-Mg (ZAM) coatings have attracted significant attention in the field of corrosion protection owing to their combination of excellent corrosion resistance and cost-effectiveness. However, the corrosion behavior of ZAM coatings was governed by the synergistic coupling effects of multiple factors, including alloy composition, coating thickness, corrosive medium, and multiphase microstructure, making it challenging for traditional empirical analysis to systematically reveal the underlying mechanisms. To address this challenge, we constructed a multidimensional corrosion dataset comprising alloy composition, corrosive medium, coating thickness, and phase composition, based on literature data from the past three decades combined with self-measured potentiodynamic polarization experimental results. After data normalization and correlation analysis, we introduced phase structure features—including the Al-rich phase, MgZn₂ phase, Mg₂Si phase, and eutectic microstructures—to enhance the model’s capability in representing microstructural factors. On this basis, we established random forest (RF), support vector regression (SVR), and artificial neural network (ANN) models to predict the corrosion current density, and subsequently conducted an interpretability analysis using the SHapley Additive exPlanations (SHAP) method. The results demonstrated that the expanded feature set significantly improved the prediction performance of the models. Among them, the RF model exhibited the best performance, achieving a determination coefficient (R²) of 0.7363 on the test set, which represented a substantial improvement over the baseline dataset. Feature importance analysis revealed that coating thickness, Mg content, NaCl concentration, and Zn content were the primary factors influencing the corrosion current density. Further SHAP analysis showed that the marginal contribution of the eutectic phase was more prominent in local samples. Meanwhile, the Mg element exhibited distinct non-linear regulation characteristics, exerting varying impacts on the corrosion behavior across different concentration ranges. This study demonstrated that the interpretable machine learning models constructed via feature engineering not only improved the prediction accuracy of the corrosion performance of ZAM coatings, but also provided a novel data-driven approach to revealing the intrinsic correlations among alloy composition, phase structure, and corrosion response. Full article

Comparing metro-station service at national scale is hindered by the scarcity of consistent ridership or mobile-phone data across cities. We propose MetroVFE, a topology-based framework for diagnosing metro-station structural service potential. It combines three coupled subsystems—vitality (network centrality, transfer capacity, and feeder-bus connectivity), function (route-type diversity, destination reachability, directional coverage, and operator diversity), and spatial coverage equity (station-spacing uniformity). All indicators are derived from publicly available transit-network topology. Each subsystem is aggregated by entropy weighting, and the three are combined through a coupling coordination degree (CCD) model that penalizes unbalanced development. We instantiate the framework on the CPTOND-2025 national bus–metro vector dataset, computing indicators for 7057 unique metro stations in 45 Chinese cities, with 42,796 feeder bus routes spatially joined to 500 m station catchments. The empirical analysis yields four main findings: (i) Coupling coordination is moderate and has positive skew: the mean station CCD is $0.46$, with $58.5\%$ of stations in antagonistic or transitional bands and only $1.5\%$ highly coordinated. (ii) vitality is the dominant lagging subsystem ($66.4\%$ of stations), and this share rises monotonically from large to small networks ($63.7\%\to 69.3\%\to 78.2\%$). (iii) Unsupervised clustering recovers five interpretable diagnostic typologies, dominated by vitality-deficient ($48.3\%$) and function-developing ($24.8\%$) stations, with only $12.3\%$balanced–coordinated. (iv) Spatial coverage equity modestly moderates the function→vitality relationship (interaction $p=1\times {10}^{-4}$, robust to controls): the marginal contribution of functional diversity to vitality is about $25\%$ larger in high-coverage-equity than low-coverage-equity station areas ($0.207$ vs. $0.165$). Mean coordination does not differ significantly across network-size tiers (ANOVA $p=0.12$), but typology composition does (${\chi }^2$$p<{10}^{-58}$). Ablation and sensitivity analyses confirm that the coordination feature and the full V-F-E feature set are necessary for stable typologies. MetroVFE is fully reproducible from open data and provides an actionable, mathematically grounded tool for prioritizing station-level interventions. Full article

Optical systems in the real world often create image problems, such as Gaussian blur from defocus or atmospheric turbulence, and radial vignetting caused by lens shape. Standard mixed-data fine-tuning helps the task head handle these issues, but it does not actually fix them. We introduce the Iterative Correction of Optical Perturbations ICOP framework, which corrects encoder feature representations before they reach the task head using a physics-aware plug-in module. ICOP does this by modeling blur as an isotropic-Gaussian point-spread function (PSF) and uses gradient-based, self-supervised optimization (Adam) to discover feature-space corrections that steer degraded representations toward their clean-data distribution. It includes a BlurEstimator that builds a degradation descriptor using fixed Laplacian and Sobel operators, and a PromptGenerator that turns this descriptor into modulation parameters for the frozen encoder output. The framework comes in two versions based on the task: an additive correction (ICOP-Add) for image classification, and a Feature-wise Linear Modulation correction (ICOP-FiLM) for object detection. We observe a convergence between clean-task performance and blur-induced degradation across datasets, consistent with greater reliance on high-frequency features in stronger backbones; we treat this as an empirical, cross-dataset observation rather than a demonstrated causal claim (task difficulty, category structure, texture, and object scale also differ across datasets). Independently of this, ICOP-FiLM’s corrective benefit does not scale with degradation severity, revealing a more nuanced relationship between backbone quality and robustness. For classification, ICOP-Add improves distorted-condition accuracy over strong mixed fine-tuning by $+2.4$, $+9.1$, and $+10.7$ percentage points on MNIST, FashionMNIST, and CIFAR-10, respectively (McNemar’s test, $p<0.001$ on all three, 5000 paired predictions per dataset). On three object detection datasets, ICOP-FiLM improves distorted-condition mAP over a mixed-fine-tuning null hypothesis by $+0.026$, $-0.005$, and $+0.003$ mAP, respectively (all values mean over 3 seeds). Against a matched-blur-ratio control that isolates the correction module’s own contribution, ICOP-FiLM wins by a consistent margin on two of the three datasets ($+0.037$ and $+0.024$ mAP, winning in every one of 3/3 seeds on each) and loses on the third ($-0.039$ mAP, losing in 3/3 seeds); it outperforms parameter-efficient (VPT, Adapter) baselines trained on identical data on the same two datasets. This dataset-dependent pattern is discussed in detail in the main text. ICOP-FiLM adds only 82,672 parameters to a 42-million-parameter Real-Time DEtection TRansformer (RT-DETR) detector. All reported results are obtained under synthetic Gaussian blur and radial vignetting applied to clean images from the six benchmark datasets studied. Full article

Tunnel traffic surveillance is challenged by uneven illumination, severe occlusion, distant small objects, and long-tailed class distributions, while practical edge deployment imposes strict computational constraints. To address these problems, this paper proposes an integrated object detection, structured compression, and edge-deployment framework based on YOLOv8. A C2f-P3A module is developed to enhance spatial and channel feature representation, while Lite-ASPP is introduced to efficiently incorporate multi-scale contextual information. WIoU v3 is adopted to optimize bounding-box regression, and a dependency-preserving channel-level LAMP strategy is employed to adaptively allocate sparsity across eligible layers. Experimental results on the tunnel-surveillance dataset show that the improved dense model achieves a Precision of 90.49%, a Recall of 75.91%, an mAP50 of 83.13%, and an mAP50:95 of 57.60%, outperforming YOLOv8s by 5.46, 6.56, 4.98, and 2.17 percentage points, respectively. Furthermore, a target channel sparsity of 60% reduces the parameter count from 11.24 M to 9.86 M and the computational cost from 28.4 to 25.7 GFLOPs, while retaining an mAP50 of 83.11%. After mixed-precision deployment on the RK3588 platform, the pruned model achieves an mAP50 of 72.87%, an average processing time of 39.04 ms/frame, and a processing rate of 25.61 FPS. These results indicate a favorable empirical trade-off among detection accuracy, model complexity, and edge-processing efficiency under the evaluated conditions. Full article

The solvent-accessible surface area (SASA) of amino acid residues is a crucial parameter for protein structure analysis; however, precise computational methods such as FreeSASA are computationally expensive. As an alternative, empirical approximations based on residue interaction network (RIN) graphs can offer high speed while maintaining acceptable accuracy. In this study, we propose and validate three empirical functions for estimating relative SASA—approx_sasa, surface_score, and exp_sasa—using node degree as the sole argument. We present a comparative analysis of two graph construction approaches: the classical Cα-graph (8 Å threshold) and the heavy-atom graph (HAG, 5.0 Å threshold). Parameters were calibrated on a dataset of 509 protein structures (128,794 residues) using the true relative SASA calculated by the FreeSASA library. An extended set of 11 topological features was also developed and validated. Ensemble models (Random Forest, XGBoost) achieved a best performance of MAE = 0.057 ± 0.033 and Pearson r = 0.915 ± 0.080 on HAG, outperforming graph neural networks (GCN, GAT, GraphSAGE) in this setting. The empirical formulas demonstrate extreme computational efficiency (0.008 ms per structure), ~26,000× faster than FreeSASA, making them suitable for large-scale pipelines requiring both speed and interpretability. Random Forest on HAG is recommended for applications requiring maximum accuracy, while GraphSAGE on HAG is a viable deep learning alternative. Full article

This study examines the forensic classification and identification of counterfeit banknotes through an integrated analysis of empirical observations obtained during field and laboratory investigations. Particular attention is given to the procedural workflow of forensic examination and to the complementary use of macroscopic, microscopic, optical, and physicochemical methods. Preliminary field assessments were conducted using document examination loupes and other optical magnification devices, while laboratory investigations employed advanced analytical systems and specialized software to characterize the substrate, printing inks, graphic elements, and security features. The findings demonstrate that reliable counterfeit banknote identification cannot be based solely on visual inspection. A multidisciplinary approach integrating morphological, optical, and compositional evidence is essential for distinguishing sophisticated counterfeits from genuine banknotes and for formulating scientifically robust forensic conclusions. This study also highlights the need for continuously updated examination procedures in response to the increasing technological sophistication of banknote counterfeiting. Full article

Artificial intelligence (AI) is increasingly used in water research. However, many AI models, particularly complex machine learning models, often generate outcomes that are difficult for humans to interpret. Explainable artificial intelligence (XAI) has been developed to address these challenges by providing transparent and human-interpretable explanations of model behavior and predictions. We conducted a structured narrative review using predefined searches of Web of Science Core Collection and Scopus to synthesize empirical XAI applications across six water-research domains: hydrological processes, water quality and pollution, groundwater systems, urban water systems, climate–water interactions, and water and wastewater treatment. The review covers feature-importance methods, SHapley Additive exPlanations (SHAP), Local Interpretable Model-agnostic Explanations (LIME), partial dependence plots (PDPs), individual conditional expectation (ICE) plots, accumulated local effects (ALE) plots, counterfactual explanations, and deep-learning attribution methods. Building on previous reviews and perspectives focused on particular water domains or methodological priorities, we provide a cross-domain synthesis of XAI spanning natural and engineered water systems, with emphasis on method selection, model and data compatibility, explanation reliability, and operational implementation. These capabilities, however, must be interpreted with appropriate caution because XAI explanations remain conditional on the data, fitted model, and explanation method, and therefore should not be treated as evidence of causal mechanisms or environmental controls. Recognizing these limitations, we provide practical guidance for selecting and evaluating XAI methods and outline priorities for developing reliable, scalable, and operationally useful AI systems for water research and management. Full article

Filtering-based stereo visual-inertial odometry often assigns fixed or uniformly scaled covariance to tracks accepted by the front-end, although accepted observations can differ in terms of tracking reliability, local directional identifiability, and agreement with the batch-level innovation model. This issue is important for UAV-based multisensor inspection platforms, where pose estimates support autonomous flight, measurement registration, repeatable survey lines, and multisensor data fusion. This paper presents RDIC-MSCKF, a Risk–Direction-Decoupled and Innovation-Calibrated MSCKF, where innovation calibration denotes bounded empirical scaling within the visual update. RDIC-MSCKF maps robust tracking diagnostics to a bounded standard-deviation multiplier and uses a robust local photometric information matrix to add penalty-only anisotropic covariance along weak image directions. The resulting observation covariance is preserved during MSCKF landmark elimination through full projected-covariance whitening. In parallel with feature-block innovation gating, bounded minimum measurement-noise inflation is estimated from the pre-gate innovation population and applied through a Kalman-equivalent modal update. On ten evaluated EuRoC MAV sequences, RDIC-MSCKF obtains lower ATE RMSE than the S-MSCKF baseline on nine sequences; averaged over five runs per sequence, the mean RMSE decreases from 0.1869 m to 0.1236 m, corresponding to a 33.8% reduction, and the sequence-mean P90 error decreases by 32.0%. Runtime profiling on five representative EuRoC sequences gives a 26.32 ms mean and 37.33 ms P95 per-frame processing time for RDIC-MSCKF, with 0.01% of profiled frames above the 50 ms reference. Outdoor UAV flights with RTK reference trajectories further demonstrate lower Sim(2)-aligned horizontal RMSE than S-MSCKF on all three evaluated flights. Full article

The rapid growth of the Internet of Things and embedded platforms increases the demand for cryptographic primitives that provide adequate security under strict resource constraints. This paper introduces FLUX-128 (Fast Lightweight Universal miXing), a new lightweight hash function targeting constrained devices. FLUX-128 is built on a modified sponge construction with a 216-bit state, a 72-bit absorption rate, and a 128-bit tag length. Its main design feature is an additional lightweight diffusion function $G$ applied before the core transformation $F$, improving mixing while keeping the architecture compact. Diffusion was evaluated via the Strict Avalanche Criterion (SAC) on 5000 messages using systematic single-bit input inversions and a corresponding probability matrix of output-bit changes. In addition, the concatenated output bitstream was tested with the NIST Statistical Test Suite, where all reported p-values exceed the 0.01 significance threshold. Overall, the results provide empirical evidence of near-ideal bit mixing and the absence of pronounced statistical defects within the adopted experimental setup. Together with the security objectives defined by the 128-bit output and the 144-bit sponge capacity, these results support FLUX-128 as a promising lightweight hash construction for integrity-oriented applications in resource-constrained IoT environments. The SAC and NIST STS results are considered complementary experimental evidence of the expected cryptographic behavior rather than stand-alone proofs of resistance to all classes of cryptanalytic attacks. Full article

Machine learning has transformed consumer credit scoring, delivering substantial gains in predictive accuracy over traditional scorecards—but whether those gains come at a cost to fairness has remained contested. The dominant assumption in the literature is that more complex, accurate models amplify bias by encoding historical patterns of disadvantage more effectively. This paper challenges that assumption with direct empirical evidence. We evaluate four model families—logistic regression, random forest, XGBoost, and a multilayer perceptron—across two real-world datasets: the UCI Credit Card Default Dataset and the 2024 US Home Mortgage Disclosure Act national loan-level data, comprising over six million mortgage applications. Using repeated cross-validation, we report predictive performance alongside two primary fairness metrics—demographic parity difference and equalized odds difference—supplemented by false positive rate difference and calibration difference, with confidence intervals across 15 estimation folds. On the UCI data, where demographic disparities are modest, model choice has negligible effect on fairness outcomes. On the HMDA mortgage data, where racial disparities are large and legally consequential, the expected accuracy–fairness tradeoff does not hold; more accurate models produce significantly fairer outcomes on equalized odds within the models, data, and fairness criteria examined here, with logistic regression occupying the worst position simultaneously on all dimensions. Persistent demographic parity disparity among the more complex models is consistent with feature-level bias that no model architecture can resolve. The findings have direct implications for the less-discriminatory-alternatives framework under US fair lending law and for the high-risk classification of credit scoring AI under the EU AI Act. Full article

Machine learning has become a widely adopted approach for URL-based phishing detection, with many studies reporting F1 scores exceeding 0.95 on benchmark datasets. However, recent adversarial machine learning research has questioned the robustness of these models, suggesting that small input perturbations can severely degrade detection performance. In this study, we argue that a substantial part of this reported vulnerability stems from the way adversarial attacks are evaluated. Specifically, many existing studies assess attacks in the feature space, where feature values are modified directly without ensuring that the resulting samples correspond to valid, functional URLs. To investigate this issue, we conduct a two-stage empirical study using both a benchmark feature dataset and a dataset of real phishing URLs. Crucially, to avoid confounding the attack space with dataset differences, we additionally evaluate both feature-space and problem-space attacks on the same real-URL dataset, using an identical model and manipulable-feature budget. Our experiments reveal a striking contrast between these evaluation settings. While feature-space attacks reduce the detection rate of a Random Forest classifier on the benchmark dataset from 0.96 to 0.36, analogous manipulations performed on real URLs have almost no effect on detection performance, as the most informative signals originate from host-related attributes that are difficult for attackers to manipulate. Building on this observation, we propose a set of robust features that capture stable domain characteristics, including lexical word validity, homoglyph disguises, brand impersonation, subdomain depth, character entropy, and transport-related signals. Incorporating these features substantially improves robustness under adversarial conditions, maintaining phishing detection rates between 0.24 and 0.76 where the lexical-only baseline deteriorates to zero under a non-adaptive attacker, while also increasing the clean-data F1 score from 0.985 to 0.994. We further evaluate an adaptive attacker that explicitly targets the proposed features; although the proposed representation raises the attacker’s cost and helps under moderate attacks, host-derived features remain the only strictly attack-invariant component, so we position the proposed features as a complement to host-based signals rather than a standalone defense. Additional analyses, including model comparison, hyperparameter sensitivity analysis, feature ablation, SHAP-based interpretation, multi-seed confidence intervals, a domain-disjoint evaluation, and host-only evaluation, consistently support the proposed approach. The findings demonstrate that problem-space evaluation provides a more realistic assessment of adversarial robustness than conventional feature-space testing and show that robust feature engineering offers a practical strategy for developing phishing detection systems that remain effective under realistic adversarial conditions. Full article

This study introduces a novel clustering approach, namely Teacher–Student-based Deep Clustering (TSDC), that relies on intra- and inter-distillation-based feature representations. In fact, TSDC distils knowledge from (i) high- to low-response channels, forcing the latter to mimic the former, and (ii) deeper to shallower layers, prompting the transfer of semantic information to enhance representational consistency. Unlike CNN-based deep clustering that relies on pseudo-labels to improve representations, TSDC introduces a newly formulated objective function to simultaneously minimize losses from clustering and intra- and inter-distillation. The proposed approach was rigorously investigated using benchmark datasets and relevant performance measures. In particular, a linear top classifier protocol was adopted to assess TSDC performance. Notably, TSDC outperformed existing deep clustering frameworks, yielding improved classification accuracy. The empirical findings highlight the efficacy of combining intra- and inter-distillation to enrich feature representations. Notably, the most prominent improvement was observed on CIFAR-100, where classification accuracy rose from 54.77 ± 0.07 to 57.58 ± 1.85. Full article

Background: Artificial intelligence (AI)-driven conversational systems are increasingly capable of simulating empathy, adapting to individual users, and fostering emotional bonds that blur the boundary between tool and companion. This scoping review maps the extent and nature of published evidence regarding psychological mechanisms underlying emotional reliance on AI chatbots and associated mental health implications. Methods: Conducted in accordance with PRISMA-ScR guidelines, we systematically searched PubMed/MEDLINE, PsycINFO, Web of Science, Scopus, and IEEE Xplore from database inception to March 2026. Two independent reviewers screened records and extracted data using the PCC (Population, Concept, Context) framework. Thematic synthesis was performed to map evidence across psychological, clinical, and developmental domains. Results: Of 1847 records identified, 46 studies met inclusion criteria. Key themes included: (1) the ELIZA effect as a foundational mechanism of human–AI attachment, with documented cases of severe dependency including fatal outcomes; (2) anthropomorphization and artificial intimacy fostered by adaptive, personalized AI design; (3) proposal of Generative AI Dependency (GAID) as a conceptual framework mapping onto behavioral addiction components, pending empirical validation; (4) particular vulnerability of adolescents and lonely individuals to exclusive affective bonds with AI; and (5) potential erosion of human relational capacities, empathy development, and tolerance for interpersonal complexity. Significant gaps were identified in longitudinal research, validated screening tools, and intervention protocols. Conclusions: Emotional reliance on AI represents an emerging clinical phenomenon with addiction-like features requiring specific diagnostic frameworks, evidence-based interventions, and ethical design guidelines. Future research should prioritize longitudinal studies examining developmental impacts and neurobiological investigations of AI-mediated reinforcement mechanisms. Full article