<?xml version="1.0" encoding="utf-8"?>
    <rss version="2.0">
      <channel xmlns:content="http://purl.org/rss/1.0/modules/content/">
        <title>Frontiers in Artificial Intelligence | Machine Learning section | New and Recent Articles</title>
        <link>https://www.frontiersin.org/journals/artificial-intelligence/sections/machine-learning</link>
        <description>RSS Feed for Machine Learning section in the Frontiers in Artificial Intelligence journal | New and Recent Articles</description>
        <language>en-us</language>
        <generator>Frontiers Feed Generator,version:1</generator>
        <pubDate>2026-10-03T09:31:55.849+00:00</pubDate>
        <ttl>60</ttl>
        <item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1854934</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1854934</link>
        <title><![CDATA[Explainable AI for phishing URL detection: a Bayesian-optimized stacking ensemble framework with SHAP-guided feature learning]]></title>
        <pubdate>2026-10-02T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Hafiz Aziz Khan</author><author>Sonia Akter</author><author>Abdur Rahman Lindon</author><author>Taslima Akter</author><author>Iftekhar Rasul</author><author>Mamunur Rahman</author><author>Nasrin Akter Tohfa</author><author>Iftekhar Hossain</author>
        <description><![CDATA[IntroductionPhishing remains one of the most persistent and financially damaging threats facing modern organizations, with over 4.7 million incidents recorded in 2023 alone. Existing AI-based phishing detection frameworks are constrained by limited benchmarking scope, absent model interpretability, and insufficient statistical validation — three limitations that collectively restrict operational utility in real-world security environments.MethodsWe present an explainable, end-to-end machine learning pipeline evaluated on a large public benchmark of 247,950 URLs described by 41 structural and lexical features. The pipeline integrates SHAP-driven feature selection (reducing 41 to 24 features via a 95% cumulative-signal rule), a systematic benchmark of 12 classifiers spanning seven algorithmic families, Bayesian hyperparameter optimization via Optuna TPE sampling (40 trials each for XGBoost and CatBoost), and a heterogeneous stacking ensemble combining Optuna-tuned XGBoost, CatBoost, Extra Trees, and Random Forest under a logistic-regression meta-learner. A four-layer statistical validation protocol — comprising a Friedman omnibus test, Wilcoxon signed-rank tests, paired t-tests, and Cohen's d effect sizes — was applied to five-fold cross-validation accuracy distributions to assess directional consistency, with the limited inferential resolution of five folds explicitly acknowledged.ResultsSHAP-driven selection reduced the feature space by 41.5% while retaining 95% of predictive signal. The stacking ensemble achieved 96.75% accuracy, 96.74% F1-score, and AUC of 0.9947, attaining the lowest Brier score among all 13 models (0.0246), indicating superior probability calibration. The Friedman omnibus test confirmed significant performance differences across models (χ2F = 59.84, p < 0.0001), and all 12 Wilcoxon pairwise comparisons yielded the minimum attainable p-value (p = 0.0313), confirming the ensemble never lost a cross-validation fold against any baseline. Post-hoc SHAP analysis identified subdomain structure, URL length, and URL entropy as the dominant phishing indicators at both ensemble and base-learner levels.DiscussionThe co-leaders — the stacking ensemble and Extra Trees — demonstrate that rigorous, interpretable AI pipelines can advance phishing detection accuracy and transparency simultaneously. The framework's calibrated risk scores, threshold flexibility, and multi-level SHAP explainability support analyst-facing decision-making in security operations, while its leakage-free]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1900815</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1900815</link>
        <title><![CDATA[Intelligent defense at the edge: a comprehensive survey of federated learning, TinyML and explainable AI for intrusion detection in IoT and IIoT ecosystems]]></title>
        <pubdate>2026-10-02T00:00:00Z</pubdate>
        <category>Review</category>
        <author>S. Sarath Kumar</author><author>M. Asha Jerlin</author>
        <description><![CDATA[The proliferation of Internet of Things (IoT) and Industrial Internet of Things (IIoT) technologies has fundamentally transformed contemporary computing infrastructures by interconnecting large heterogeneous devices, sensors, embedded systems, and cyber-physical platforms. These ecosystems support diverse applications including smart cities, intelligent transportation, industrial automation, medical surveillance, precision agriculture and home automation. Despite their operational advantages, the vast connectivity and inherent resource constraints of IoT devices significantly expand the cyberattack surface. Limited processing power, limited memory and restricted energy budgets render conventional cybersecurity solutions infeasible for IoT environments. Consequently, IoT systems have become prime targets for advanced threats, including malware, botnets, Distributed Denial of Service (DDoS) attacks, ransomware, insider attacks and data manipulation. Traditional Intrusion Detection Systems (IDS), designed for enterprise networks, rely on centralized data collection and computationally intensive machine learning models that are unsuitable for IoT deployments due to scalability limitations, communication overhead, latency constraints and privacy concerns. Recent research has therefore shifted toward lightweight, intelligent and distributed cybersecurity frameworks capable of providing robust protection under severe resource limitations. This survey comprehensively reviews recent advances in AI-based intrusion detection and malware analysis for IoT ecosystems, covering lightweight machine learning models (ML), deep learning (DL) architectures for edge deployment, federated learning (FL) based privacy-preserving detection, TinyML on-device intelligence, blockchain-based trust management, Explainable AI (XAI) driven transparent monitoring and adaptive cyber defense. A unified conceptual framework is proposed integrating lightweight detection, distributed collaborative learning, trust-based orchestration and adaptive defense.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1904459</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1904459</link>
        <title><![CDATA[Hybrid neuro-symbolic graph explanation framework for fraud detection in financial transaction networks]]></title>
        <pubdate>2026-10-01T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Neeraj S. Kumar</author><author>Idhikash J.</author><author>Jannath Nisha O. S.</author><author>Shreenidhi S. R.</author>
        <description><![CDATA[IntroductionThis research focuses on the development of a hybrid neuro-symbolic system for fraud detection in rapidly expanding digital financial systems.MethodsA bipartite customer-merchant graph is constructed from the Nigerian Financial Transactions dataset, and a 3-layer Graph Attention Network (FraudGAT) is used to model structural dependencies while a symbolic reasoning layer and constrained LLM prompting generate auditor-readable explanations.ResultsOn the curated evaluation split, FraudGAT achieves AUC 0.874, AUPRC 0.771, and F1 0.609. We additionally benchmark against GCN, GraphSAGE, GIN, a graphtransformer baseline, and tabular baselines (Random Forest, XGBoost, LightGBM): tabular ensembles perform best on this sampled split, while FraudGAT remains stronger than several graph baselines and supports an interpretable detect-explain workflow.DiscussionsTherefore, we position the contribution as an auditable neuro-symbolic graph framework and an empirical foundation for scaling to larger, richer transaction graphs.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1900887</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1900887</link>
        <title><![CDATA[Enhancing vehicle routing problem through a qubit-efficient Symmetric Arc Reduction using Quantum Approximate Optimization Algorithm]]></title>
        <pubdate>2026-10-01T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Jayaraja A</author><author>Senthil Kumaran U</author>
        <description><![CDATA[IntroductionFinding the most efficient routes for a fleet of delivery vehicles remains a critical and costly challenge for nearly any logistics business. This core idea is known as the Vehicle Routing Problem (VRP). The primary goal is to minimize operational costs, such as the total travel distance or the number of trucks deployed, while ensuring all the customers are served. Because VRP is NP-hard, finding a perfect solution is often impossible, forcing industries to rely on classical approximate algorithms. However, quantum variational algorithms offer a new paradigm by iteratively optimizing parameters to navigate the solution space, presenting a new and promising avenue for finding better and faster solutions.MethodsIn this paper, we introduce Symmetric Arc Reduction in Quantum Approximate Optimization Algorithm (SAR-QAOA), which systematically eliminates redundant symmetric routing variables in the binary integer programming formulation prior to the Quadratic Unconstrained Binary Optimization (QUBO) formulation. The proposed SAR-QAOA encodes edge-based VRP while reducing the qubit requirement from n(n−1) to n(n−1)/2, thereby improving the resource efficiency of the QUBO formulation by prioritizing regions of the solution space that are more likely to contain high-quality solutions; SAR enhances both search accuracy and computational efficiency.ResultsExperimental results on noiseless quantum simulators validate the feasibility of the proposed formulation and its integration with the QAOA framework, providing a promising foundation for future investigations on larger routing problems and evolving quantum hardware.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1919564</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1919564</link>
        <title><![CDATA[UniPET: a unified approach for whole-body CT to PET translation]]></title>
        <pubdate>2026-10-01T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Francesco Di Feola</author><author>Valerio Guarrasi</author><author>Mikael Johansson</author><author>Paolo Soda</author>
        <description><![CDATA[Positron emission tomography (PET) provides critical metabolic information for oncological imaging, yet its use is constrained by radiation exposure, cost, and limited availability. Synthesizing PET-like images from computed tomography (CT) has been proposed as a way to approximate metabolic information; however, existing approaches either fail to capture region-specific variability or rely on multiple organ-specific models that do not scale to whole-body imaging. In this work, we introduce UniPET, a unified approach for whole-body CT-to-PET translation based on curriculum learning. By progressively incorporating anatomical regions of increasing complexity, UniPET enables a single network to learn coherent morpho-metabolic relationships across different regions. We evaluate UniPET on a public dataset of 900 patients and an external cohort of 579 patients. The model achieves image quality and metabolic consistency comparable to region-specific approaches, while improving over conventional whole-body models and maintaining stable performance under distribution shifts. By capturing region-specific variability within a single model, UniPET provides a promising approach for augmenting CT-based workflows with synthetic metabolic information, with potential applications in screening triage, in low-resource environments where PET is unavailable, and as an additional input for multimodal analysis pipelines. The code is available at: https://github.com/arco-group/UniPET.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1928178</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1928178</link>
        <title><![CDATA[Robust multi-class DDoS and fraud detection using a Context-Adaptive Deep Evolutionary Game Framework with CNN–BiGRU]]></title>
        <pubdate>2026-09-30T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Ashwini V.</author><author>Priya V.</author>
        <description><![CDATA[Distributed Denial of Service (DDoS) attacks have become a serious threat in modern computer networks, which can commonly be integrated with other types of attacks to evade detection by conventional devices. Existing deep learning-based intrusion detection methods have high identification capability. However, they are also highly susceptible to adversarial attacks, in which even slight perturbations to features can fool the model into misclassification. To address these issues, this study introduced a hybrid deep learning framework. It integrates Convolutional Neural Networks (CNN) and Bidirectional Gated Recurrent Units (BiGRU) to detect diverse types of attacks in real time. CNN extracts spatial interactions to model traffic characteristics, while BiGRU, a recurrent neural network, models temporal dependencies in data flows. This combination enables the model to capture complex spatiotemporal behavioral patterns in network traffic. To address class imbalance, we propose a new Neuro-Evolutionary Resampling Algorithm (NERA). This method uses neural density estimation and evolutionary optimization to generate balanced, noise-filtered training samples. Additionally, a Context-Adaptive Deep Evolutionary Game Framework (CADEGF) is designed to improve resistance to attacks. It does this by simulating interactions between attackers and defenders. It also adjusts detection thresholds using game-theoretic ideas. Experimental results show that this approach achieves high classification accuracy, effectively detects minority classes, and highlights trade-offs between minority-class detection and adversarial robustness, indicating contexts where NERA benefits or harms resilience. The integrated synergy of NERA, CNN-BiGRU, and CADEGF establishes a context-aware, adaptive defense model that advances intelligent intrusion detection with balanced performance and adversarial robustness.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1886221</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1886221</link>
        <title><![CDATA[SkinNet-X: a task coupled feature interaction framework for joint skin lesion segmentation and classification]]></title>
        <pubdate>2026-09-30T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>E. Babu</author><author>S. Murali</author>
        <description><![CDATA[Well trained segmentation and classification of skin lesions from dermoscopic images are crucial to an early detection of melanoma and other malignant skin disease. While many deep learning models reach impressive overall performance, most existing approaches treat segmentation and classification as separate tasks, leading to limited interaction between spatial lesion localization and semantic disease knowledge. This segregation results in a high variance prediction and would reduce the overall robustness of automated diagnostic systems. To mitigate this limitation, we proposed SkinNet-X: a unified multitask framework to reformulate the skin lesion segmentation and classification tasks jointly. We achieve this with a shared hierarchical feature representation to transfer knowledge between both tasks while introducing new Cross Task Consistency Learning (CTCL) strategy that enforces semantic consistency between lesion localization and disease prediction during joint optimization. In order to enhance the discriminative capability of feature expression and alleviate redundant representations, a dual stage multi label feature optimization module is further exploited, accompanied by an adaptive PolyLoss based optimization strategy stabilizing multitask learning. The effectiveness of the proposed SkinNet-X framework was evaluated using segmentation and classification experiments on the ISIC 2018, ISIC 2019, and HAM10000 benchmark datasets. We also performed additional analyses, namely ablation studies, computational complexity evaluation, sensitivity analysis and statistical validation to explore the contribution of each architectural component. Experimental results show that the proposed multitask learning strategy can consistently improve accuracy of lesion localization and performance of classification with robust generalizability across multiple datasets. These findings demonstrate that SkinNet-X is an effective and potentially useful framework for automated skin lesion analysis and computer-aided dermatological diagnosis.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1834939</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1834939</link>
        <title><![CDATA[Deepfake anomaly detection using kernel density estimation based convolutional autoencoder]]></title>
        <pubdate>2026-09-30T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Manas Ranjan Prusty</author><author>Sahil Amritkar</author><author>Ananthakrishnan Balasundaram</author><author>Subhra Rani Patra</author>
        <description><![CDATA[The rise of deepfake technology has fuelled the need for robust detection methods to counter the increasing sophistication of manipulated multimedia content. Research has shown that humans struggle to identify deepfakes, highlighting the pressing need for AI-based solutions to safeguard the integrity of the information we consume daily. This paper proposes a novel approach CAE-DAD (Convolutional Autoencoder Deepfake Anomaly Detector) which facilitates deepfake detection through visual anomaly identification using convolutional autoencoders (CAEs). Methods like Kernel Density Estimation are used in the latent space to widen the anomaly score gap between real and fake images. The model is trained only on real face images. Existing models like VGG16 which are known for their feature extraction capabilities as incorporated as part of our encoder. The model is benchmarked against state-of-the-art methods and establish a comprehensive set of evaluation metrics, including accuracy, precision, recall, F1 score. The model achieves comparable performance to existing two-class training methods, and achieves 91% accuracy. The proposed anomaly identification using CAEs serves as a valuable addition to the arsenal of tools aimed at mitigating the risks associated with the proliferation of deepfake content.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1934604</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1934604</link>
        <title><![CDATA[TL-CRANN: a transfer-learning-based hybrid CNN-RNN model with attention mechanism for cross-domain IoT attack detection]]></title>
        <pubdate>2026-09-29T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>L. Sankar Ganesh</author><author>P. Mohan Kumar</author>
        <description><![CDATA[IntroductionThe rapid and pervasive growth of Internet of Things (IoT) devices has significantly expanded the attack surface of current network environments, making them vulnerable to advanced and evolving cyber threats. However, current intrusion detection systems (IDS) usually suffer from weak cross-domain generalization, high false positive rates, and poor performance against novel attacks.MethodsTo overcome these issues, this paper proposes TL-CRANN, a transfer-learning-based hybrid deep learning framework utilizing a convolutional neural network (CNN), bidirectional long short-term memory (Bi-LSTM), and a multi-head self-attention mechanism for cross-domain IoT attack detection. Transfer learning is leveraged to transfer knowledge acquired from source-domain traffic and achieve rapid adaptation to heterogeneous target IoT environments with scarce labeled data. The proposed model is evaluated using three benchmark datasets: CICIDS2017, TON_IoT, and Bot-IoT.ResultsExperimental results show that TL-CRANN achieved detection accuracies of 98.37%, 96.21%, and 97.14%, respectively, on the three datasets, and significantly outperformed existing state-of-the-art machine learning and deep learning baselines.DiscussionThe results show that transfer learning improves the cross-domain detection performance and training cost. Moreover, the ablation analysis shows that CNN, Bi-LSTM and attention mechanisms-based modules are contribution to effectiveness of proposed framework which also confirms the applicability of TL-CRANN for cross-domain IoT attack detection.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1957727</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1957727</link>
        <title><![CDATA[Leakage-free evaluation of task-aware Nyström sampling for IoMT intrusion detection]]></title>
        <pubdate>2026-09-28T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Hiba A. Tarish</author><author>Rosilah Hassan</author><author>Khairul Akram Zainol Ariffin</author><author>Mustafa Musa Jaber</author>
        <description><![CDATA[Standard Nyström kernel approximation samples landmark points uniformly at random, ignoring that boundary-adjacent and minority-class samples often matter most for imbalanced intrusion detection. We propose a task-aware, importance-weighted Nyström landmark-selection mechanism that scores each training sample's classifier uncertainty, boundary distance, representativeness, and minority-class importance, with an optional quantum-amplitude reparameterization (QI-AWNS) shown to be algebraically equivalent to direct score-proportional sampling; the contribution is the importance scoring and its rigorous evaluation, not the amplitude framing or a quantum advantage. We benchmark QI-AWNS against uniform, K-means, and approximate ridge-leverage-score Nyström sampling, five tabular baselines, and a full-kernel RBF-SVM on two leakage-free IoMT benchmarks (ECU-IoHT, WUSTL-EHMS-2020) under a strict split-before-fit protocol, with each test partition evaluated once. Across 10 reseeded splits, per-split retuning of the importance weights improves QI-AWNS's stability relative to a frozen-weight configuration and yields a mean balanced-accuracy gain of 0.18 points over uniform sampling on ECU-IoHT. This is not significant after Holm correction for multiple comparisons (raw Wilcoxon p = 0.022, adjusted p = 0.129), and no advantage appears over K-means or approximate ridge-leverage-score sampling on either dataset. An 8-variant ablation and weight-search analysis shows that a validation-based ranking of two weight candidates reverses on WUSTL-EHMS-2020's test set. These findings show that task-aware landmark weighting offers limited, dataset-dependent benefit, and that single-split conclusions can overstate it.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1936035</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1936035</link>
        <title><![CDATA[ANN-driven analysis of Cattaneo–Christov heat transfer and entropy generation in Maxwell hybrid nanofluid flow over a vertical cone]]></title>
        <pubdate>2026-09-25T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Bhuvana M</author><author>Durgaprasad P</author>
        <description><![CDATA[IntroductionThis research investigates Cattaneo-Christov non-Fourier heat transfer and entropy generation in Maxwell hybrid nanofluid flow over a non-Darcy porous conical surface using numerical, sensitivity, and Artificial neural networks (ANN) approaches.MethodsA coupled momentum-energy model that includes fluid relaxation, magnetic field effects, nanoparticle loading, porous resistance, and thermal relaxation is transformed using similarity variables and solved by a boundary-value numerical method. Sobol' global sensitivity analysis is employed to quantify the direct and interaction effects of the governing parameters on skin friction, Nusselt number, and entropy generation.ResultsThe analysis identifies the porous-medium parameter as the dominant contributor to skin friction and heat transfer, while its contribution to entropy generation remains substantial at both buoyancy conditions. The total-order indices further reveal increased parameter interactions in entropy generation at higher buoyancy. The artificial neural network, trained using the numerically generated dataset, accurately predicts skin friction, Nusselt number, and entropy generation, with test R2 values of 0.99479, 0.99642, and 0.99033, respectively, along with low prediction errors.DiscussionThe combined numerical-Sobol' sensitivity-ANN framework enables rapid evaluation of nonlinear viscoelastic heat-transfer systems. The results underscore the utility of Maxwell hybrid nanofluids in thermally non-Fourier porous media from an engineering perspective and provide a computational framework for evaluating their thermal performance.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1894859</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1894859</link>
        <title><![CDATA[Mind versus machine: can biological models of the mind inspire machine intelligence design?]]></title>
        <pubdate>2026-09-25T00:00:00Z</pubdate>
        <category>Hypothesis and Theory</category>
        <author>Michael Sharwood Smith</author>
        <description><![CDATA[Models of human cognition that look biologically plausible might in principle be a source of insights that could prove applicable to machine intelligence. The Modular Cognition Framework fits that description in terms of its biological credentials and its potential to identify the challenges confronting GenAI developers. However, it is an account of cognition in the mind and not in the brain. Accordingly, the account includes a clear delineation of the two intimately related concepts, mind and brain, and a definition of their relationship. If either brain or mind can contribute to this debate, a model of the mind could offer a more immediate promising source of inspiration than its complex neural underpinnings. Key features of the theoretical framework include its fundamentally heterarchical architecture having no single locus of control. In addition, general and domain-specific processing principles permit rapid, highly flexible, adaptable interaction between its network of modular systems as they collaborate in the execution of multiple tasks in a constantly changing environment. Eleven types of knowledge are constructed from experience using a similar number of inherited toolkits. Questions are raised as to whether the adaptable and highly creative character of evolved intelligence is reproducible in artificial systems, the best strategies for deriving benefit from studying biological systems and whether neural organisation is really the only place to look for biological insights.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1910483</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1910483</link>
        <title><![CDATA[Enhancing ZTMAF through federated trust learning, post quantum cryptography integration and seamless cross domain authentication handover]]></title>
        <pubdate>2026-09-25T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Jeevan Yoganand</author><author>Riya Bansal V.</author><author>Dishal L. S.</author><author>Jeipratha P. N.</author>
        <description><![CDATA[Vehicular Fog Computing (VFC) has a major role to play in facilitating low-latency and real-time communication in Intelligent Transportation Systems (ITS). Nevertheless, there exist major issues in trust management, security, and cross-domain authentication, considering the dynamic nature of vehicular networks. The Zero-Trust Mobility-Aware Authentication Framework (ZTMAF) has been proposed to overcome issues in trust management, security, and cross-domain authentication. Nevertheless, there exist major issues in distributed intelligence, trust transmission, and handover, which have been addressed in this proposed framework. The proposed framework extends ZTMAF with three major advancements: Federated Trust Learning, PQC-inspired Secure Transmission, and Cross-Domain Authentication Handover. The proposed framework uses fog nodes to train a trust model in a distributed setting, allowing for scalability and privacy preservation. Additionally, Post-Quantum Cryptography inspired encryption would be utilized for secure trust transmission, as well as to remain compliant with future post-quantum encryption standards. A token-based handover mechanism will allow vehicles to make use of previously verified trust, thus lowering latency in cross-domain authentication. The proposed framework has been validated via multiple simulation scenarios, and its performance has been evaluated against different types of attacks (e.g., stealth, on/off, Sybil). The results of the evaluations show that trust stability has improved, that detection of behavior-based attacks has improved, and that the use of the token-based handover mechanism has reduced the authentication delay by 45%. The proposed framework represents a secure and scalable solution for next-generation vehicular networks, and it sheds light on the issues associated with detecting Sybil attacks as part of the process.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1922367</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1922367</link>
        <title><![CDATA[Photonic AI for deeper understanding of the Planck epoch]]></title>
        <pubdate>2026-09-25T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Aleksandr Raikov</author>
        <description><![CDATA[This article addresses the requirements for a photonic artificial intelligence (AI) capable of supporting a deeper understanding of the Planck epoch in the origin of the Universe. The article continues the previous author’s articles, which outlined requirements for AI to support studies of protein evolution on Earth, cosmic strings (CS) detection, and a non-standard view on the photon’s nature. Modern digital generative AI cannot support its solution due to a lack of training datasets and limited energy resources for conducting adequate physical experiments. This article analyses the states of the Planck epoch, when there was no space–time, and when entropy began to emerge. It outlines additional requirements for an AI and proposes an adequate full-analogue photonic AI (PAI) that can be helpful for the research. The main idea of the PAI is to reject digitising the natural analogue signal and to use single-step Fourier convolution on many images simultaneously for representing states of the Planck epoch. To create PAI, a single holographic plate must replace a multilevel digital neural network, thereby eliminating the need for traditional multi-step machine learning. PAI is also expected to provide effective solutions to complex tasks such as Markov chains, genetic algorithms, and Kolmogorov (Fokker-Planck) equations, which are necessary for this research. The author’s convergent methodology helps make the research focused and stable. This methodology is based on fundamental thermodynamics and the inverse-problem-solving method in topological spaces. The PAI project has a starting technology readiness level: methodology, principles, and design.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1896192</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1896192</link>
        <title><![CDATA[Leakage-aware pre-event machine learning evaluation for horse-race prediction under temporal validation]]></title>
        <pubdate>2026-09-24T00:00:00Z</pubdate>
        <category>Brief Research Report</category>
        <author>Shuichi Sugiura</author>
        <description><![CDATA[Horse-racing prediction was used as a case study to develop and evaluate a leakage-aware machine-learning framework based on information available after race-entry finalization and before outcomes were known. Japanese Racing Association flat-race data obtained through JRA-VAN Data Lab. and exported using TARGET frontier JV were split temporally: 2015–2022 for training, 2023–2024 for validation, and January 5, 2025 to May 10, 2026 for independent testing. The test set contained 63,910 horse-level observations from 4,556 races. Confidence intervals were estimated using 10,000 race-level bootstrap resamples. Two horse-level binary outcome labels were evaluated: a win outcome and a JRA place-rule-compatible place outcome derived from the processed field_size variable. Compared with the augmented current_full_model, the hyperparameter-matched no-theory sensitivity-analysis model matched_no_theory_model performed better for both outcomes. For the win outcome, ROC AUC was 0.7543 (95% CI: 0.7475–0.7609) for matched_no_theory_model and 0.7293 (95% CI: 0.7224–0.7362) for current_full_model. For the JRA place-rule-compatible place outcome, ROC AUC was 0.7513 (95% CI: 0.7469–0.7558) and 0.7164 (95% CI: 0.7118–0.7212), respectively. An exploratory race-level uncertainty diagnostic was evaluated separately and was not incorporated into the horse-level prediction models. These findings indicate that the evaluated theory-score block did not improve temporal test-set performance. Domain-derived feature blocks should be evaluated incrementally through leakage audits, temporal validation, feature-block ablation, and matched sensitivity analyses before deployment.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1873285</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1873285</link>
        <title><![CDATA[ Performance evaluation of ARIMA, VAR, ANN, GRU, and LSTM for multivariate healthcare time-series prediction: a rigorous, reproducible benchmarking study with ablation analysis on MIMIC-III and eICU]]></title>
        <pubdate>2026-09-24T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Md. Tazul Islam</author><author>Muhammad Hanif</author><author>Samuel Kim</author>
        <description><![CDATA[IntroductionAccurate forecasting of physiological time-series in the intensive care unit (ICU) is critical for early-warning systems, deterioration detection, and dynamic resource allocation. Yet the comparative literature suffers from five persistent methodological deficiencies: non-unified experimental conditions, heterogeneous preprocessing, univariate classical baselines compared against multivariate neural models, single-run point estimates without variance, and underspecified prediction tasks.MethodsThis paper addresses all five deficiencies through a standardized, reproducible benchmarking study comparing Autoregressive Integrated Moving Average (ARIMA), Vector Autoregression (VAR), Artificial Neural Networks (ANN), Gated Recurrent Units (GRU), and Long Short-Term Memory (LSTM) networks on one-step-ahead (1-h-ahead) prediction of heart rate, respiratory rate, systolic and diastolic blood pressure, and oxygen saturation across two large-scale ICU cohorts (8,000 MIMIC-III and 15,000 eICU admissions). VAR is introduced as a multivariate linear baseline alongside univariate ARIMA, enabling a three-way decomposition of performance gains into multivariate modeling, nonlinear approximation, and gated temporal memory. All experiments are repeated over 10 independent seeds and reported as mean ± 95% confidence interval with Diebold-Mariano, paired t-test, and Bonferroni-corrected ANOVA significance testing. Seven categories of ablation study, a clinical alarm simulation, a leave-one-hospital-out protocol, and a subgroup fairness analysis are additionally conducted.ResultsLSTM achieves the highest performance on MIMIC-III (RMSE 0.082 ± 0.004, R2 0.940 ± 0.008, directional accuracy 90.1%±0.9%), with GRU reaching near-equivalent accuracy at 64% of training cost. The layered decomposition reveals that gated temporal memory provides the largest marginal gain (−47% to −53% RMSE over ANN), followed by nonlinear approximation (−19% over VAR) and multivariate modeling (−13% over ARIMA). Ablation identifies normalization as the single most critical preprocessing step (+79.3% RMSE without it) and a window length of l = 10 as optimal. The clinical alarm simulation demonstrates a 38.4% false alarm rate reduction for LSTM relative to threshold-only monitoring, and the leave-one-hospital-out experiment confirms cross-institutional generalizability with consistent model ranking.DiscussionSubgroup fairness analysis identifies a performance gap for severely ill patients (Sequential Organ Failure Assessment score ≥ 8, ratio 1.27) warranting targeted future work. These results establish a reusable, methodologically rigorous benchmark for ICU physiological forecasting and provide actionable model-selection guidance for clinical deployment.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1963509</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1963509</link>
        <title><![CDATA[Hybrid machine learning framework for electrical grid stability prediction using comparative model evaluation]]></title>
        <pubdate>2026-09-23T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>N. Subramanian</author><author>Albert Alexander Stonier</author>
        <description><![CDATA[IntroductionAccurately predicting electrical grid stability is essential to ensuring the reliability, resilience, and sustainable operation of modern smart energy systems. The nonlinear interactions among the grid operating parameters make stability prediction challenging, particularly when conventional machine learning models depend primarily on statistical relationships, without clearly incorporating physically meaningful characteristics.MethodsThis study presents a systematic comparative evaluation of machine-learning and deep-learning approaches for electrical grid stability prediction using the Electrical Grid Stability Simulated Dataset, comprising 10,000 observations from a decentralized four-node star-topology power system. Six conventional machine learning models - Logistic Regression, Random Forest, Gradient Boosting, XGBoost, LightGBM, and Support Vector Machine with an RBF kernel are compared with four deep-learning architectures, namely Deep Multilayer Perceptron, 1D-CNN, CNN-LSTM, and Bi-LSTM. We also develop a hybrid stacking ensemble to integrate complementary representations from high-performing boosting models through a linear meta-learner. To incorporate domain knowledge, seven physics-informed features are engineered from the original 12 grid parameters, producing a 19-dimensional feature space. Model performance is evaluated using accuracy, weighted F1-score, ROC- AUC, Matthews correlation coefficient, and class-wise precision and recall, evaluated using stratified fivefold cross-validation. SHAP-based analysis is employed to quantify feature contributions and improve model interpretability.ResultsThe proposed hybrid stacking ensemble achieves the best overall predictive performance, obtaining an accuracy of 96.85%, weighted F1 score of 0.9685, ROC-AUC of 0.9953, and MCC of 0.9319. The results demonstrate that integrating complementary ensemble learners provides more reliable stability discrimination than individual conventional and deep learning models. SHAP analysis further indicates that physics- informed interaction features, particularly τ−P interaction and power balance characteristics, provide substantial discriminatory information when combined with the original grid reaction time parameters.DiscussionThe findings demonstrate that hybrid ensemble learning combined with physics-informed feature engineering will provide accurate, interpretable, and computationally practical solutions for electrical grid stability prediction. The strong predictive discrimination and consistent cross-validation performance indicate its potential for real-time stability monitoring and decision support in smart-grid environments. The integration of domain-informed representations with data-driven learning offers a promising direction for developing more transparent and reliable intelligent grid-monitoring systems.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1938382</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1938382</link>
        <title><![CDATA[LEFTNet: low-frequency enhanced feature transmission network for dual-branch collaborative single image dehazing]]></title>
        <pubdate>2026-09-23T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Beesetti Vanitha</author><author>Chintakindi Balaram Murthy</author>
        <description><![CDATA[Image dehazing is a basic low-level vision task designed to recover high-fidelity images from hazed images. In autonomous driving, aerial surveillance, and outdoor scene recognition, image quality is significantly degraded by light attenuation due to scattering in the propagation medium, especially under hazy, foggy, or smoggy conditions. While remarkable advances have been made with existing deep learning methods, they remain constrained by over-smoothing, ineffective use of haze frequency, and insufficient consideration of multi-scale feature interactions. To meet these challenges, we introduce a novel single-image dehazing framework, called Low-Frequency Enhanced Feature Transmission Network (LEFTNet), which consists of four special modules: the Haze Suppression Block (HSB), the Dual-Branch Block (DBB), the Cross-Scale Fusion Block (CSFB), and the Channel Mining Mechanism (CMM). In the first stage, a hazy input image is decomposed into two complementary low-frequency representations: a haze-related component (IhazeL) and a frequency-domain component (IfreqL), which are jointly processed by the HSB for early, channel-selective haze attenuation. Then, the encoder extracts multi-scale hierarchical features, which are cleaned up at each stage by the Channel Mining module. The DBB's residual refinement enables intra-image global-to-local feature interactions, while the CSFB's learnable fusion coefficients enable inter-scale feature fusion. A dual fully connected bottleneck with Global Average Pooling (GAP) suppresses haze-dominant channels while enhancing structure-sensitive channels. Two symmetric branches ensure complementary representation learning, and a hybrid loss function comprising L1 loss and perceptual LSSIM loss assures both pixel-wise accuracy and perceptual fidelity. Comprehensive experiments on four benchmarks, RESIDE-6K, I-Haze, O-Haze, and NH-Haze, demonstrate that LEFTNet achieves competitive performance across the evaluated Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), Average Mean Brightness Error (AMBE), Gradient Magnitude Similarity Deviation (GMSD), and Visual Saliency-Induced Index (VSI) metrics compared with recent dehazing methods.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1906686</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1906686</link>
        <title><![CDATA[MPDRL: difficulty-aware representation learning with medical priors for vision–language pre-training]]></title>
        <pubdate>2026-09-22T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Juntao Liu</author><author>Zhuoyi Tan</author><author>Weiqi Chen</author><author>Xiaoyong Liu</author>
        <description><![CDATA[IntroductionMedical vision-language pre-training (MedVLP) learns transferable cross-modal representations by aligning chest X-ray images with radiology reports. However, most existing CLIP-style methods have two limitations. First, they treat all unpaired image-report samples within a mini-batch as negatives, which may incorrectly penalize medically similar samples as false negatives. Second, they assume that all samples contribute equally to optimization, overlooking the heterogeneous difficulty of medical data arising from factors such as disease rarity, severity, comorbidity, and report complexity.MethodsTo address these limitations, we propose Medical Prior-Driven Difficulty-Aware Representation Learning (MPDRL), which incorporates sample-level medical difficulty priors into multimodal representation learning. This method first constructs an eight-dimensional medical prior feature vector for each image-report pair to characterize sample heterogeneity. Based on these priors, it then incorporates difficulty-aware feature modulation, adaptive temperature scaling, medical prior-softened contrastive learning, and difficulty consistency regularization to reduce false-negative penalties among medically similar samples and preserve meaningful difficulty structures in multimodal representations.ResultsExtensive experiments on five medical imaging datasets across three downstream tasks demonstrate the effectiveness of MPDRL. For zero-shot classification, MPDRL achieves an AUC of 87.7% on RSNA, outperforming the strongest baseline by 0.2 percentage points, and an AUC of 84.1% on CheXpert, which is 0.1 percentage points lower than the best-performing baseline, IMITATE. On CheXpert14, MPDRL achieves the best F1-score and accuracy, reaching 26.9% and 86.5%, respectively, although its AUC of 71.8% remains lower than those of several competing methods. For zero-shot image-text retrieval on CheXpert 8 × 200, MPDRL achieves Precision@1 scores of 39.9% for image-to-text retrieval and 64.7% for text-to-image retrieval. For report generation, MPDRL achieves BLEU-1 scores of 0.508 on IU X-Ray and 0.403 on MIMIC-CXR. Importantly, clinical evaluation using CheXbert and RadGraph further shows improved medical factual consistency. On IU X-Ray, MPDRL achieves CheXbert precision, recall, and F1 scores of 0.462, 0.414, and 0.437, respectively, while obtaining an RG-F1 score of 0.256. On MIMIC-CXR, it achieves CheXbert precision, recall, and F1 scores of 0.405, 0.354, and 0.378, respectively, together with an RG-F1 score of 0.238.DiscussionOverall, these results indicate that medical prior-driven difficulty modeling can improve the robustness of MedVLP representations across diverse downstream tasks.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1857281</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1857281</link>
        <title><![CDATA[DST-SGR: a Mamba-inspired dual-stream spectro-temporal network for remaining useful life prediction of rotating machinery]]></title>
        <pubdate>2026-09-22T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>A. S. Gaurish Kumar</author><author>S. Madhavan</author><author>R. Dhanalakshmi</author>
        <description><![CDATA[IntroductionPredicting the Remaining Useful Life (RUL) of rolling bearings is essential in predictive maintenance. High-frequency vibration signals present a dual-domain challenge, since impulses are most visible in the time domain while harmonics are most visible in time-frequency analysis. Existing approaches are limited by vanishing gradients in LSTM networks and by the quadratic O(L2) cost of transformer models.MethodsWe introduce DST-SGR, a Mamba-inspired dual-stream spectro-temporal network in which selective temporal modeling is realized through optimized gated recurrent (GRU-based) state interactions with linear O(L) computational cost in a fully cross-platform PyTorch implementation. A temporal branch detects impulses from raw time-series data while a spectral branch learns harmonic signatures from an STFT spectrogram, and three Spectro-Temporal Interaction Blocks (STIBs) couple the streams through bidirectional cross-gated fusion.ResultsOn the NASA IMS bearing dataset, under a leakage-free, file-level evaluation protocol with piecewise RUL labels, DST-SGR attains a test-set MAE of 0.152 and RMSE of 0.191, with a five-fold file-level cross-validation MAE of 0.128 ± 0.004, and generalizes to IMS Tests 2 and 3 under changed fault locations.DiscussionThe model contains 131,137 trainable parameters and trains in approximately 0.97 h on an NVIDIA RTX 4060 GPU, showing that competitive prognostic accuracy is achievable with modest computational resources.]]></description>
      </item>
      </channel>
    </rss>