<?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 and Artificial Intelligence section | New and Recent Articles</title>
        <link>https://www.frontiersin.org/journals/artificial-intelligence/sections/machine-learning-and-artificial-intelligence</link>
        <description>RSS Feed for Machine Learning and Artificial Intelligence 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>2026-08-18T16:34:30.433+00:00</pubDate>
        <ttl>60</ttl>
        <item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1867177</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1867177</link>
        <title><![CDATA[GSARC: a group shuffle activated redundant connection framework for post-hoc feature space adaptation in skin lesion classification on the HAM10000 dataset]]></title>
        <pubdate>2026-08-18T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Kumar Abhishek</author><author>Muralitharan Aishwaryaa Shree</author><author>Vinesh Kannaa Balaji</author><author>Rathna R</author>
        <description><![CDATA[IntroductionDeep convolutional neural networks learn rich feature representations; however, the final classification head may not fully align with this feature space after end-to-end training, leading to underutilization of discriminative information. To address this limitation, we propose a lightweight post-hoc redundant head mechanism that improves feature-space adaptation without modifying or retraining the backbone network.MethodsThe proposed approach introduces an additional shallow classification head trained from scratch on frozen feature representations augmented with frozen class logits. This design preserves training stability while enabling broader learning-rate exploration within an expanded signal space. The inclusion of frozen logits provides class-aware priors that stabilize optimization and support improved adaptation within the frozen feature manifold, allowing recovery of residual discriminative information with negligible computational overhead. The method was evaluated using a DenseNet-121 backbone enhanced with a Group–Squeeze-Excitation–Shuffle (GSSE) feature block on the HAM10000 skin lesion dataset through seven-class training and clinically relevant binary evaluation.ResultsThe proposed method improved binary classification accuracy from *91.06% to 92.42%* (+1.36 percentage points) and macro-averaged F1-score from *0.8588 to 0.8748* (+0.0160). It also outperformed the strongest baseline accuracy of *91.11%* by *1.31 percentage points*, while introducing only negligible computational overhead.DiscussionThese findings demonstrate that lightweight post-hoc redundant head learning with logit-aware feature augmentation provides a stable, computationally efficient, and practical strategy for enhancing feature-space adaptation and improving medical image classification performance without requiring modification or retraining of the backbone network.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1834688</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1834688</link>
        <title><![CDATA[Digital twin framework for personalized behavioral health optimization with Q-learning and graph-based planning]]></title>
        <pubdate>2026-08-18T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Ayan Chatterjee</author><author>Nurilla Avazov</author>
        <description><![CDATA[Promoting healthy lifestyle behaviors, including physical activity, sleep, diet, stress management, and healthy habits, requires adaptive systems capable of responding to dynamic changes in human behavior. Sustained behavioral change improves individual wellbeing, reduces disease risk, and contributes to healthier societies. However, developing personalized behavioral intervention systems is challenged by demographic heterogeneity, limited and fragmented datasets, reporting inconsistencies, and scarce high-quality labeled data. Ethical, privacy, and cost constraints further restrict the collection of large-scale longitudinal behavioral data. Consequently, there is a need for robust simulation and synthetic data generation frameworks that enable the development and evaluation of adaptive decision-making systems capable of optimizing personalized behavioral interventions over time. This study presents a digital twin framework integrated with tabular Q-learning for personalized behavioral recommendation under World Health Organization (WHO) lifestyle constraints. The framework combines synthetic behavioral data generation, reinforcement learning, and a TSP-inspired planning mechanism to investigate long-term behavioral adaptation in privacy-preserving simulated environments. The digital twin environment models user adherence variability, misreporting, dropout, and behavioral drift, enabling the evaluation of intervention strategies under realistic conditions. Experimental evaluation on synthetic populations demonstrates that Q-learning achieves competitive reward performance while maintaining favorable computational efficiency and stability compared with heuristic and reinforcement learning baselines. Statistical analysis indicates that reward differences among the evaluated methods are not significant; however, the proposed framework provides a flexible platform for adaptive behavioral recommendation and simulation-based experimentation. Furthermore, a real-time recommendation interface illustrates how simulation knowledge can be translated into actionable behavioral guidance. The proposed framework offers a scalable foundation for future digital health systems, particularly in scenarios where data scarcity, privacy constraints, and personalization requirements limit the use of real-world datasets.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1922250</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1922250</link>
        <title><![CDATA[Leakage-aware race-level upset-risk diagnostics for prospective horse-race prediction]]></title>
        <pubdate>2026-08-17T00:00:00Z</pubdate>
        <category>Brief Research Report</category>
        <author>Shuichi Sugiura</author>
        <description><![CDATA[Prospective prediction systems should not only generate individual-level predictions but also indicate when event-level conditions suggest high upset risk or difficult-to-act-on cases. This study developed and evaluated a leakage-aware race-level upset-risk diagnostic for prospective horse-race prediction under temporal validation. Japanese flat-racing data were analyzed using a fixed pre-event evaluation framework. A race-level binary upset label was constructed retrospectively from post-event results, payouts, and final popularity for label construction and evaluation only, whereas prediction features were restricted to pre-event race-structure variables. The temporal split used 2015–2022 races for training, 2023–2024 races for validation, and races from January 5, 2025 to May 10, 2026 for independent testing. Candidate-model assessment was performed using the validation split only, and the primary revised model was a prespecified unweighted standardized logistic regression model. In the temporal test set of 4,556 races, the primary revised model achieved ROC AUC: 0.6459, PR-AUC: 0.6308, Brier score: 0.2325, and log loss: 0.6567. Upset-risk stratification showed that the top 10% highest-upset-probability races had an observed upset rate of 0.6820, compared with a baseline rate of 0.5151. Conversely, the bottom 10% lowest-upset-probability races had an observed upset rate of 0.2193. These findings suggest that a race-level upset-risk diagnostic can stratify races into high-risk and lower-risk strata while remaining separate from horse-level prediction scores. The proposed framework emphasizes prediction-time information constraints, leakage prevention, validation-only candidate-model assessment, calibration assessment, and selective-use evaluation as practical components of deployment-oriented machine-learning evaluation.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1846713</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1846713</link>
        <title><![CDATA[A blockchain-empowered AutoML reputation system for anonymous customer service in SMEs]]></title>
        <pubdate>2026-08-17T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Selvalakshmi Annamalai</author><author>Logesh Ravi</author>
        <description><![CDATA[A typical manual customer service model for small and medium-sized enterprises (SMEs) is transitioning to automated customer service, using various computing resources to deliver more effective and high-quality services. The prerequisites for automated customer service include extensive data and strong data analytics capabilities. Unfortunately, because of their small size and lack of resources, many SMEs, particularly small and medium-sized companies, lack adequate data and ability. SMEs are forced to rely on external computing sources, which prevents them from developing core competencies in customer service. To address this issue, we propose a blockchain-enabled and automated machine learning (AutoML) model that uses decentralized datasets for customer services. Blockchain, a secure and transparent digital ledger, ensures the integrity and traceability of data, making it an ideal solution for storing and managing customer service data. Blockchain smart contracts allow customer services to be carried out automatically in a pipeline, producing review scoring reputation systems. The proposed scheme includes the following: model implementation, enterprise registration, review aggregation, verification, and review evaluation integrated into blockchain. To guarantee that the customer service system is traceable and that the data of each reviewing enterprise is safely, effectively, and impenetrably stored on the blockchain nodes, we first present a mechanism for review data administration and storage utilizing blockchain. Next, we provide a pipeline that efficiently integrates the crucial processes of model creation, review collection, verification, and model assessment using an AutoML model. The experimental findings show that our proposed system reliably and effectively assesses customer service in SMEs.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1826705</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1826705</link>
        <title><![CDATA[Hybrid fuzzy clustering and temporal deep learning framework for multi-parameter forecasting in industrial thermal processes]]></title>
        <pubdate>2026-08-17T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>V. Vignesh</author><author>G. V. Narendran</author><author>R. Senthil Kumar</author><author>R. Sitharthan</author>
        <description><![CDATA[Blast furnace (BF) operation involves strongly coupled thermal and chemical processes governed by nonlinear heat transfer, gas–solid reactions, and dynamic operational control. Accurate real-time prediction of key thermal and gas parameters is essential for maintaining furnace stability, improving energy efficiency, and reducing carbon emissions. However, most existing studies focus on single-parameter forecasting and fail to adequately capture the heterogeneous temporal dynamics and strong process coupling inherent in blast furnace operations. In this paper, a hybrid fuzzy clustering and temporal deep learning framework is proposed for multi-parameter forecasting. This framework integrates Fuzzy C-Means (FCM) clustering with temporal deep learning models, such as Nonlinear Autoregressive with Exogenous Inputs (NARX), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU). The dataset consists of 43,396 industrial Distributed Control System (DCS) samples collected from an operating BF. FCM is employed to identify different operating regimes and generate fuzzy membership values, which are subsequently utilized as sample weights during model training. The performance of the proposed models is evaluated using Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and coefficient of determination (R2). Among the developed models, the FCM-GRU model achieved the best overall forecasting performance, attaining R2 values of 0.602, 0.355, 0.656, and 0.912 for the prediction of Hot Metal Temperature (HMT), silicon content (Si), CO, and CO₂ respectively. The novelty of the proposed work lies in integrating fuzzy membership-based operational-state identification with temporal deep learning architectures for simultaneous forecasting of multiple blast furnace parameters. The obtained results demonstrate the feasibility of the proposed framework for real-time process monitoring and predictive decision support in blast furnace operations.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1800407</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1800407</link>
        <title><![CDATA[LATTICE: a governance-first architecture for authorized autonomous AI operations]]></title>
        <pubdate>2026-08-14T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Elias Calboreanu</author>
        <description><![CDATA[Deploying autonomous AI agents in high-consequence operational environments requires organizational authorization, yet few frameworks provide end-to-end, testable governance mechanisms suitable for such authorization decisions. This paper introduces LATTICE (Layered Agentic Triad Topology for Intelligent Coordinated Execution), a governance-first architecture that reframes the authorization question from “do we trust this AI?” to “do we trust this architecture?” The latter question is answerable through engineering validation rather than assumptions about model behavior. LATTICE enforces separation of concerns across planning, execution, and governance functions through a 1+3 Grid Cell pattern, so that no single component can both decide actions and judge compliance. The architecture implements policy-as-code enforcement with deterministic verdicts, gated execution paths that, under stated trusted-infrastructure assumptions (A1–A5), prevent unauthorized actions, confidence-based escalation to human operators, and cryptographic audit trails that preserve complete decision provenance. Empirical results characterize the AEGIS reference implementation; architecture-level properties are analytic, under stated assumptions. The governance engine is released as open source and reproduces its core results on commodity hardware: deterministic verdicts with zero deviations across 13 configurations repeated 10,000 times each, and no bypass in a 21-vector adversarial suite (0/21 observed; one-sided 95% upper bound 13.3%). In a pre-specified, planner-invariant safety evaluation (not an autonomy benchmark) across four frontier planner families (GPT-5, Claude Sonnet 4.6, Gemini, Grok-4; 4,000 trajectories), a confidence-threshold baseline's false-allow rate ranged from 0.03 to 0.998 across planners, whereas the AEGIS reference implementation admitted zero unsafe actions (false-allow 0.0, recall 1.0) invariant to the planner, at a conservative operating point that auto-allowed no action; a separate live run additionally governed real operating-system actions with zero unsafe executions. Governance latency is low and host-specific (on an Apple M4 Pro: policy evaluation p50 ≈ 6.2 μs; full gated enforcement p50 ≈ 0.7 ms including audit I/O). LATTICE provides a pathway for responsible deployment of autonomous AI in defense, critical infrastructure, and regulated industries where authorization requires verifiable governance rather than trust in AI behavior.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1826633</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1826633</link>
        <title><![CDATA[A systematic literature review exploring the application of deep learning in electric vehicles from 2015 to 2025]]></title>
        <pubdate>2026-08-11T00:00:00Z</pubdate>
        <category>Systematic Review</category>
        <author>John Vianney Ssennono</author><author>Javeed Kittur</author><author>Sabah-Ud-Din Waqar</author>
        <description><![CDATA[IntroductionElectric vehicles (EVs) are rapidly gaining popularity and global recognition, driven by their reliability, flexibility, simplicity, and scalability. This paper provides a systematic literature review of research at the intersection of electric vehicles and deep learning, aiming to identify current advancements and explore their potential for future scalability.MethodsA total of 92 publications from 2015 to 2025 were included in the final synthesis phase of the review. These works were categorized into five key themes: data-driven research on electric vehicles and deep learning, societal integration of electric vehicles, implications of electric vehicle adoption, software considerations, and challenges and solutions enabled by deep learning. Crucially, the scope of this synthesis extends into state-of-the-art frameworks spanning 2025 and 2026, evaluating deep learning’s dual footprint in vehicle-level mechanical safety systems, such as machine learning-driven brake-blending policies optimizing regenerative energy capture and fleet-level performance logistics via neural network-driven predictive maintenance optimization.ResultsThe findings for each theme and their implications for research and practice are thoroughly discussed. Additionally, a descriptive analysis of research trends shows: (1) a steady increase in publications each year; (2) a majority of contributions originating from China; (3) diverse deep learning approaches being applied to tackle various challenges within the electric vehicle industry; and (4) significant opportunities for the development, testing, and deployment of deep learning technologies and algorithms in the electric vehicle domain.DiscussionThe findings highlight the growing applicwation of deep learning across the electric vehicle domain and demonstrate significant opportunities for the continued development, testing, and deployment of deep learning technologies and algorithms to support future advancements and scalability in electric vehicles.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1880282</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1880282</link>
        <title><![CDATA[Improved graph-based model for phishing website detection using multi-level web page graphs and dynamic heterogeneous graph attention network]]></title>
        <pubdate>2026-08-11T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>S. Kavya</author><author>D. Sumathi</author>
        <description><![CDATA[Modern phishing pages are challenging to detect because they can appear like legitimate brand sites using dynamic document object model (DOM) structures, misleading visual displays, and changing hyperlinks. Most conventional blacklists and URL-based methods do not capture the relationships at multiple levels of the web page structure. In this paper, we introduce a new graph-based phishing detection method to represent a web page with a Multi-Level Web Page Graph (MLWPG). MLWPGs incorporate DOM hierarchies, rendered visual blocks, and hyperlink relations into one heterogeneous graph. We use a novel Deep Learning Method called DHGAN that has Type-Aware Attention and Dynamic Convolution to learn how to differentiate discriminative interaction amongst structural, spatial, and navigation elements. APDA will be used in feature space to enhance robustness to evasive phishing versions. Using a balanced data set of 50,000 web pages, our complete pipeline showed a 97.0% accuracy, 96.8% F1-Score, 3.0% False Positive Rate, and 95.5% Robustness to Adversaries. The proposed MLWPG-DHGAN-APDA framework outperformed all baseline models. Our experimental results demonstrate that multi-level graph models provide improvements to both detection accuracy and reliability of operation for real-time phishing defenses.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1886896</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1886896</link>
        <title><![CDATA[Predicting influenza in the post-COVID era: assessing LSTM, GRU, and transformer robustness to covariate shift]]></title>
        <pubdate>2026-08-11T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Atiqa Naeem Alam Din</author><author>Woldegebriel Assefa Woldegerima</author><author>Jianhong Wu</author>
        <description><![CDATA[Forecasting influenza has become increasingly challenging due to post-COVID disruptions in seasonality and strain circulation. This work compares the performance of Long Short Term Memory Networks (LSTM), Gated Recurrent Unit (GRU), and transformer models in forecasting influenza spread using multivariate epidemiological and environmental data, with a focus on robustness under post-COVID non-stationarity. We compare LSTM, GRU, and transformer architectures within a multivariate deep learning framework using influenza and temperature data from Ontario (2014–2025), with data split into training, validation, and testing periods. Although recurrent models outperform transformers on limited, noisy data, all architectures exhibit marked performance collapse under post-COVID non-stationarity. The GRU and LSTM track pre-COVID seasonal peaks more closely, yet both substantially under-estimate the post-COVID resurgence, indicating that none of the models generalize across the regime shift. These findings position our study as a diagnostic of how architectural inductive biases break down under covariate shift. Furthermore, this manuscript assesses how the COVID-19 pandemic affected the accuracy and performance of machine learning algorithms and notes the integration of transfer learning and attention mechanisms to improve model performance.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1784973</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1784973</link>
        <title><![CDATA[What really happens when a dev vibes with the code? An empirical study on LLM behavioral divergence in response to expressive code comments]]></title>
        <pubdate>2026-08-11T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Angela N. Johnson</author>
        <description><![CDATA[IntroductionWe investigate how expressive inline code comments written in various developer styles, functional to progressively poetic, philosophical, and misleading, affect large language model (LLM) behavior during code optimization.MethodsIn this pilot study, we used a controlledmerge sort implementation across five stylistic variants and evaluated GPT-5 and Claude Opus 4.1 under standardized console prompts, isolating the effect of embedded comment semiotic variation. Seven expert developers (three senior, four mid-level) scored model outputs against adapted ISO/IEC 25010 criteria and novel LLM suggestibility index (LSI) framework.ResultsSemiotic character of comments measurably altered code quality, with consensus-score reliability ICC(2, k) = 0.65–0.81 for six of seven dimensions; single-rater Krippendorff's α = 0.232 reflects substantial interpretive variability. Claude exhibited higher interpretive sensitivity (mean behavioral divergence 4.00; SD 1.16), while GPT-5 maintained stronger architectural fidelity (mean divergence 3.58; SD 1.26). Reflective comments (philosophical, conversational) were associated with Claude's highest maintainability scores in our panel (both M = 4.00, ~8% above stock M = 3.71), while the same philosophical comments reduced GPT-5 maintainability (M = 2.86), suggesting asymmetric model responses to expressive context.ConclusionsThese findings position inline comments as model-sensitive latent semantic prompts, with implications for AI-in-the-loop development and design of comment conventions for AI-assisted maintenance.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1884843</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1884843</link>
        <title><![CDATA[Automated evaluation of dental cavity preparation quality using deep learning and anatomically informed geometric analysis]]></title>
        <pubdate>2026-08-11T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Abdullah F. Alshammari</author><author>Bassam A. Anazi</author><author>Mahvish Khan</author><author>Hamdan A. Alshammari</author><author>Najmah A. Almowina</author><author>Yousef E. Alenezi</author><author>Saif Khan</author><author>Shafiul Haque</author><author>Ahmed A. Madfa</author>
        <description><![CDATA[BackgroundThe quality of cavity preparation critically influences the longevity and success of restorative dental treatments. Current assessment methods remain largely subjective, relying on visual inspection and examiner judgment, which are prone to variability and limited reproducibility. Although three-dimensional (3D) imaging enables quantitative evaluation, its routine use in clinical and educational settings is limited by cost, accessibility, and workflow complexity.ObjectiveThis study aimed to develop an automated, objective, and clinically interpretable framework for evaluating dental cavity preparation quality using standard two-dimensional (2D) images, with optional integration of 3D depth information.MethodsA deep learning pipeline based on enhanced U-Net architectures was developed to automatically segment cavity and cusp regions from 2D molar photographs. Anatomically informed geometric analyses were applied to quantify cavity-shape similarity, intercuspal distance, isthmus width, and cavity proportionality. Global cavity-shape conformity was assessed using Elliptic Fourier Descriptors (EFDs), enabling scale-, rotation-, and translation-invariant comparisons with reference preparations. When 3D STL data were available, cavity depth and cavity-bed smoothness were additionally quantified. These measurements were integrated into a transparent Cavity Quality Score (CQS) ranging from 1 to 10.ResultsThe cavity segmentation model achieved an internal validation Dice coefficient of 0.81 and an Intersection-over-Union of 0.74, while cusp segmentation achieved a Dice coefficient of 0.83. External validation using measurements from three independent experts demonstrated close agreement between automated predictions and expert consensus for EFD cavity-shape similarity (MAE = 1.32 percentage points; r = 0.981), pooled isthmus-width measurements (MAE = 0.03 mm; r = 0.995), pooled cusp-pair distances (MAE = 0.08 mm; r = 0.999), and cavity depth estimation (absolute error ≈ 0.01 mm).ConclusionThis study presents a hybrid, explainable artificial intelligence framework for objective assessment of dental cavity preparation using widely available 2D images. By integrating deep learning with anatomically informed geometric analysis, the proposed CQS offers a transparent and scalable tool for formative feedback in clinical and competency-based dental education. Further validation against expert summative grading is required before high-stakes implementation.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1895239</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1895239</link>
        <title><![CDATA[GA-AFedOD: gradient-aligned active federated learning for resource-aware object detection in edge industrial IoT]]></title>
        <pubdate>2026-08-10T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Zepeng Wang</author><author>Xiaogang Yuan</author><author>Jie Chen</author>
        <description><![CDATA[Visual object detection is essential for defect inspection and process monitoring in edge-deployed Industrial Internet of Things (IIoT). Yet, training accurate detectors across distributed factories faces stringent constraints on data privacy, annotation budgets, and uplink communication. Standard federated learning (FL) preserves locality but often wastes labeling resources on redundant frames and overlooks detection-specific gradient alignment when scheduling clients. To bridge this gap, we propose Gradient-Aligned Active Federated Object Detection (GA-AFedOD), a unified framework that jointly optimizes annotation selection, client participation, and model aggregation as a constrained stochastic program. A novel utility metric integrates box-level uncertainty, prototype diversity, gradient alignment, and resource pricing, enabling edge clients to perform locally guided active querying while the server solves a lightweight primal-dual problem for budget-aware client scheduling. We prove a submodular approximation guarantee for the greedy sampling rule and establish a non-convex convergence bound that explicitly captures the impact of label budgets, client drift, and compression noise. This article further clarifies the relationship with recent federated active learning and industrial detection studies, adds parameter and theory-diagnostic analyses, and distinguishes controlled simulation evidence from real-world deployment validation on industrial datasets such as RasPiDets, Electric Power Fitting Dataset (EPFD), and Diverse Insulator Dataset (DINS). Controlled simulation results show that GA-AFedOD achieves considerably higher mean average precision (mAP) while reducing both annotation costs and uplink consumption by over 40% compared with competitive baselines.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1868693</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1868693</link>
        <title><![CDATA[Fear-driven predator–prey dynamics with prey refuge: analytical framework and physics-informed neural network approach]]></title>
        <pubdate>2026-08-10T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>G. Ramraj</author><author>T. Poornima</author>
        <description><![CDATA[Ecological communities are governed not only by direct consumption but also by indirect behavioral responses triggered by perceived predation risk. Predator-induced fear substantially suppresses prey reproductive output and foraging efficiency even when lethal predation is absent, a mechanism documented across a wide range of taxa including songbirds, ungulates, and marine invertebrates. Motivated by this observation, we formulate a deterministic two-species model that simultaneously incorporates fear-mediated prey growth reduction, partial prey refuge, density-dependent intraspecific regulation, and predator self-interference. The proposed model is distinguished from existing fear–refuge frameworks by jointly embedding four ecological mechanisms within a single functional-response denominator 1 + kv + αu, producing qualitatively novel stability thresholds absent in models incorporating only subsets of these effects. Biological admissibility is rigorously established through positivity and uniform boundedness proofs. The boundedness condition cβ(1-δ)<2aη is derived from first principles by applying Sylvester's criterion to the cross-interaction quadratic form. Three ecologically meaningful equilibria are identified and their local stability is characterized via carefully re-derived Jacobian linearization and the Routh–Hurwitz criterion. Numerical experiments via the fourth-order Runge–Kutta method reveal convergence to a stable coexistence equilibrium across the explored parameter ranges, with the approach transitioning from a stable node to a stable focus as predation intensifies; no sustained oscillations are observed. A physics-informed neural network (PINN) is constructed with four hidden layers of 64 neurons each, tanh activations, Adam followed by L-BFGS training over 10,000 iterations, and 200 collocation points, achieving maximum absolute errors of 7.98 × 10−3 (prey) and 5.83 × 10−3 (predator) relative to the RK4 reference. Comparison with a data-driven neural network of identical architecture shows a fivefold accuracy improvement from the physics-informed loss. Numerical evidence for global stability is reported; rigorous Lyapunov-based analysis is identified as future work.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1881543</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1881543</link>
        <title><![CDATA[Generative AI-enhanced synthetic X-ray augmentation with gradient-based selection for battery detection in WEEE]]></title>
        <pubdate>2026-08-06T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Farhan Mahmood</author><author>George Chryssinas</author><author>Myrto Inglezou</author><author>Evangelos Katralis</author><author>Panagiotis Chatzakos</author><author>Antonios Porichis</author>
        <description><![CDATA[Automated detection of batteries in Waste Electrical and Electronic Equipment (WEEE) using X-ray imaging is critical for safe recycling, yet collecting large annotated real-world datasets remains prohibitively expensive and hazardous. This paper proposes a three-stage synthetic data pipeline to improve battery detection under limited labeled data conditions. First, dual-energy X-ray images are generated using physics-based ray-casting in Blender with automatic pixel-level annotation. Second, the synthetic-to-real domain gap is reduced using unpaired CycleGAN-based image translation. Finally, a class-conditional gradient-alignment criterion is introduced to rank synthetic training candidates by their cosine similarity to reference gradients computed from real validation data, ensuring that only the most informative synthetic samples are injected into training. The pipeline is evaluated on a real X-ray dataset of 127 scanned WEEE devices annotated across four battery categories. Under a limited-data regime of 400 real training images, our best configuration, gradient-selected CycleGAN-translated synthetic data at +30 images per class, achieves 0.621 mAP50:95 and 0.864 mAP50, surpassing the limited-data real-only baseline (0.563/0.832) evaluated on the same held-out test set, and reaching a performance level comparable to that of a full-data reference model trained on 800 real images (0.590/0.850, evaluated on a separate test split). Ablation studies confirm that gradient-based selection consistently outperforms random sampling under matched budgets, and that domain translation provides additional complementary gains. Overall, the main outcome of this research is that adding only 120 curated synthetic images yields an absolute gain of +0.058 mAP50:95, a 10.3% relative improvement over the real-only baseline under identical training and evaluation conditions. These results demonstrate that carefully curated synthetic augmentation can compensate for real data scarcity in industrial X-ray inspection, with direct implications for scalable automated WEEE recycling.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1925121</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1925121</link>
        <title><![CDATA[Correction: Exploiting explanations for model extraction via knowledge distillation and mitigation with private counterfactuals]]></title>
        <pubdate>2026-08-06T00:00:00Z</pubdate>
        <category>Correction</category>
        <author>Fatima Ezzeddine</author><author>Silvia Giordano</author><author>Omran Ayoub</author>
        <description></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1856630</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1856630</link>
        <title><![CDATA[Federated spatio-temporal graph neural network for privacy-preserving vehicle trajectory prediction in autonomous driving]]></title>
        <pubdate>2026-08-06T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Aditi Joshi</author><author>Amrutha P</author><author>Vaidehi Prajapati</author><author>Sudha Anbalagan</author><author>Suganeshwari G</author>
        <description><![CDATA[Accurate vehicle trajectory prediction plays a vital role in autonomous driving and intelligent transport systems. Deep learning models like LSTM, CNN, GNN, etc., have shown remarkable performance but often operate in a centralized setting, aggregating raw trajectory data at the server. Furthermore, the majority of models focus on either spatial or temporal features alone, but overlook the information that can be obtained by combining spatio-temporal features. This leads to major privacy concerns, issues with centralized data, and scalability problems. To overcome these challenges, we introduce a Federated Spatio-Temporal Synchronous Dynamic Graph Neural Network (Federated STSDGNN) framework for privacy-preserving and adaptive trajectory prediction. Using the highD dataset, trajectories are segmented into spatiotemporal sequences and represented as dynamic interaction graphs. Each client (vehicle or roadside unit) locally trains an STSDGNN consisting of a pre-processing module, a spatial-temporal synchronization module (GCN/GAT with GRU) and a prediction module (CNN with MLP). Clients send only model updates, which are aggregated by the server using a federated learning algorithm. This design improves privacy, achieves approximately 23.5% lower RMSE compared to the centralized STSDGNN baseline in the conducted experiments, and has the potential to support future privacy-preserving V2X (Vehicle-to-Everything) applications.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1872217</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1872217</link>
        <title><![CDATA[EdgeLane-SEG: an energy-efficient embedded edge AI framework for real-time road marking and lane lines detection with instance segmentation in ADAS and autonomous driving]]></title>
        <pubdate>2026-08-05T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Mohammed Chaman</author><author>Anas El Maliki</author><author>Wiame Bouyoussef</author><author>Abdelmounaim Belaaribi</author><author>Younes Laababid</author><author>Zouhair Sadoune</author><author>Abdelkader Mezouari</author><author>Hamad Dahou</author><author>Abdelkader Hadjoudja</author>
        <description><![CDATA[ObjectiveAccurate and energy-efficient perception of road-surface markings is essential for Advanced Driver Assistance Systems (ADAS) and autonomous driving, particularly under real-time embedded constraints. This study proposes EdgeLane-SEG, a unified framework designed to achieve high instance segmentation accuracy while maintaining low computational cost and power consumption on resource-constrained edge platforms.MethodsThe proposed framework integrates two single-stage models, YOLO11-SEG and YOLO26-SEG, for joint object detection and instance segmentation. A dedicated dataset of 10,542 annotated images with 23,420 labeled instances was constructed, covering lane markings, directional arrows, and pedestrian crossings. Both models were trained under identical conditions using a unified multi-task loss combining IoU-based regression, objectness, classification, and hybrid Binary Cross-Entropy and Dice segmentation losses. Performance was evaluated using precision, recall, F1-score, mAP@0.5, mAP@0.5–0.95, FPS, and FPS/W. Deployment was conducted on NVIDIA Jetson Nano, Raspberry Pi 5, Raspberry Pi 5 with Intel Movidius VPU, and Raspberry Pi 5 with Hailo-8 NPU.ResultsBoth models achieved high detection and segmentation performance, with mAP@0.5 exceeding 98%. YOLO26-SEG demonstrated superior inference speed and energy efficiency across all platforms, achieving higher FPS and FPS/W than YOLO11-SEG. The Hailo-8 NPU configuration achieved the best embedded performance, reaching 50–52.6 FPS and 18.85 FPS/W for YOLO26-SEG.ConclusionEdgeLane-SEG effectively balances accuracy, efficiency, and deployment feasibility. YOLO26-SEG with Hailo-8 NPU acceleration is particularly suitable for real-time embedded ADAS applications, enabling energy-efficient and reliable road perception in resource-constrained environments.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1892739</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1892739</link>
        <title><![CDATA[AdaK: adaptive KV cache budget estimation framework for analyzing long-context large language model inference]]></title>
        <pubdate>2026-08-05T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Tianjun Shao</author>
        <description><![CDATA[IntroductionThe deployment of LLMs on resource-constrained hardware is hindered by the memory-intensive KV Cache mechanism.MethodsWe propose AdaK, an adaptive KV cache budget estimation framework with three strategies: entropy-based thresholding, task-aware lookup table, and a lightweight policy network.ResultsAdaK reveals estimated KV cache reductions of up to 17.9% relative to fixed-k = 2048 baselines across 16 settings on Qwen3-4B, Qwen3-8B, and Mistral-7B.DiscussionAdaK's decoupled design enables safe budget estimation as a dynamic ceiling for downstream sparse attention kernels.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1886098</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1886098</link>
        <title><![CDATA[Trust-driven consensus reaching in human-AI hybrid large-scale group decision-making]]></title>
        <pubdate>2026-07-29T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Xinyu Wang</author><author>Xuanhua Xu</author><author>Weiwei Zhang</author>
        <description><![CDATA[IntroductionTo address the issues of insufficient trust representation, lack of feedback in opinion conflicts, and low consensus convergence efficiency in human-AI hybrid group decision-making, this paper proposes a trust-driven consensus-reaching method for human-AI collaborative decision-making.MethodsCentered on trust modeling, the proposed method integrates human experts and large language models into a unified collaborative framework. By constructing dynamic trust relationships among multiple agents, it realizes the coupled evolution of trust mechanisms and opinion dynamics. Furthermore, a differentiated opinion updating mechanism is designed based on trust propagation, and combined with consensus measurement and feedback regulation to form an iterative process from initial opinions to a stable consensus solution.ResultsBased on the Zhengzhou “7.20” rainstorm case, and in comparison with traditional weight adjustment methods and dynamic regulation methods based on deep reinforcement learning, the proposed method preliminarily validates its potential advantages in the above indicators.DiscussionUnder the conditions of this case, the proposed method exhibits a trend of achieving higher consensus quality with lower intervention costs.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1861374</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1861374</link>
        <title><![CDATA[A hierarchical federated learning framework with FedNova, game-theoretic matching, and QKD-assisted privacy for the internet of vehicles]]></title>
        <pubdate>2026-07-29T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>L. Jai Vinita</author><author>V. Vetriselvi</author>
        <description><![CDATA[The Internet of Vehicles (IoV) supports essential intelligent transportation applications but encounters challenges in federated learning (FL) due to non-independent and identically distributed (non-IID) data, vehicle mobility, resource heterogeneity, and strict privacy requirements in latency-sensitive scenarios such as misbehavior detection and accident response. Traditional FL methods, such as random client selection and standard FedAvg, often experience slow convergence and reduced performance under non-IID conditions. We introduce a hierarchical federated learning framework for software-defined vehicular fog computing. The framework incorporates FedNova (a normalized-averaging aggregation method for heterogeneous federated optimization) to produce normalized model updates under data heterogeneity, a Reward-Based Payoff Strategy (RBPS) for incentive-aware client selection, and game-theoretic vehicle-aggregator matching based on the college admissions problem. Privacy is strengthened through quantum key distribution (QKD)-assisted secure key establishment and classical gradient masking, with quantum circuit simulation used to assess future enhancements. The three-layer architecture includes vehicles, Roadside Unit (RSU)/ Base Station (BS)-level aggregators, and a Software-Defined Network Controller (SDNC) global aggregator. The framework uses both monetary and service-based incentives, such as toll exemptions, to encourage vehicle participation. Hybrid simulations using OMNeT++, Veins, SUMO, and the VeReMi misbehavior detection dataset show that the proposed approach achieves 94.8% classification accuracy [95% Confidence Interval (CI): 92.7–97.0 over 10 runs], converges in 120 rounds (33% faster than FedAvg), and reduces average latency by 29% (320 ms compared to 450 ms for FedAvg), with statistically significant improvements (p < 0.05). These gains enable faster model adaptation to evolving attacks (5–10 min shorter training cycles) and support real-time safety applications where delays above 400 ms can compromise road safety. Ablation studies confirm the complementary roles of FedNova, RBPS, and matching. Although quantum operations are currently simulated classically, the design remains compatible with future quantum hardware.]]></description>
      </item>
      </channel>
    </rss>