• Please read through all the templates before choosing • Pick the most relevant text template(s) from the following page and delete all others. • Edit the text as necessary, ensuring that the original incorrect text is included for the record, please see the below. • Please do not use any extra formatting when editing the templates, and only modify the red text unless absolutely necessary • Submit to Frontiers following the instructions on this page.When the original text contained incorrect information, to preserve the scientific record, please include that text when editing the below templates. For example:There was a mistake in the Funding statement, an incorrect number was used. The correct number is "2015C03Bd051.". The publisher apologizes for this mistake.The original version of this article has been updated. 2026). Are Müller glial cells gatekeepers of neuroprotection and regeneration in age-related macular degeneration? Unraveling their roles in pathophysiology and therapeutic innovation. Prog. Retin. Eye Res. 112:101471. doi: 10.1016/j.preteyeres.2026.101471" will be removed. The paragraph will now read: "The contributions collected in this Research Topic collectively indicate that the next phase of digital medicine should focus less on algorithmic novelty and more on generating clinically actionable evidence that supports routine healthcare implementation (Figure 1). Although the included studies demonstrate substantial advances across AI-enabled diagnostics, digital therapeutics, medical imaging, large language models, implementation science, and healthcare informatics, they also reveal several common priorities that should guide future research. Building upon these advances, future innovation is expected to further benefit from emerging technologies that complement the themes represented in this Research Topic, including mobile health (mHealth) (Wang et al., 2025), LLMs, generative artificial intelligence (Generative AI), wearable sensing systems (Song et al., 2026), federated learning (Fu et al., 2026), and multimodal foundation models (Wang et al., 2024b). However, their successful translation into clinical practice will depend not only on technological innovation but also on rigorous prospective validation, demonstrated clinical utility, seamless workflow integration, fairness, privacy-preserving learning, LLM safety, benchmark standardization, and appropriate regulatory oversight."The original version of this article has been updated.
Mini Han Wang, Norberto Peporine Lopes, Nuno S. Osório· Frontiers in Artificial Inte...· 0 citations
Decentralised machine learning requires authenticated transport to secure exchanged model updates. To this end, this study presents a healthcare-focused proof of concept integrating a quantum key distribution (QKD) simulation with HKDF-SHA-256, AES-256-GCM, role-based access control, replay detection, and sample-weighted federated averaging (FedAvg), and evaluates it on 42,989 medical records across 1107 synthetic patients. A trusted server decrypted authorised updates before plaintext aggregation. Centralised logistic regression achieved an F1 score of 0.7483, while a five-client FedAvg achieved 0.7333. Encrypted and unencrypted pipelines demonstrated identical model parameters. The simulated QKD protocol produced a mean quantum bit error rate of 0.2486 and successfully aborted all sessions under a full intercept constraint. Under ten independent data splits, the five-client FedAvg achieved mean F1 scores of 0.6973 (IID) and 0.6104 (non-IID). Demonstrating that transport security does not mitigate adversarial machine learning, an authenticated sign-flip attack bypassed AES-256-GCM protection and reduced the mean F1 score to 0.1185. Computationally, the QKD-assisted path averaged 0.386 ms per update compared to 0.486 ms for fresh TLS 1.3 and 0.006 ms for persistent TLS while isolating software execution from hardware constraints. Overall, this study demonstrates a reproducible proof of concept and a practical baseline for a secure machine learning approach for healthcare collaborations.
This paper presents a novel approach to differential privacy in federated learning, termed Dynamic Differential Privacy (DDP). The core idea is to adapt the privacy budget dynamically, responding to the varying sensitivity of local data. Traditional federated learning frequently utilizes a static privacy budget, which can lead to either overly conservative privacy protection, significantly impacting model accuracy, or insufficient protection when confronted with malicious adversaries. DDP addresses these limitations by iteratively adjusting the privacy budget (ε, δ) based on gradient variance and update divergence. The algorithm reduces the privacy budget when gradients exhibit high variance, indicating sensitive data, and increases it when a node provides substantially different updates, signaling potential adversarial behavior. We demonstrate that this dynamic adaptation provides a stronger privacy guarantee while maintaining acceptable accuracy levels compared to fixed-budget approaches. This work contributes to a more robust and practical implementation of federated learning in privacy-sensitive environments.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
Federated learning (FL) offers a promising approach to training machine learning models on decentralized data sources without direct data sharing. However, traditional FL methods are vulnerable to information leakage and model divergence, particularly when dealing with highly heterogeneous client datasets. This research proposes a novel framework that combines differential privacy (DP) with personalized model aggregation (PMA) to mitigate these issues. The core claim is that integrating DP safeguards client data while PMA allows for adaptation of the global model to individual client characteristics. We introduce a rigorous methodology for quantifying the trade-offs between privacy, model accuracy, and computational overhead. The proposed approach enhances both the robustness and effectiveness of FL systems, paving the way for more secure and adaptable distributed learning applications. Specifically, we formulate the aggregation process using the following notation: Let $m_i$ represent the model received from client *i*, $s_i$ be the local data used by client *i*, and $\epsilon$ and $\delta$ be the privacy parameters. The personalized aggregation function is defined as: $m_{aggregated} = \frac{\sum_{i=1}^{K} \alpha_i * m_i}{\sum_{i=1}^{K} \alpha_i}$ where $\alpha_i$ are weights determined based on client data characteristics and privacy constraints. We leverage differential privacy by adding noise to the model updates, represented as: $m_i' = m_i + \epsilon * N(0, I_d)$ where $N(0, I_d)$ denotes Gaussian noise with mean 0 and covariance matrix $I_d$ (d-dimensional identity matrix).
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
Chronic kidney disease (CKD) is a major public health concern that requires early and reliable diagnosis to reduce disease progression and associated complications. Existing machine learning approaches often rely on centralized training, limiting their applicability in healthcare environments where data privacy, heterogeneity, and resource constraints are critical concerns. To address these challenges, this study proposes a privacy-preserving federated meta-ensemble stacking framework for CKD prediction. The proposed approach integrates a hybrid preprocessing pipeline, federated learning with prediction-level aggregation, differential privacy, and SHAP-based explainability to enable secure, interpretable, and decentralized model learning. Experimental evaluation demonstrates a mean cross-validation accuracy of 98.85% with an F1-score of 98.58%. The framework maintains robust performance under Gaussian noise, achieving 97.50% accuracy, while cross-client evaluation reaches 99.25%, demonstrating strong generalization across distributed healthcare institutions. The model also exhibits excellent discrimination and calibration, achieving a ROC-AUC of 0.998, PR-AUC of 0.997, and a Brier score of 0.00038. Furthermore, communication overhead is reduced from 0.82 MB to 0.25 MB through compression and sparse updates. These findings demonstrate that the proposed framework provides an accurate, privacy-aware, and interpretable solution for decentralized CKD prediction.
Komal Kumar Napa, D. Sathyanarayanan, Raguraman Purushothaman et al.· Discover Artificial Intellig...· 0 citations
"This dataset contains the source code and evaluation logs supporting the manuscript \"FedWave: Stealthy Wavelet-Domain Backdoor Attacks in Federated Learning,\" submitted to the IEEE Transactions on Information Forensics and Security.FedWave is an anchor-aligned wavelet-subband backdoor framework for federated learning (FL). It combines three mechanisms: (i) frequency-domain trigger embedding in the LH\/HL\/HH subbands of a one-level 2D Haar DWT, restricted to the blue channel for visual stealth; (ii) subband-wise trigger decomposition with probabilistic cross-client association, governed by an association probability p; and (iii) Anchor-Based Projection (ABP), which aligns each poisoned update to a safety band centred on a locally simulated, same-round benign update norm.The release includes the FedWave implementation, the baseline attack implementations used for comparison (BadNets, DBA, FIBA-FL, Neurotoxin), the evaluation harness for the nine defenses studied in the manuscript, and the per-round metric logs from which every table and figure is produced. All experiments use ResNet-18 on CIFAR-10 and GTSRB and a lightweight SimpleCNN on EMNIST, under Dirichlet non-IID client partitioning.The input datasets (CIFAR-10, GTSRB, EMNIST) are third-party public resources and are NOT redistributed here; links to their official sources are provided under Links. These artifacts are released to support reproducibility and follow-on defense research. See README for responsible-use terms. "
Xin Ai, Yang Cao, Mengli Wei· IEEE DataPort· 0 citations
The insurance industry is undergoing a profound digital transformation driven by advances in machine learning (ML), artificial intelligence (AI), big data analytics, and cloud computing. Traditional claim prediction models, which primarily rely on demographic and historical claim information, often struggle to capture the complex behavioral patterns and dynamic risk factors that influence claim occurrence and severity. Recent developments in customer analytics, telematics, Internet of Things (IoT) devices, wearable technologies, and digital customer interactions have created opportunities to develop more accurate, adaptive, and explainable predictive models. This paper presents a comprehensive conceptual analysis of machine learning-based insurance claim prediction using customer behavioral and policy data. Drawing upon recent research published between 2018 and 2026, the study critically examines the evolution of predictive techniques, including Logistic Regression, Decision Trees, Random Forest, Support Vector Machines, Gradient Boosting, Extreme Gradient Boosting (XGBoost), LightGBM, CatBoost, Artificial Neural Networks, and Deep Learning architectures. Beyond algorithmic performance, the paper evaluates emerging issues such as data quality, class imbalance, model interpretability, algorithmic fairness, privacy preservation, regulatory compliance, and deployment challenges in real-world insurance environments. A conceptual framework is proposed to illustrate the integration of customer behavioral characteristics, policy attributes, and advanced machine learning models into an intelligent decision-support system for claim prediction. The framework highlights the importance of explainable AI, continuous model monitoring, and human-in-the-loop decision-making to enhance transparency and operational trust. The study contributes to the literature by synthesizing recent developments, identifying critical research gaps, and outlining future research directions involving federated learning, graph neural networks, multimodal data fusion, causal machine learning, and generative AI. The findings provide valuable insights for researchers, insurance practitioners, and policymakers seeking to improve underwriting accuracy, fraud detection, claims management, and customer-centric risk assessment while supporting sustainable and data-driven insurance ecosystems. Keywords: Machine Learning, Insurance Claim Prediction, Customer Behavioral Analytics, Policy Data, Explainable Artificial Intelligence (XAI), Risk Assessment, Predictive Analytics
Pragadeesh Roopchander· International Journal of Cre...· 0 citations
This paper proposes a novel approach to decentralized federated learning (DFL) that leverages secure multi-party computation (SMPC) to guarantee data privacy during collaborative model training. Traditional federated learning methods, while promoting data sharing for model improvement, inherently expose individual datasets to the central server, raising significant privacy concerns. Our framework addresses this limitation by employing SMPC protocols, enabling model updates to be aggregated securely without revealing the underlying data. This approach provides a strong privacy guarantee, combining the benefits of federated learning with robust privacy protection. We present a detailed description of the system architecture, the SMPC protocols utilized, and the mathematical formulation underpinning the aggregation process. The core claim of this work is the ability to enable collaborative model training across multiple parties without revealing individual data. The core mechanism relies on implementing a federated learning framework based on secure multi-party computation protocols, where model updates are aggregated securely without exposing the underlying data. This research represents a significant advancement in the field of privacy-preserving machine learning.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper proposes a novel approach to decentralized federated learning (DFL) that incorporates Byzantine fault tolerance (BFT) using blockchain consensus. Traditional DFL systems are vulnerable to malicious clients injecting poisoned model updates, compromising the overall model accuracy and data integrity. Our system leverages a blockchain network to provide a robust and verifiable mechanism for validating model updates. Each client's update is cryptographically signed and submitted to the blockchain. The blockchain then employs a consensus algorithm to verify the update's validity, ensuring that only legitimate updates are added to the global model. This approach significantly enhances the security and reliability of DFL, particularly in environments with untrusted participants. The core claim is that blockchain-based consensus mechanisms can provide robust BFT for DFL, while the core mechanism involves using a blockchain network to verify and validate model updates. This addresses the security vulnerabilities of traditional DFL by utilizing the inherent properties of blockchain technology. We demonstrate the feasibility and effectiveness of this system through a detailed theoretical analysis and design considerations.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
Blockchain consensus still forces a hard choice among scalability, security, and energy use. Redbelly, a leaderless Byzantine Fault Tolerance protocol, largely settles the scalability-versus-energy side of that tension, yet it accepts any cryptographically valid transaction without judging whether it is economically fraudulent or whether the node behind it is misbehaving. That blind spot is what we target. We present ML-Redbelly, a formally specified extension that attaches four learning components to the Redbelly pipeline: a LightGBM gradient-boosted fraud classifier, an Isolation Forest behavioural anomaly detector, a tabular Q-Learning agent for adaptive committee selection, and a Paillier-based federated learning aggregator that keeps model updates private. We prove that this layer leaves Redbelly’s safety and liveness intact, give pseudocode and complexity bounds for every component, and measure the system on the IEEE-CIS Fraud Detection benchmark (400,000 transactions) paired with a faithful discrete-event Redbelly simulator parameterised from measured inputs and validated against the published Redbelly deployment. LightGBM reaches an F1 of 0.783 (precision 0.858, recall 0.719, AUROC 0.963), a 34 percent relative F1 gain over the conference-baseline Random Forest at five times the inference speed. The Isolation Forest detector attains recall 0.885 at a false-positive rate of 0.047, and the Q-Learning agent settles into a stable policy within about 200 rounds across normal, bursty, and Byzantine-attack conditions. End to end, the framework sustains 48,844 TPS on 32 validators (mean over 30 seeds), and because the leaderless superblock commits every proposer’s block in parallel, this throughput advantage over leader-based BFT grows with the validator count (5.0 times PBFT and 2.9 times HotStuff at 32 validators). The learning layer costs only about 4 percent in throughput, since the measured ML inference is small next to the geo-distributed consensus round. Per-transaction energy is comparable across BFT protocols, being dominated by signature verification, and is orders of magnitude below proof-of-work chains, which expend energy on mining. One federated update epoch takes 36 s across 10 nodes with 2048-bit Paillier keys and reconstructs gradients with negligible error. All performance figures are emergent outputs of the discrete-event simulation, which reproduces the published Redbelly benchmark to within a conservative factor of about 1.7. Taken together, these results outline a simulation-validated design for making consensus intelligent as well as fast and identify the steps needed toward real-cluster deployment.
F. Ben Fredj Ismail, Khadija Louzaoui, Khalid Benlhachmi· Computers· 0 citations
The heterogeneity of data across Industrial Internet of Things (IIoT) devices poses significant challenges to federated learning-based intrusion detection systems, where the Non-IID data distribution leads to poor detection performance, particularly on rare attack classes. To address these issues, this letter proposes FedMASA, a deep reinforcement learning-assisted federated learning framework with minority-aware selection and aggregation for Non-IID IIoT intrusion detection. Specifically, we formulate client selection as a Markov Decision Process and employ Deep Deterministic Policy Gradient to dynamically select optimal clients based on real-time state observations. A minority-aware aggregation mechanism with class-balanced weighting is designed to amplify the influence of scarce attack classes while penalizing high-latency clients. Extensive experiments on the Edge-IIoTset and ToN-IoT datasets demonstrate that FedMASA consistently surpasses the standard FedAvg baseline across all Non-IID settings. In the most challenging $\alpha=0.1$ scenario, FedMASA outperforms FedAvg by 2.96\% in accuracy and 8.87\% in Macro-F1 on Edge-IIoTset, and by 4.24\% and 10.34\% on ToN-IoT, respectively.
Xiaofei Huang, Fei Shu, Jaydar Jingus et al.· 0 citations
"This dataset contains Ultra-Wideband (UWB) ranging and localization measurements collected for the evaluation of adaptive and federated learning methods for indoor positioning in challenging environments. The data were acquired in a controlled two-room laboratory deployment designed to capture realistic line-of-sight (LOS), non-line-of-sight (NLOS), multipath, and location-dependent ranging errors. Measurements include UWB anchor-to-tag distance estimates, corresponding reference positions, anchor information, and associated localization data suitable for training and evaluating range-error compensation and positioning algorithms.The dataset was developed in support of the study \u201cFederated Adaptive UWB Localization with Random-Forest Error Compensation\u201d and enables the investigation of centralized, distributed, and privacy-preserving machine-learning approaches for UWB localization. In particular, it can be used to study anchor-specific ranging-error correction, adaptive anchor weighting, geometric localization, random-forest regression, federated learning, and robustness to spatially nonuniform propagation conditions. The data are intended to facilitate reproducible research and comparative evaluation of indoor localization techniques for Internet-of-Things, industrial, edge-computing, and cyber-physical-system applications."
Christos Sad, Kostas Siozios· IEEE DataPort· 0 citations
A new machine-learning framework aims to improve the success rate of computational protein design while moving away from results that reproduce sequences found in nature.
MIT News · Artificial Intelligence· news.mit.eduAug 24, 2026
A new method for surgically removing training examples from a model reveals that as datasets grow, the link between what a model learns and what it produces dissolves.
