Decentralized Federated Learning (DFL) offers a scalable paradigm for collaborative intelligence at the edge, yet its practical efficacy is severely constrained by system heterogeneity. Traditional synchronous protocols enforce rigid, lockstep aggregation barriers, where the training velocity of the entire collective is strictly dictated by the slowest straggler node, inevitably leading to significant idle time and resource underutilization. While asynchronous strategies mitigate latency, they often introduce complex pathologies, such as unbounded staleness and systemic unfairness, because high-performance nodes disproportionately bias the global model toward local data distributions, thereby marginalizing slower contributors. To rigorously reconcile these conflicting trade-offs, this work presents CAFF , a novel asynchronous communication framework for DFL that decouples local optimization from global synchronization via a topology-aware, event-driven protocol. By implementing a topology-aware cache with a strict per-neighbor replacement policy, the mechanism limits per-peer dominance by enforcing a one-slot-per-neighbor cache and exclusive replacement, preventing any peer from contributing multiple updates within a single aggregation event. Furthermore, a configurable staleness filter and a dynamic aggregation threshold ensure robust convergence stability across diverse federation topologies. Extensive empirical evaluations using MNIST, FashionMNIST, CIFAR-10, and SVHN, conducted on a high-fidelity, virtualized testbed across fully connected, star, and ring topologies, demonstrate that CAFF significantly outperforms synchronous baselines. Specifically, in dense network configurations, the framework reduces wall-clock training time by up to 39% and network traffic by up to 75%, while maintaining competitive predictive fidelity with controlled accuracy degradation. These results position CAFF as a robust and scalable efficiency-oriented solution for heterogeneous peer-to-peer learning environments.

Decentralized Federated Learning (DFL) enables nodes to collaboratively train models without a central server, introducing new vulnerabilities since each node independently selects peers for model aggregation. Malicious nodes may exploit this autonomy by sending corrupted models (model poisoning), delaying model submissions (delay attack), or flooding the network with excessive messages, negatively affecting system performance. Existing solutions often rely on rigid configurations or additional infrastructure, such as blockchains, which can incur computational overhead, introduce scalability issues, or limit adaptability. To overcome these limitations, this paper proposes RepuNet, a decentralized reputation system that categorizes threats in DFL and dynamically evaluates node behavior using metrics like model similarity, parameter changes, message latency, and communication volume. Node influence in model aggregation is adjusted based on each node’s reputation score. RepuNet was integrated into the Nebula platform and experimentally evaluated with MNIST and CIFAR-10 datasets under non-IID distributions, using federations of up to 25 nodes in both fully connected and random topologies. The evaluation considers different attack intensities, frequencies, and activation intervals, and includes comparisons with Byzantine-resilient aggregation mechanisms (Krum and Trimmed Mean), stronger structured poisoning strategies (Signed Neuron Remapping and GLL Neuro Inversion), as well as an ablation study of the exclusion threshold and a communication overhead analysis. Results demonstrate that RepuNet effectively detects and mitigates malicious behavior, achieving F1 scores above 95% on MNIST and approximately 76% on CIFAR-10. These outcomes highlight the adaptability, robustness, and practical potential of RepuNet for mitigating threats in decentralized environments.

Federated Learning (FL) offers a paradigm for collaborative AI that mitigates raw data exposure, yet the statistical heterogeneity of client data severely constrains its practical application. This non-independent and identically distributed (non-IID) setting induces client drift, leading to unstable optimization and degraded generalization, particularly for under-represented classes. Existing solutions present a difficult trade-off: iterative, regularization-based methods suffer from high communication overhead and a centralized bottleneck, while knowledge-distillation-based approaches rely on impractical artifacts, such as shared public datasets. This work introduces FedEnD , a novel framework that addresses the previous challenge through an efficient, fully decentralized architecture. FedEnD employs a two-stage protocol that decouples local specialist training from a collaborative fusion stage. Following a communication-free training phase, clients perform a one-shot peer-to-peer broadcast that shares (optionally privatized) specialist parameters and lightweight class-count statistics. Each client then uses these statistics to construct a class-distribution-weighted teacher ensemble from the received specialists. Crucially, this ensemble’s knowledge is distilled into a robust global model on each client, using only their local data as unlabeled inputs, obviating the need for a central server or auxiliary data. Extensive experiments on MNIST, FashionMNIST, SVHN, and CIFAR-10 demonstrate that FedEnD outperforms baselines, surpassing robust methods such as SCAFFOLD by +5.7% on complex datasets in pathologically skewed settings. This superior accuracy is achieved while reducing communication bandwidth by 68.6% compared to standard iterative averaging, and by up to 84% compared to gradient-correction methods like SCAFFOLD, highlighting a favorable trade-off between accuracy and communication bandwidth in decentralized learning under severe non-IID partitions.

Decentralized Federated Learning (DFL) trains models in a collaborative and privacy-preserving manner while removing model centralization risks and improving communication bottlenecks. However, DFL faces challenges in efficient communication management and model aggregation within decentralized environments, especially with heterogeneous data distributions. Thus, this paper introduces ProFe, a novel communication optimization algorithm for DFL that combines knowledge distillation, prototype learning, and quantization techniques. ProFe utilizes knowledge from large local models to train smaller ones for aggregation, incorporates prototypes to better learn unseen classes, and applies quantization to reduce data transmitted during communication rounds. The performance of ProFe has been validated and compared to the literature by using benchmark datasets like MNIST, CIFAR10, and CIFAR100. Results showed that the proposed algorithm reduces communication costs by up to$\approx 40-50 \%$while maintaining or improving model performance. In addition, it adds$\approx 20 \%$training time due to increased complexity, generating a trade-off.

Decentralized Federated Learning (DFL) enables collaborative, privacy-preserving model training without relying on a central server. This decentralized approach reduces bottlenecks and eliminates single points of failure, enhancing scalability and resilience. However, DFL also introduces challenges, such as suboptimal models with non-IID data distributions, increased communication overhead, and resource usage. Thus, this work proposes S-VOTE, a voting-based client selection mechanism that optimizes resource usage and enhances model performance in federations with non-IID data conditions. S-VOTE considers an adaptive strategy for spontaneous local training that addresses participation imbalance, allowing underutilized clients to contribute without significantly increasing resource costs. Extensive experiments on benchmark datasets demonstrate the S-VOTE effectiveness. More in detail, it achieves lower communication costs by up to 21%, 4-6% faster convergence, and improves local performance by 9-17% compared to baseline methods in some configurations, all while achieving a 14-24% energy consumption reduction. These results highlight the potential of S-VOTE to address DFL challenges in heterogeneous environments.

The metaverse has gained tremendous popularity in recent years, allowing the interconnection of users worldwide. However, current systems in metaverse scenarios, such as virtual reality glasses, offer a partial immersive experience. In this context, Brain–Computer Interfaces (BCIs) can introduce a revolution in the metaverse, although a study of the applicability and implications of BCIs in these virtual scenarios is required. Based on a limited number of publications, this work reviews the applicability of BCIs in the metaverse, analyzing the current status of this integration based on different categories related to virtual worlds and the evolution of BCIs in these scenarios in the medium and long term. This work also proposes the design and implementation of a general framework that integrates BCIs with different data sources from sensors and actuators (e.g., VR glasses) based on a modular design to be easily extended. This manuscript also validates the framework in a demonstrator consisting of driving a car within a metaverse, using a BCI for neural data acquisition, a VR headset to provide realism, and a steering wheel and pedals. Four use cases (UCs) are selected, focusing on cognitive and emotional assessment of the driver, detection of drowsiness, and driver authentication while using the vehicle. The results demonstrate the applicability of BCIs to metaverse scenarios using the proposed framework, achieving over 80% F1-score for all UCs, with performance close to 100% for detecting emotions and authenticating users. Moreover, this manuscript offers an analysis of BCI trends in the metaverse, also identifying future challenges that the intersection of these technologies will face. Finally, it reviews the concerns that using BCIs in virtual world applications could generate according to different categories: accessibility, user inclusion, privacy, cybersecurity, physical safety, and ethics.

Decentralized Federated Learning (DFL) emerges as an innovative paradigm to train collaborative models, addressing the single point of failure limitation. However, the security and trustworthiness of FL and DFL are compromised by poisoning attacks, negatively impacting its performance. Existing defense mechanisms have been designed for centralized FL and they do not adequately exploit the particularities of DFL. Thus, this work introduces Sentinel, a defense strategy to counteract poisoning attacks in DFL. Sentinel leverages the accessibility of local data and defines a three-step aggregation protocol consisting of similarity filtering, bootstrap validation, and normalization to safeguard against malicious model updates. Sentinel has been evaluated with diverse datasets and data distributions. Besides, various poisoning attack types and threat levels have been verified. The results improve the state-of-the-art performance against both untargeted and targeted poisoning attacks when data follows an IID (Independent and Identically Distributed) configuration. Besides, under non-IID configuration, it is analyzed how performance degrades both for Sentinel and other state-of-the-art robust aggregation methods.

In 2016, Google proposed Federated Learning (FL) as a novel paradigm to train Machine Learning (ML) models across the participants of a federation while preserving data privacy. Since its birth, Centralized FL (CFL) has been the most used approach, where a central entity aggregates participants’ models to create a global one. However, CFL presents limitations such as communication bottlenecks, single point of failure, and reliance on a central server. Decentralized Federated Learning (DFL) addresses these issues by enabling decentralized model aggregation and minimizing dependency on a central entity. Despite these advances, current platforms training DFL models struggle with key issues such as managing heterogeneous federation network topologies, adapting the FL process to virtualized or physical deployments, and using a limited number of metrics to evaluate different federation scenarios for efficient implementation. To overcome these challenges, this paper presents Fedstellar, a novel platform designed to train FL models in a decentralized, semi-decentralized, and centralized fashion across diverse federations of physical or virtualized devices. Fedstellar allows users to create federations by customizing parameters like the number and type of devices training FL models, the network topology connecting them, the machine and deep learning algorithms, or the datasets of each participant, among others. Additionally, it offers real-time monitoring of model and network performance. The Fedstellar implementation encompasses a web application with an interactive graphical interface, a controller for deploying federations of nodes using physical or virtual devices, and a core deployed on each device, which provides the logic needed to train, aggregate, and communicate in the network. The effectiveness of the platform has been demonstrated in two scenarios: a physical deployment involving single-board devices such as Raspberry Pis for detecting cyberattacks and a virtualized deployment comparing various FL approaches in a controlled environment using MNIST and CIFAR-10 datasets. In both scenarios, Fedstellar demonstrated consistent performance and adaptability, achieving F1scores of 91%, 98%, and 91.2% using DFL for detecting cyberattacks and classifying MNIST and CIFAR-10, respectively, reducing training time by 32% compared to centralized approaches.

Brain–Computer Interfaces (BCIs) allow the acquisition of brain activity using non-invasive techniques such as Electroencephalography (EEG). Since BCI devices do not commonly interpret the acquired EEG signals, external software applications play a critical role in the BCI lifecycle. Despite the literature offering a great variety of platforms and frameworks, they present several limitations, such as implementing old software architectures or needing more functionality to cover all phases of the BCI lifecycle. Based on these limitations, this work proposes the design and implementation of NeuronLab, a secure, multi-platform, standalone, multi-paradigm, and web-based framework that defines all BCI lifecycle phases and provides novel functionality compared to current open-source BCI software, such as sharing experiments between researchers and storing data on the cloud. This framework has been validated in two experiments common in BCI literature: P300 identification and limb movements detection. This verification has been performed based on performance metrics, such as CPU and RAM consumption, highlighting that NeuronLab is a promising solution for BCI scenarios requiring a distributed and collaborative platform for researchers and practitioners.

Abstract In response to the global safety concern of drowsiness during driving, the European Union enforces that new vehicles must integrate detection systems compliant with the general data protection regulation. To identify drowsiness patterns while preserving drivers’ data privacy, recent literature has combined Federated Learning (FL) with different biosignals, such as facial expressions, heart rate, electroencephalography (EEG), or electrooculography (EOG). However, existing solutions are unsuitable for drowsiness detection where heterogeneous stakeholders want to collaborate at different levels while guaranteeing data privacy. There is a lack of works evaluating the benefits of using Hierarchical FL (HFL) with EEG and EOG biosignals, and comparing HFL over traditional FL and Machine Learning (ML) approaches to detect drowsiness at the wheel while ensuring data confidentiality. Thus, this work proposes a flexible framework for drowsiness identification by using HFL, FL, and ML over EEG and EOG data. To validate the framework, this work defines a scenario of three transportation companies aiming to share data from their drivers without compromising their confidentiality, defining a two-level hierarchical structure. This study presents three incremental Use Cases (UCs) to assess detection performance: UC1) intra-company FL, yielding a 77.3% accuracy while ensuring the privacy of individual drivers’ data; UC2) inter-company FL, achieving 71.7% accuracy for known drivers and 67.1% for new subjects, ensuring data confidentiality between companies but not intra-organization; and UC3) HFL inter-company, which ensured comprehensive data privacy both within and between companies, with an accuracy of 71.9% for training subjects and 65.5% for new subjects.

Abstract The rise of Decentralized Federated Learning (DFL) has enabled the training of machine learning models across federated participants, fostering decentralized model aggregation and reducing dependence on a server. However, this approach introduces unique communication security challenges that have yet to be thoroughly addressed in the literature. These challenges primarily originate from the decentralized nature of the aggregation process, the varied roles and responsibilities of the participants, and the absence of a central authority to oversee and mitigate threats. Addressing these challenges, this paper first delineates a comprehensive threat model focused on DFL communications. In response to these identified risks, this work introduces a security module to counter communication-based attacks for DFL platforms. The module combines security techniques such as symmetric and asymmetric encryption with Moving Target Defense (MTD) techniques, including random neighbor selection and IP/port switching. The security module is implemented in a DFL platform, Fedstellar, allowing the deployment and monitoring of the federation. A DFL scenario with physical and virtual deployments have been executed, encompassing three security configurations: (i) a baseline without security, (ii) an encrypted configuration, and (iii) a configuration integrating both encryption and MTD techniques. The effectiveness of the security module is validated through experiments with the MNIST dataset and eclipse attacks.The results showed an average F1 score of 95%, with the most secure configuration resulting in CPU usage peaking at 68% (± 9%) in virtual deployments and network traffic reaching 480.8 MB (± 18 MB), effectively mitigating risks associated with eavesdropping or eclipse attacks.

This paper presents Fedstellar, a platform for training decentralized Federated Learning (FL) models in heterogeneous topologies in terms of the number of federation participants and their connections. Fedstellar allows users to build custom topologies, enabling them to control the aggregation of model parameters in a decentralized manner. The platform offers a Web application for creating, managing, and connecting nodes to ensure data privacy and provides tools to measure, monitor, and analyze the performance of the nodes. The paper describes the functionalities of Fedstellar and its potential applications. To demonstrate the applicability of the platform, different use cases are presented in which decentralized, semi-decentralized, and centralized architectures are compared in terms of model performance, convergence time, and network overhead when collaboratively classifying hand-written digits using the MNIST dataset.

Abstract Traffic accidents are the leading cause of death among young people, a problem that today costs an enormous number of victims. Several technologies have been proposed to prevent accidents, being brain–computer interfaces (BCIs) one of the most promising. In this context, BCIs have been used to detect emotional states, concentration issues, or stressful situations, which could play a fundamental role in the road since they are directly related to the drivers’ decisions. However, there is no extensive literature applying BCIs to detect subjects’ emotions in driving scenarios. In such a context, there are some challenges to be solved, such as (i) the impact of performing a driving task on the emotion detection and (ii) which emotions are more detectable in driving scenarios. To improve these challenges, this work proposes a framework focused on detecting emotions using electroencephalography with machine learning and deep learning algorithms. In addition, a use case has been designed where two scenarios are presented. The first scenario consists in listening to sounds as the primary task to perform, while in the second scenario listening to sound becomes a secondary task, being the primary task using a driving simulator. In this way, it is intended to demonstrate whether BCIs are useful in this driving scenario. The results improve those existing in the literature, achieving 99% accuracy for the detection of two emotions (non-stimuli and angry), 93% for three emotions (non-stimuli, angry and neutral) and 75% for four emotions (non-stimuli, angry, neutral and joy).

As recently reported by the World Health Organization (WHO), the high use of intelligent devices such as smartphones, multimedia systems, or billboards causes an increase in distraction and, consequently, fatal accidents while driving. The use of EEG-based Brain–Computer Interfaces (BCIs) has been proposed as a promising way to detect distractions. However, existing solutions are not well suited for driving scenarios. They do not consider complementary data sources, such as contextual data, nor guarantee realistic scenarios with real-time communications between components. This work proposes an automatic framework for detecting distractions using BCIs and a realistic driving simulator. The framework employs different supervised Machine Learning (ML)-based models on classifying the different types of distractions using Electroencephalography (EEG) and contextual driving data collected by car sensors, such as line crossings or objects detection. This framework has been evaluated using a driving scenario without distractions and a similar one where visual and cognitive distractions are generated for ten subjects. The proposed framework achieved 83.9% F1-score with a binary model and 73% with a multiclass model using EEG, improving 7% in binary classification and 8% in multi-class classification by incorporating contextual driving into the training dataset. Finally, the results were confirmed by a neurophysiological study, which revealed significantly higher voltage in selective attention and multitasking.