federated learning

Integrating Particle Swarm Dynamics into Federated Learning: A Promising Hybrid to Counter Client Drift under Non-IID Conditions

Federated learning (FL) enables distributed model training without centralizing raw data, making it attractive for privacy-sensitive Internet of Things (IoT) environments.However, conventional FL algorithms suffer from slow convergence and degraded performance under statistically heterogeneous (non-IID) data distributions due to client drift.This paper proposes a Hybrid Federated-Swarm Optimization (HFSO) algorithm, which integrates particle swarm optimization (PSO) dynamics-velocity smoothing, personal-best memory, and global-best coordination-into the client update rule of stochastic gradient descent.The primary objective is to accelerate convergence and improve robustness under severe non-IID conditions without increasing per-round communication payload.Experimental evaluation is conducted using a unified PyTorch implementation on CIFAR-10 (α=0.1, 500 rounds, 10 clients/round) and FEMNIST for ablation and sensitivity studies.At round 500, the accuracy of HFSO on CIFAR-10 is 74.84% ± 0.62 while FedAvg is 62.44% ± 1.15 with the same settings.This final accuracy differs from but follows threshold crossing speed; while the time required for the initial threshold crossing varies between 180 ± 8 rounds on average across each of 5 independent runs (seeds {42, 123, 256, 512, 1024}), it is shown that the time when the threshold is crossed remains below 75% accuracy by round 500 as a result of non-IID-induced variance among clients.The pattern of transient-crossing is the same in FedAvg, reaching 75% accuracy at 320 ± 15 rounds on average, and then decreasing to the reported accuracy level within the budget of 500 rounds.Under the same CIFAR-10 conditions, HFSO outperforms recent heterogeneity-aware methods (FedNova, MOON, CCVR) within this specific experimental setup.Distributional analysis shows improved worst-client accuracy and reduced inter-client variance.An empirical privacy evaluation under a limited, honest-but-curious, white-box threat model shows reduced attack success rates relative to FedAvg, though no formal differential privacy guarantees are claimed.Additional memory overhead is approximately 3× (storing velocity and personal-best vectors), and computational overhead is 10-15% per local epoch.These results indicate that swarm-based coordination is a promising hybrid direction for improving convergence stability and communication efficiency in FL under non-IID conditions, within the specific experimental contexts evaluated (CIFAR-10/FEMNIST, α=0.1, 500 rounds).Generalization beyond these settings requires further validation.

Execution Governance 5.0 Research Architecture: From Governed Effect Fabrics to Governed Effect Regimes

Execution Governance 5.0 (EG5) Research Architecture v0.1.4 proposes a candidate major-version research direction extending Execution Governance from authority-preserving Governed Effect Fabrics to the time-evolving Governed Effect Regime that determines how such fabrics, authority roots, policies, comparators, composition rules, reconciliation rules, adaptation mechanisms, and evidence requirements may themselves be created, changed, combined, suspended, or replaced. The central research question is: Who authorizes a material change to the governance system that determines what counts as authorized? EG5 treats a material governance-state transition as consequential when it can alter the future admissible effect space or the authority, semantic, commitment, reconciliation, or evidence rules applied to future effects. Its candidate governing principle is: Governance may evolve, but no governance change may create the authority that legitimizes itself. A companion composition principle is: Valid governed fabrics do not imply a valid governance composition. EG5 retains the effect-centered discipline of earlier Execution Governance generations and does not introduce a new source of normative authority. It does not claim to invent administrative authorization, policy administration, compositional authorization, recursive governance, governance-of-governance, governed runtime mutation, learning/authority separation, non-widening composition, cryptographic authorization proofs, or evidence-chain composition. These areas have substantial antecedent and adjacent work. The proposed research distinction is narrower: the effect-centered conjunction of authorization requirements at a material governance-state boundary. The candidate EG5-Core v0.1 defines six Recursive Integrity properties: Governance-Change Authority Provenance Non-Self-Authorization Authority Composition Closure Multi-Root Reconciliation Integrity Consequence-Bounded Adaptation Evolution-Witnessed Closure The accompanying deterministic executable guard-ablation harness provides six hand-constructed minimal counterexamples, one for each property. In each named scenario, omission of the relevant property admits the bad state while the complete candidate guard blocks it. A separate regression confirms that participating EG4 fabric validity remains a non-substitutable prerequisite. These results are executable falsification evidence only. They are not bounded model checking, exhaustive state-space exploration, formal proof, independent reproduction, certification, production assurance, or evidence of governance completeness. The first proposed implementation profile is the Minimal Federated Governance-Evolution Profile v0.1, designed around two independently administered EG4-class digital fabrics, two authority roots, one separate verifier, one reversible synthetic cross-fabric effect, one material governance change, explicit composition and reconciliation rules, and bounded consequence-feedback semantics. The principal next evidence milestone is to demonstrate an executable case in which: Fabric A is EG4-valid.Fabric B is EG4-valid.Their composite effect is not authorized.EG5 correctly blocks the composition. Stable EG5.0 status is intentionally not claimed by this publication. It should be earned through a profile-bounded formal model, two-domain implementation, public reproducibility, and independent reconstruction. EG4 baseline:KU, H. W. (2026). Execution Governance 4.0: From Authorization-Bound Execution to Governed Effect Fabrics (Version 0.3.6.2). Zenodo.https://doi.org/10.5281/zenodo.22157731 Research programme: https://executiongovernance.org Status: Independent research and pre-standardization candidate. This publication does not constitute a standard, certification scheme, legal authorization determination, production-safety claim, or proof of governance completeness.

Execution Governance 5.0 Research Architecture: From Governed Effect Fabrics to Governed Effect Regimes

Execution Governance 5.0 (EG5) Research Architecture v0.1.4 proposes a candidate major-version research direction extending Execution Governance from authority-preserving Governed Effect Fabrics to the time-evolving Governed Effect Regime that determines how such fabrics, authority roots, policies, comparators, composition rules, reconciliation rules, adaptation mechanisms, and evidence requirements may themselves be created, changed, combined, suspended, or replaced. The central research question is: Who authorizes a material change to the governance system that determines what counts as authorized? EG5 treats a material governance-state transition as consequential when it can alter the future admissible effect space or the authority, semantic, commitment, reconciliation, or evidence rules applied to future effects. Its candidate governing principle is: Governance may evolve, but no governance change may create the authority that legitimizes itself. A companion composition principle is: Valid governed fabrics do not imply a valid governance composition. EG5 retains the effect-centered discipline of earlier Execution Governance generations and does not introduce a new source of normative authority. It does not claim to invent administrative authorization, policy administration, compositional authorization, recursive governance, governance-of-governance, governed runtime mutation, learning/authority separation, non-widening composition, cryptographic authorization proofs, or evidence-chain composition. These areas have substantial antecedent and adjacent work. The proposed research distinction is narrower: the effect-centered conjunction of authorization requirements at a material governance-state boundary. The candidate EG5-Core v0.1 defines six Recursive Integrity properties: Governance-Change Authority Provenance Non-Self-Authorization Authority Composition Closure Multi-Root Reconciliation Integrity Consequence-Bounded Adaptation Evolution-Witnessed Closure The accompanying deterministic executable guard-ablation harness provides six hand-constructed minimal counterexamples, one for each property. In each named scenario, omission of the relevant property admits the bad state while the complete candidate guard blocks it. A separate regression confirms that participating EG4 fabric validity remains a non-substitutable prerequisite. These results are executable falsification evidence only. They are not bounded model checking, exhaustive state-space exploration, formal proof, independent reproduction, certification, production assurance, or evidence of governance completeness. The first proposed implementation profile is the Minimal Federated Governance-Evolution Profile v0.1, designed around two independently administered EG4-class digital fabrics, two authority roots, one separate verifier, one reversible synthetic cross-fabric effect, one material governance change, explicit composition and reconciliation rules, and bounded consequence-feedback semantics. The principal next evidence milestone is to demonstrate an executable case in which: Fabric A is EG4-valid.Fabric B is EG4-valid.Their composite effect is not authorized.EG5 correctly blocks the composition. Stable EG5.0 status is intentionally not claimed by this publication. It should be earned through a profile-bounded formal model, two-domain implementation, public reproducibility, and independent reconstruction. EG4 baseline:KU, H. W. (2026). Execution Governance 4.0: From Authorization-Bound Execution to Governed Effect Fabrics (Version 0.3.6.2). Zenodo.https://doi.org/10.5281/zenodo.22157731 Research programme: https://executiongovernance.org Status: Independent research and pre-standardization candidate. This publication does not constitute a standard, certification scheme, legal authorization determination, production-safety claim, or proof of governance completeness.

Distributed Graph Learning via Federated Bayesian Networks

This paper presents a novel approach to distributed graph learning utilizing Federated Bayesian Networks (FBNs). The core challenge in training large graph neural networks (GNNs) lies in the substantial computational resources required, often necessitating centralized training environments. Federated Bayesian Networks offer a decentralized solution, enabling learning across multiple clients without direct data sharing. The proposed method involves local training of Bayesian Networks on individual client graph subsets, followed by parameter aggregation by a central server to refine a global Bayesian Network model. This architecture addresses the limitations of traditional GNN training while prioritizing data privacy and mitigating computational demands. The key innovation lies in the synergistic combination of federated learning principles with the probabilistic inference capabilities of Bayesian Networks, resulting in a robust and scalable framework for distributed graph learning. This approach demonstrates the potential for efficient learning from decentralized graph data sources.

DAP-FedGL: A decentralized adaptive protection framework for client-level privacy in federated graph learning

Research on Privacy Protection of Digital Twin Intelligence Based on Big Data in 5G System Security

Digital twin (DT) technology is becoming a foundational component of next-generation 5G and 6G networks, enabling real-time monitoring, predictive analytics, and collaborative research. However, integrating DTs with big data and AI introduces significant privacy, security, and governance challenges. This article presents a privacy-preserving DT intelligence framework that integrates federated learning (FL) and differential privacy (DP) within a secure, standards-based architecture. The design follows ISO/IEC 27001 for information security management, complies with EU GDPR principles for data minimization and consent, and aligns with 3GPP TS 33.501 security procedures for 5G networks. The architecture combines TLS-secured data ingestion, blockchain-backed audit trails, and policy-driven compliance monitoring to ensure trusted DT operations without centralizing sensitive data. By mapping key threat vectors to corresponding international standards, the framework offers a practical blueprint for secure and interoperable DT deployments. The article concludes with deployment considerations and a roadmap for extending existing standards to support AI-native digital twins in future 6G environments.

Efficient Backdoor Mitigation in Federated Learning With Contrastive Loss

The rapid adoption of Internet of Things (IoT) devices has accelerated the need for privacy-preserving machine learning techniques, such as federated learning (FL). However, the decentralized and collaborative nature of FL makes it vulnerable to backdoor attacks, where adversaries locally update their malicious models before contributing to the global aggregation, subtly injecting backdoors without degrading the normal performance. An affected model behaves as expected during regular operations but exhibits malicious behavior when an embedded trigger is presented. In this article, we propose a novel self-supervised contrastive-learning-based approach to detect and mitigate backdoor attacks in FL within IoT environments. Unlike conventional reverse-engineering methods that iterate through each class in the dataset to reconstruct triggers, our approach directly regenerates triggers from compromised global models without class iteration. This is achieved by comparing last-layer feature representations of a potentially compromised model with those of a relatively clean model under the guidance of contrastive loss. The reverse-engineered trigger is then leveraged to patch the global model and remove the backdoors. We evaluate our method on three benchmark datasets under two federated backdoor attack scenarios, simulating IoT device collaborations. Extended experiments are also conducted on a transformer-based model and two mitigation methods to assess the robustness of our approach. Our results demonstrate that while traditional reverse-engineering techniques are effective in centralized settings, they struggle to detect backdoors in FL. Comparatively, our method is resilient against backdoor attacks across various settings. In addition, our method is more time-efficient because of its capability of generating the backdoor trigger directly without iterating through all classes.

A Blockchain-Based Federated Learning Approach for Electricity Theft Detection Through Dual-Verification

Malicious clients participating in data collection and interaction may launch attacks such as model and data poisoning to degrade the performance of the global model and conceal their electricity theft behaviors. Although existing studies have introduced blockchain technology to achieve decentralization, they still suffer from limited pre-aggregation validation dimensions. To address these issues, this paper proposes a blockchain-based federated learning approach with dual-verification (BFL-DV) for electricity theft detection. In the pre-aggregation stage, a multi-metric reputation-based consensus committee verification strategy is designed, which effectively mitigates the impact of malicious participants. In the post-aggregation stage, a dynamic threshold-based blockchain verification strategy is developed to counter security risks during the transmission process, which can refuse malicious global updates adaptively. Experimental results demonstrate that BFL-DV can accurately reduce the impact of all malicious clients under the data poisoning attack. Notably, across various proportions of malicious clients, the proposed framework achieves an average AUC improvement of 32.68% compared with SOTA methods, demonstrating its consistent performance advantage.

A Spatiotemporal Coupling-Based Clustered Federated Learning Scheme for Low Latency Digital Twin Within Heterogeneous IIoT

Data-driven deep learning (DL) techniques have increasingly been employed to construct digital twin (DT) models for the intelligent industrial Internet of Things (IIoT) systems. Within DTs, federated learning (FL) offers a decentralized framework that enables distributed entities to collaboratively update global models without sharing raw data. Clustered FL (CFL) further enhances training efficiency by grouping clients, thereby reducing communication overhead and accelerating model convergence. However, data heterogeneity arising from spatial distribution differences and system heterogeneity, resulting in straggler clients, jointly hinder the convergence and efficiency of CFL. The interplay of these factors introduces spatiotemporal coupling, which further degrades model training. To address these challenges, we propose a spatiotemporal coupling-based CFL scheme that jointly optimizes client clustering and aggregation strategies to minimize overall training latency. A semisynchronous aggregation mechanism is introduced, allowing clients to update at different frequencies based on their delay tiers. Furthermore, client clustering is performed according to location similarity to improve convergence, while clients with higher delay tiers and greater data diversity are prioritized for cluster head selection. To mitigate the impact of imbalanced cluster sizes under data heterogeneity, a balanced matching optimization is formulated to evenly distribute remaining clients to the nearest cluster heads. Within each cluster, adaptive bandwidth allocation is employed to satisfy delay-tier constraints and shorten communication rounds. Extensive simulations on CIFAR-10 and Fashion-MNIST with nonindependent and identically distributed settings show that the proposed scheme can reduce the total training latency by up to 38.71% and 8.87%, respectively, to reach a fixed target accuracy, while achieving comparable model accuracy to existing baselines. These results confirm the effectiveness of the proposed scheme in heterogeneous IIoT environments.

CSFL: Communication-Efficient Semi-Asynchronous Federated Learning Method in Resource-Constrained Edge Computing

Federated learning (FL) is a distributed machine learning (ML) paradigm that has been widely used to train ML models on massive amounts of data in edge computing (EC) environments. However, FL faces significant challenges from device heterogeneity, edge dynamics, and limited communication resources. To address these challenges, we propose a communication-efficient semi-asynchronous FL (CSFL) framework. First, the work introduces a threshold adaptive gradient compression (TAGC) algorithm, which can reduce redundant communication rounds and accelerate model convergence by appropriately increasing local computation. Second, we propose an adaptive weight adjustment mechanism (AWAM), which employs a staleness-based decay function and, based on varying data distributions, sets different weight coefficients to mitigate the impact of statistical and system heterogeneity. To tackle edge dynamics, a dynamic node selection algorithm based on deep reinforcement learning (DRL) is proposed. This algorithm enables adaptive adjustment of the number of local models participating in global model aggregation according to environmental changes. Finally, we analyze the convergence bound of CSFL theoretically and conduct extensive experiments on classical datasets to demonstrate the effectiveness of our algorithm. Compared with baseline algorithms, the experimental results indicate that CSFL can effectively decrease bandwidth resource consumption and total training time during the training of edge intelligence models across various datasets and data distributions.

Hybrid Distributed Learning With Knowledge Distillation for Resource-Efficient Intrusion Detection in Distributed Networks

A novel AI-driven distributed NIDS that considers the computing capabilities of resource-constrained nodes while enabling efficient learning in distributed environments is proposed and can achieve accuracy comparable to a centralized model while reducing local computational overhead and maintaining stable convergence under realistic data distribution scenarios.

Twin2Twin: Digital Twin–Driven Orchestration of Federated Learning in Medical IoT Environments

Sixth-generation (6G) networks are expected to integrate intelligence as a native capability, enabling advanced verticals such as digital health (eHealth) supported by large-scale Internet of Medical Things (IoMT) deployments. In this context, Federated Learning (FL) is emerging as a promising paradigm for collaborative model training, allowing distributed medical devices to learn from data while preserving privacy. Nevertheless, conventional approaches such as Federated Averaging (FedAvg) often suffer from unstable convergence and performance degradation in non-independent and non-identically distributed (non-IID) environments, especially under partial and resource-constrained client participation. To address these challenges, we propose a Digital Twin (DT)-enabled FL framework in which the server evolves from a passive aggregator into an active orchestrator of the learning process. The DT consists of a virtual representation of participating devices and their learning dynamics, thereby enabling an informed selection of clients and coordinated updates based on both resource conditions and predictive uncertainty. Building on this orchestration layer, we introduce Twin2Twin (T2T), a hierarchical aggregation strategy that groups clients into similarity-based cohorts and combines their updates in a structured manner, improving training stability in dynamic environments. Specifically, the proposed framework integrates: (i) DT-driven client orchestration that accounts for both device heterogeneity and data informativeness, and (ii) a similarity-aware aggregation mechanism that reduces update variability across training rounds. Experimental results demonstrate that the proposed strategy achieves faster and more stable convergence compared to traditional FL, highlighting the potential of DT-driven orchestration to support reliable and scalable learning in IoMT systems deployed in 6G networks.