Related Experiment Videos
BRIDGE-T: addressing temporal unreliability in federated learning for edge-enabled IoT networks
Fahmida Islam1, Adnan Mahmood2, Yingxun Wang3
1School of Computing, Macquarie University, Sydney, 2109, Australia.
Scientific Reports
|June 29, 2026
Summary
Federated Learning in IoT networks faces challenges from unreliable clients. The new BRIDGE-T framework enhances training stability and robustness by managing client participation and data drift.
Area of Science:
- * Edge computing and the Internet of Things (IoT).
- * Distributed machine learning systems.
- * Network reliability and data integrity.
Background:
- * Federated Learning (FL) in edge-enabled IoT networks is hindered by intermittent client availability and data drift, impacting global model optimization and training stability.
- * Existing FL frameworks often address these issues in isolation, neglecting their combined effects, especially during client rejoining.
- * This leads to temporal unreliability and impaired training performance in dynamic IoT environments.
Purpose of the Study:
- * To introduce BRIDGE-T, a novel reliability-aware federated learning framework designed to tackle temporal unreliability in edge-enabled IoT networks.
- * To address the coupled impact of intermittent client participation and distributional drift on FL training stability.
- * To enhance the robustness and convergence of FL models in dynamic IoT settings.
Main Methods:
- * Prototype Contrastive Drift Alignment (PCDA): Constrains representation divergence across clients with evolving non-Independent and Identically Distributed (non-IID) data.
- * Prototype Query Agreement (PQA): Assesses client reliability per round using cross-client prediction consistency on shared prototypes.
- * Reliability-Weighted Asynchronous-aware Aggregation (RWAA): Regulates client influence, mitigating updates from unreliable or misaligned clients.
Main Results:
- * BRIDGE-T demonstrates smoother convergence compared to state-of-the-art FL frameworks.
- * The framework exhibits enhanced robustness, particularly during client reintegration processes.
- * Experiments on diverse datasets (CIFAR-10, CIFAR-100, MNIST, TON-IoT) validate performance under varying intermittency and drift.
Conclusions:
- * BRIDGE-T effectively addresses the challenges of temporal unreliability in federated learning for edge-enabled IoT networks.
- * The proposed framework improves training stability and model robustness against client intermittency and data drift.
- * BRIDGE-T offers a significant advancement over existing FL approaches for dynamic IoT environments.