Scholarlink

<p>Federated Learning (FL) enables edge devices to collaboratively train a shared prediction model while retaining their training data locally, eliminating the need to centralize data in the cloud for machine learning. Efficient resource management in cloud and edge environments is a critical challenge for cloud service providers. Near accurate workload prediction plays a pivotal role in addressing this issue. In this regard, leveraging the AzurePublicDatasetV2 and Alibaba datasets a novel multi-horizon hybrid neural architecture integrating Gated Recurrent Units (GRUs), multi-head attention mechanisms, and Kalman filter is proposed. To offer client’s privacy the model is integrated with client-side differential privacy. The proposed methodology is implemented across 10 clients using a dirichlet distribution, simulating non-IID data scenarios in FL to preserve the privacy of the data produced in edge data centers and centralized setup. The results demonstrate that the model achieves an MSE of 0.00492 and an MAE of 0.0562 for a 60-minute window size in FL settings for Azure. In contrast, the centralized models yield an MSE of 0.01008 and an MAE of 0.0582 for a 60-minute window size showing consistency across different temporal window sizes, where as for Alibaba it achieves MSE of 0.0165 and an MAE of 0.0962 for 60-minute window size in FL setting.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

FedMV: Multivariate Privacy Preserving Edge Workload Forecasting using Federated Learning

  • 期刊论文

  • 发表日期：

    2026年6月30日

查看全文
  • Naga Surya Randhi,
  • B. Annappa,
  • M. R. Naveen Kumar

摘要

Federated Learning (FL) enables edge devices to collaboratively train a shared prediction model while retaining their training data locally, eliminating the need to centralize data in the cloud for machine learning. Efficient resource management in cloud and edge environments is a critical challenge for cloud service providers. Near accurate workload prediction plays a pivotal role in addressing this issue. In this regard, leveraging the AzurePublicDatasetV2 and Alibaba datasets a novel multi-horizon hybrid neural architecture integrating Gated Recurrent Units (GRUs), multi-head attention mechanisms, and Kalman filter is proposed. To offer client’s privacy the model is integrated with client-side differential privacy. The proposed methodology is implemented across 10 clients using a dirichlet distribution, simulating non-IID data scenarios in FL to preserve the privacy of the data produced in edge data centers and centralized setup. The results demonstrate that the model achieves an MSE of 0.00492 and an MAE of 0.0562 for a 60-minute window size in FL settings for Azure. In contrast, the centralized models yield an MSE of 0.01008 and an MAE of 0.0582 for a 60-minute window size showing consistency across different temporal window sizes, where as for Alibaba it achieves MSE of 0.0165 and an MAE of 0.0962 for 60-minute window size in FL setting.