Online training data acquisition for federated learning in cloud-edge networks
Online training data acquisition for federated learning in cloud-edge networks
复制标题
DOI：
10.1016/j.comnet.2023.109556
复制
发表时间：
2023-01
期刊：
影响因子：
--
通讯作者：
Konglin Zhu;Wentao Chen;Lei Jiao;Jiaxing Wang;Yuyang Peng;Lin Zhang
中科院分区：
文献类型：
--
作者：
Konglin Zhu;Wentao Chen;Lei Jiao;Jiaxing Wang;Yuyang Peng;Lin Zhang
联邦学习（FL）是一种利用不同数据源进行协作模型训练，同时维护用户隐私的有效方法。为了提高联邦学习的准确性，需要足够的数据。然而，所有数据源很难为联邦训练的每次迭代获取足够的数据。在本文中，我们研究了云边网络中的数据获取问题，其中边缘从用户处获取数据并进行本地训练，而云则聚合本地模型。由于用户获取的数据具有不确定性，在线为FL做出数据采集决策并不容易。更糟糕的是，由于数据成本未知，很难做出宏观时间尺度的决策。我们提出了一种联邦学习的两个时间尺度的在线调度来应对数据获取的不确定性。该方法通过精心设计的Lyapunov虚拟队列学习系统的经验状态信息，并以在线方式协调不同时间尺度的数据采集，最小化联邦学习的数据采集成本，减少数据传输延迟，加快联邦学习的收敛速度。严格的理论分析表明，所提出的两时间尺度 Lyapunov 优化算法具有强大的性能保证，而广泛的跟踪驱动的实验结果表明，该算法比现有基准实现了显着的性能提升。
Federated learning (FL) is an effective approach to exploiting different data sources for collaborative model training while maintaining the privacy of users. Adequate data is necessary to improve the accuracy of the federated learning. However, it is difficult for all data sources to acquire sufficient data for each iteration of federated training. In this paper, we study the data acquisition problem in cloud–edge networks, where edges acquire data from users and conduct the local training while the cloud aggregates the local models. Since data acquired by users with uncertainty, it is not easy to make the data acquisition decision for the FL online. Even worse, due to the unknown cost of data, it is difficult to make a macro-timescale decision. We propose a two-timescale online scheduling for federated learning to confront the uncertainty of the data acquisition. By learning empirical state information of the system with a carefully designed Lyapunov virtual queue and coordinating the data acquisition in different timescales in an online manner, the proposed approach minimizes the data acquisition cost of federated learning, reduces the data transmission delay and accelerates the convergence speed of federated learning. Rigorous theoretical analysis shows strong performance guarantees of the proposed two-timescale Lyapunov optimization algorithm and extensive trace-driven experimental results suggests that the algorithm achieves outstanding performance gains over existing benchmarks.