Federated learning refers to conducting training on multiple distributed devices and collecting model weights from them to derive a shared machine-learning model. This allows the model to get benefit from a rich source of data available from multiple sites. Also, since only model weights are collected from distributed devices, the privacy of those data is protected. It is useful in a situation where collaborative training of machine learning models is necessary while training data are highly sensitive. This study aims at investigating the implementation of lightweight federated learning to be deployed on a diverse range of distributed resources, including resource-constrained edge devices and resourceful cloud servers. As a resource management framework, the FogBus2 framework, which is a containerized distributed resource management framework, is selected as the base framework for the implementation. This research provides an architecture and lightweight implementation of federated learning in the FogBus2. Moreover, a worker selection technique is proposed and implemented. The worker selection algorithm selects an appropriate set of workers to participate in the training to achieve higher training time efficiency. Besides, this research integrates synchronous and asynchronous models of federated learning alongside with heuristic-based worker selection algorithm. It is proven that asynchronous federated learning is more time efficient compared to synchronous federated learning or sequential machine learning training. The performance evaluation shows the efficiency of the federated learning mechanism implemented and integrated with the FogBus2 framework. The worker selection strategy obtains 33.9% less time to reach 80% accuracy compared to sequential training, while asynchronous further improve synchronous federated learning training time by 63.3%.
翻译:联邦学习是指对多种分布式设备进行培训,并从中收集模型重量,以获得一个共享的机器学习模式。这样,模型就能从多个地点的丰富数据来源中获益。此外,由于仅仅从分布式设备收集模型重量,这些数据的隐私受到保护;如果在培训数据高度敏感的情况下,有必要对机器学习模式进行协作培训,联邦学习是有用的;这项研究旨在调查轻量化联合学习的实施情况,以在多种分布式资源,包括资源紧张的边缘装置和资源丰富的云端服务器上进行部署;作为资源管理框架,FogBus2框架(即集装箱化分布式资源管理框架)被选为实施的基础框架。由于仅从分布式设备收集模型重量,这些数据的隐私受到保护;在FogBus2 中,需要对机器学习模式进行架构和轻度实施;此外,工人选择算法选择了一组合适的工人参加培训,以达到更高的培训时间效率;此外,这项研究通过同步和同步化的学习模型,将同步的模型与超时速化的系统学习框架一起被选为学习80年期学习周期的工人选择。