Federated Learning (FL) is a novel distributed privacy-preserving learning paradigm, which enables the collaboration among several participants (e.g., Internet of Things devices) for the training of machine learning models. However, selecting the participants that would contribute to this collaborative training is highly challenging. Adopting a random selection strategy would entail substantial problems due to the heterogeneity in terms of data quality, and computational and communication resources across the participants. Although several approaches have been proposed in the literature to overcome the problem of random selection, most of these approaches follow a unilateral selection strategy. In fact, they base their selection strategy on only the federated server's side, while overlooking the interests of the client devices in the process. To overcome this problem, we present in this paper FedMint, an intelligent client selection approach for federated learning on IoT devices using game theory and bootstrapping mechanism. Our solution involves the design of: (1) preference functions for the client IoT devices and federated servers to allow them to rank each other according to several factors such as accuracy and price, (2) intelligent matching algorithms that take into account the preferences of both parties in their design, and (3) bootstrapping technique that capitalizes on the collaboration of multiple federated servers in order to assign initial accuracy value for the newly connected IoT devices. Based on our simulation findings, our strategy surpasses the VanillaFL selection approach in terms of maximizing both the revenues of the client devices and accuracy of the global federated learning model.
翻译:联邦学习联盟(FL)是一个新颖的分布式隐私保护学习模式,它使若干参与者(例如,Things互联网设备)之间能够合作,以培训机器学习模式。然而,选择有助于这一合作培训的参与者是极具挑战性的。采用随机选择战略将因数据质量、计算和通信资源方面各有差异而给参与者带来重大问题。虽然文献中提出了几种办法,以克服随机选择问题,但大多数这些办法都遵循单方面选择战略。事实上,他们选择战略的基础只是联盟服务器的准确性,而忽视了在这一过程中客户设备的利益。为了克服这一问题,我们在本文中提出了使用游戏理论和制导机制在IOT设备上进行联合学习的明智客户选择方法。我们的解决办法包括设计:(1) 客户IOT模型和Federederm 服务器的偏好功能,以便他们能够根据若干因素,例如准确性和价格,相互排名;(2) 智能匹配算法,考虑到在FIFL系统初步设计中,既要将我们FL的客户对FL的超级服务器的精度的精度,又要将我方对FLA的精准性原则进行。