Incentive Framework for Cross-Device Federated Learning and Analytics With Multiple Tasks Based on a Multi-Leader-Follower Game
Yue Yu, Dawei Chen, Xiao Tang, Tiecheng Song, Choong Seon Hong, Zhu Han
Abstract
Federated learning and analytics are proposed to collaboratively learn models or statistics from decentralized data. However, the accuracy of global models relies heavily on the participation of data owners. Furthermore, for a multi-task system, data owners need to allocate limited computation resources efficiently among different local tasks. In this paper, based on the federated optimization, we use a multi-leader-follower game to build an incentive framework for the federated learning and analytics system with multiple tasks. In the proposed framework, each task owner at the upper layer decides its reward rate to incentivize data owners to participate in the federated learning or analytics. Then, based on the reward rates of tasks, each data owner (e.g., mobile device or organization) at the lower layer determines its accuracy level of local tasks' solutions. We analyze the optimal computation strategies of data owners. Because the closed-form expression of the lower-layer solution cannot be obtained, we leverage the variational inequality (VI) theory to analyze the existence of the proposed game's solution. Moreover, we provide a block coordinate descent (BCD) algorithm to solve the proposed game. Simulation results show that our proposed framework can achieve the performance improvement of the whole system.