Bertrand-Game-Based On-Device Federated Learning Framework

Federated learning (FL) has been proposed as a popular learning framework to protect the users' data privacy but it has difficulties in motivating the users to participate in task training. This paper proposes a Bertrand-game-based framework for FL in wireless networks, where the model server as a resource buyer can issue an FL task, whereas the employed user equipment (UEs) as the resource sellers can help train the model by using their local data. Specially, the influence of time-varying \textit{task load} and \textit{channel quality} on UE's motivation to participate in FL is considered. Firstly, we adopt the finite-state discrete-time Markov chain (FSDT-MC) method to predict the \textit{existing task load} and \textit{channel gain} of a UE during a FL task. Depending on the performance metrics set by the model server and the estimated overall energy cost for engaging in the FL task, each UE seeks the best price to maximize its own profit in the game. To this end, the Nash equilibrium (NE) of the game is obtained in closed form, and a distributed iterative algorithm is also developed to find the NE. Simulation result verifies the effectiveness of the proposed approach.