Stackelberg Game Based Resource Allocation Algorithm for Federated Learning in MEC Systems

Introducing Federated Learning (FL) into the mo- bile edge computing (MEC) system can effectively deal with delay-sensitive tasks and protect end devices (EDs) data privacy. In the process of participating in FL, the EDs will carry out a large number of local iterations and multiple rounds of communication with the MEC server to achieve a target model accuracy. These will bring delay and energy cost which may reduce EDs’ willingness to participate. In this paper, a resource allocation algorithm considering EDs incentives is proposed. We model the resource allocation of the MEC server and EDs as a two-layer Stackelberg game model and design two-layer utility functions. In EDs layer, we provide rewards to incentive EDs to contribute computing resource to achieve higher local model accuracy and weigh it against energy consumption of ED. In MEC server layer, the tradeoff between global model accuracy and system delay is conducted. We take utilities maximization as the optimization objective, and optimize the number of local iterations and bandwidth of EDs to achieve joint computing and communication resource allocation in the MEC system. Then, according to the solution of the optimization problems, we propose a resource allocation algorithm. Finally, the simulation results show that the proposed algorithm is superior to the benchmark schemes in reducing EDs’ energy consumption and system delay, which can achieve the purpose of encouraging EDs to participate.