Incentive-Aware Resource Allocation for Multiple Model Owners in Federated Learning

A user (model owner) in federated learning builds a learning model by aggregating local learning models trained by independent workers with their private datasets. A fundamental issue of federating learning is allocating resource from workers to the training task. As the allocation causes extra costs and overheads, workers are inherently reluctant to participate. Therefore, it is crucial to design an incentive-based resource allocation mechanism (incentive mechanism) that motivates workers to contribute their resources. Though some incentive mechanisms have been proposed for federating learning, none has devoted to the case when multiple users coexist and compete for worker service whereas a worker can contribute to multiple training tasks at the same time. For this scenario, this paper proposes an auction-based approach, where multiple users as buyers place bids for worker's service. We devise two algorithms attempting to find an auction result that maximizes social welfare, together with a pricing rule that ensures incentive compatibility and individual rationality. Simulation results show that one of the algorithms, which is based on the alternating direction method of multipliers (ADMM), outperforms the other greedy algorithm in terms of social welfare particularly when workers do not have adequate computing resource for all the training tasks.