Incentive Mechanism Design for Multi-Round Federated Learning With a Single Budget

Federated learning (FL) is a popular distributed learning paradigm. In practical applications, FL faces two major challenges: (1) Participants inevitably incur computational and communication costs during training, which may discourage their participation; (2) The local data of participants is usually non-IID, which significantly affects the global model's performance. To address these challenges, in this paper, we model the FL incentive processas a budget-constrained cumulative quality maximization problem (BCQM). Unlike most existing works that focus on a single round of FL, BCQM fully encompasses the entire multi-round FL process with a single budget. Then, we propose a comprehensive incentive mechanism named Reverse Auction for Budget-constrained nOn-IID fedeRated learNing (RABORN) to solve BCQM. RABORN covers the entire FL process while ensuring several desirable properties. We also prove RABORN's theoretical performance. Moreover, compared to baselines on real-world datasets, RABORN exhibits significant advantages. Specifically, on MNIST, Fashion-MNIST, and CIFAR-10, RABORN achieves final accuracies that are respectively 2.94%, 5.94%, and 21.75% higher than baselines. Correspondingly, when the final model accuracies on MNIST, Fashion-MNIST, and CIFAR-10 converge to 80%, 70%, and 40%, RABORN reduces communication rounds by over 33%, 45%, and 74% compared to baselines, while increasing the remaining budget by over 30%, 19%, and 130%, respectively.