Incentive Mechanism for Heterogeneous Federated Learning With Budget Constraint

Federated learning (FL) has recently received considerable attention due to its capability of collaboratively training machine learning models among clients, without compromising data privacy. However, encouraging client participation remains challenging, particularly in heterogeneous environments where clients possess different training capabilities. To address this issue, this paper formulates a heterogeneous-aware incentive problem in FL, aiming at optimizing both training intensities of selected clients and the training time of each communication round. We design an efficient auction-based mechanism that selects suitable participants and determines their compensation within a fixed budget in each communication round. This approach ensures several desirable properties, including individual rationality, budget balance, truthfulness and computational efficiency. Notably, the proposed mechanism guarantees an approximation ratio of at least one-third of the optimal solution. Simulations results based on three widely adopted real-world datasets validate its effectiveness in enhancing FL performance, compared with baseline algorithms.