Incentive Mechanisms in Federated Learning and A Game-Theoretical Approach

Federated learning (FL) represents a new machine learning paradigm, utilizing various resources from participants to collaboratively train a global model without exposing the privacy of training data. The learning performance critically depends on various resources provided by participants and their active participation. Hence, it is essential to enable more participants to actively contribute their valuable resources in FL. In this article, we present a survey of incentive mechanisms for FL. We identify the incentive problem, outline its framework, and categorically discuss the state-of-the-art incentive mechanisms in Shapley value, Stackelberg game, auction, contract, and reinforcement learning. In addition, we propose three multi-dimensional game-theoretical models to study the economical behaviors of participants and demonstrate their applicability in cross-silo FL scenarios.