Incentive Mechanism Design for Federated Learning with Stackelberg Game Perspective in the Industrial Scenario

Federated Learning (FL) has received attention due to its data privacy-preserving mechanism. As a decentralized Machine Learning framework, all clients collectively train a model and a server gathers the model parameters to form a global model. Every client can apply the global model to tackle its problem which could bring some undesirable phenomena and affect the FL procedure, such as clients with high-quality data unwilling to invest resources for FL or clients with low-quality data obtaining the trained global by means of deception. Therefore, Incentive Mechanism Design (IMD) is a basic and important research direction in FL. However, IMD to stimulate clients to invest more resources in training models has been largely overlooked. In this paper, we designed the incentive mechanism as a Stackelberg game in which the server acts as the leader and the clients act as followers. The decision variables of the server are reward allocation and data size of the model parameters, and the decision variables of a client are training data quantity and compensation to other clients. The server has two utility functions on value increment and time difference and the clients have one utility function on payoffs. We leveraged modified NSGA-II to find the Nash equilibrium based on their decision variables and utility functions. Finally, a real industrial scenario of pre-baked carbon anode quality prediction is applied to verify the performance of the proposed method. The results show that the incentive mechanism can stimulate the clients to invest their resources for model training when compared with the scenario without an incentive mechanism, and some enlightenments are discovered