Incentive Mechanism Design for Vertical Federated Learning

In vertical federated learning (VFL), participants with different features of the same group of samples can train a global model cooperatively. Currently, no existing studies discuss the economic mechanism between the VFL participants. To fill this research gap, we study the incentive mechanism design with a linear reward scheme for VFL. Specifically, we model the interactions between the label owner and the data owner as a twostage Stackelberg game. In Stage I, the label owner strategically chooses its processing speed and linear reward parameter for the data owner. In response to the label owner's decisions, the data owner will choose its optimal processing speed in Stage II. By characterizing the threshold structure of the reward parameter that incentivizes the maximum processing speed from a data owner in Stage II, we can derive the equilibrium of the two-stage Stackelberg game in closed-form. Finally, our simulation results show that as the cost coefficient of the data owner increases, the label owner will increase its reward but reduce its processing speed due to the linear cost growth but concave revenue growth.