Long-term Incentive Mechanism for Federated Learning: A Dynamic Repeated Game Approach

Federated learning (FL) is capable of using the local data sets from large-scale nodes for distributed model training. In FL tasks, training and updating are usually repeated ping-pong processes, in that the model training process between devices (workers) and task publisher (TP) needs to be repeated for multiple rounds towards the global model convergence. However, a worker is typically selfish to save its local resource, and even with an incentive mechanism in place at the beginning of model training, a worker may not be honest to participate in all training rounds, leading to poor performance in global model convergence. To enable long-term cooperation in FL, however, has rarely been considered in the existing literature which motivates our work. In this paper, the multi-round FL is modeled as a dynamic repeated game. To exploit the long-term cooperation gain, a general trigger strategy is deployed as the punishment for free-riding and the Nash equilibrium (NE) of the repeated game is derived. Based on the game theoretic analysis, we develop a NE-driven incentive mechanism to guide the TP selects the most effective wages to motivate workers towards long-term cooperation and avoid midway free-riding. Simulation results show the effectiveness of our proposal.