The Force of Compensation, a Multi-stage Incentive Mechanism Model for Federated Learning

Han Xu, Priyadarsi Nanda, Jie Liang, Xiangjian He

Abstract: In federated learning, data owners ‘provide’ their local data to model owners to train a mature model in a privacy-preserving way. A critical factor in the success of a federated learning scheme is an optimal incentive mechanism that motivates all participants to fully contribute. However, the privacy protection inherent to federated learning creates a dual ethical risk problem in that there is information asymmetry between the two parties, so neither side’s effort is observable. Additionally, there is often an implicit cost associated with the effort contributed to training a model, which may lead to self-interested, opportunistic behaviour on both sides. Existing incentive mechanisms have not addressed this issue. Hence, in this paper, we analyse how dual ethical risk affects the performance of federated learning schemes. We also derive an optimal multi-stage contract-theoretic incentive mechanism that minimises this risk, and experiment with calculating an optimal incentive contract for all participants. To our best knowledge, this is the first time that dual ethical risk for federated learning participants has been discussed. It is also the first time that an optimal incentive mechanism to overcome this issue has been developed.