A Fair Incentive Mechanism for Federated Auctioning Networks

Incentivizing data owners to contribute to federated learning (FL) is crucial to the sustainable operation of an FL ecosystem. Existing incentive mechanisms are designed assuming that all data owners are known to FL task publishers, which may not always hold in practical scenarios. As the domain of auction-based FL (AFL) continues to grow in importance, we are rethinking this assumption to better suit the realities of Federated Auctioning Networks (FANs). We propose an incentive mechanism named FIM-FAN, which is designed to perform data owner selection based on Lyapunov optimization through referrals among data owners in FANs, without relying on the existence of an entity that has information about all data owners. In FIM-FAN, a data owner can perform the dual roles of a worker and a referrer. Both activities can be incentivized monetarily. To solve the data owner selection problem, an online greedy client selection algorithm is proposed considering reputation, bidding price, and fairness. Reputation is also involved in the compensation calculation process to encourage honest behaviors in FAN. Theoretical analysis shows that FIM-FAN satisfies budget feasibility and individual rationality. Extensive experiments on MNIST and CIFAR-10 datasets against 10 baselines demonstrate that, on average, FIM-FAN outperforms them by 6.54% and 13.63% in terms of test accuracy and fairness, respectively.