Fairness-Aware Reverse Auction-Based Federated Learning

Auction-based federated learning (AFL) has garnered significant research attention recently. However, existing methods for AFL data consumers (DCs) primarily focus on improving FL model performance by recruiting data owners (DOs) with high reputations and low ask prices, disregarding fair treatment for DOs. The challenge of striking a balance between performance and fairness when recruiting DOs remains unaddressed. To tackle this issue, we propose the Fairness-aware Reverse AFL for DCs (FAR-AFL). FAR-AFL leverages Lyapunov optimization to dynamically adjust selection probabilities for potential DOs, taking into account dynamic changes in participation rates and reputation. FAR-AFL adopts a reverse auction-based DO recruitment mechanism to determine candidate selection and pricing. By combining these components, FAR-AFL improves FL model accuracy while minimizing overall recruitment costs. Crucially, FAR-AFL ensures equitable DO treatment, providing them with fair participation opportunities. Theoretical analysis shows the computational efficiency, individual rationality, and truthfulness of FAR-AFL. Extensive experimental evaluation against six alternative strategies on four benchmark datasets demonstrates that FAR-AFL outperforms the best alternative strategy by 1.99%, 6.60%, 1.97%, and 23.31% in terms of test accuracy, root mean square error, cost reduction, and fairness improvement, respectively.