Yardstick competition regulation for incentive mechanisms in federated learning: balancing cost optimization and fairness

Official URL: https://dx.doi.org/10.23919/WiOpt58741.2023.10349890

Federated learning (FL) is a global machine-learning model that is trained using several participating nodes, where data is kept private. In this paper, we propose a cost-effective dynamic joint load balancing and pricing algorithm (CELBP) for heterogeneous workers that utilizes the Yardstick Competition Regulation (YCR) to encourage workers to disclose their costs honestly and provide incentives based on relative contributions. In the dynamic case, we aim to minimize the total cost of the server, considering the positive value of subtasks assigned to workers and their limited full load, and a new cost function for workers that accounts for their efficiency level and total communication and computation time required to complete assigned tasks. We show that CELBP outperforms other schemes in terms of accuracy and training time while reducing the server’s total cost, using experiments with the MNIST and CIFAR-10 datasets. Additionally, we show that using the proposed online algorithms improves accuracy and reduces latency when compared to other algorithms mentioned in the paper.