Loss-Convergence- Driven Federated Learning with Energy Optimization: A Stackelberg Game Approach
Yanbo Fang, Tengfei Cao, Yiming Zhang · 2024
In Federated Learning (FL), the accuracy of each client model directly impacts the performance of the aggregated global model. Increasing the number of local training rounds can improve the accuracy of the aggregated model, thereby reducing the number of global training rounds and minimizing communication overhead. However, existing methods need help in incentivizing clients for training, particularly in evaluating their contributions. To address these issues, the paper introduces a Stackelberg game-based incentive mechanism that focuses on loss convergence and power optimization. By leveraging the Monte Carlo algorithm to solve the game equilibrium, we optimize the training strategy to accelerate the FL process and reduce energy consumption effectively. Experimental results demonstrate that, while maintaining the same level of global model accuracy, our approach reduces energy consumption by up to 33% and shortens the number of global iterations by 27 %.