联邦学习中基于 Chebyshev 定理的模型性能感知逆向拍卖

文章研究多服务器、多客户端联邦学习(federated learning, FL)场景中的激励机制,并将任务分配和定价问题建模为多个逆向拍卖问题。根据切比雪夫(Chebyshev)定理对客户端每一轮的本地模型性能进行评估,并进一步利用指数衰减函数评估其本地模型的总体性能;设计基于本地模型性能的逆向拍卖(local model performance based reverse auction, LPRA)算法解决任务分配和定价问题以激励更多高性能的客户端参与,并从理论上证明LPRA算法满足个体理性、真实性和计算高效性;通过仿真实验验证LPRA算法的有效性。

This paper studies the incentive mechanism in a multi-server, multi-client federated learning (FL) scenario and models the task allocation and pricing as multiple reverse auction problem. Firstly, local model performance of clients is evaluated in each round according to Chebyshev's theorem, and exponential decay function is used to evaluate the historical performance of clients. Then, a local model performance based reverse auction (LPRA) algorithm is designed to solve the task allocation and pricing problems with the goal of maximizing the overall performance of clients participating in FL. Through theoretical analysis, it is confirmed that LPRA algorithm satisfies individual rationality, truthfulness and computational efficiency. Finally, the effectiveness of the LPRA algorithm is verified by simulated experiments.