基于多维拍卖与主从博弈的联邦学习激励机制  

An incentive mechanism of multi-dimensional auction and Stackelberg game for federated learning

作　　者：江帆 陈紫东 王军选 禹忠 JIANG Fan;CHEN Zidong;WANG Junxuan;YU Zhong(School of Communications and Information Engineering,Xi’an University of Posts and Telecommunications,Xi’an 710121,China;Shaanxi Key Laboratory of Information Communication Network and Security,Xi’an 710121,China)

机构地区：[1]西安邮电大学通信与信息工程学院,陕西西安710121 [2]陕西省信息通信网络及安全重点实验室,陕西西安710121

出　　处：《西安邮电大学学报》2024年第2期27-37,共11页Journal of Xi’an University of Posts and Telecommunications

基　　金：国家自然科学基金项目(62071377,62101442,62201456);陕西省工业攻关项目(2023-YBGY-036);西安邮电大学创新基金项目(CXJJZL2022002)。

摘　　要：为了在复杂的用户成本与有限的基站预算下保证训练用户持续参与整个联邦学习训练过程,提出一种基于多维拍卖与主从博弈模型的联邦学习激励机制。该机制结合用户自身的数据质量、数量及历史声誉等指标,采用多维拍卖法从所有拟参与拍卖的用户中选出排名前K名的用户参与联邦学习训练过程。利用主从博弈法得到基站的最优奖励和训练用户的最佳训练成本,并确定双方博弈的纳什均衡解。仿真结果表明,与固定激励机制及无激励机制相比,在数据独立同分布(Independent Identically Distributed,IID)和非独立同分布(Non-Independent Identically Distributed,Non-IID)情况下,所提机制的准确率与全局训练损失均优于对比机制,能够保证训练用户持续参与整个联邦学习训练过程。In order to ensure that the training users can continuously participate in the whole federated learning and training process under the limited base station budget,a federated learning incentive mechanism based on multi-dimensional auction and Stackelberg game model is proposed.Combined with the user’s own data quality,quantity and historical reputation,the mechanism uses the multi-dimensional auction method to select the top K users from all the users who intend to participate in the auction,to participate in the federated learning training process.The Stackelberg game method is used to obtain the optimal reward of the base station and the optimal training cost of the training user,and the Nash equilibrium solution of the game between the two sides is determined.Simulation results show that compared with the fixed incentive mechanism and the unincentive mechanism,the accuracy and global training loss of the proposed mechanism are better than those of the comparison mechanisms,which can guarantee that the training users will continuously participate in the whole federated learning and training process with data of independent identically distribution(IID)and non-independent identically distribution(Non-IID)data,which ensures that the training user can continuously participate in the whole federated learning and training process.

关 键 词：移动边缘计算 联邦学习 多维拍卖 主从博弈 纳什均衡 

分 类 号：TN929.5[电子电信—通信与信息系统]