基于激励机制的联邦学习优化算法

Federated learning optimization algorithm based on incentive mechanism

针对联邦学习的训练过程迭代次数多、训练时间长、效率低等问题,提出一种基于激励机制的联邦学习优化算法。首先,设计与时间和模型损失相关的信誉值,基于该信誉值,设计激励机制激励拥有高质量数据的客户端加入训练。其次,基于拍卖理论设计拍卖机制,客户端通过向雾节点拍卖本地训练任务,委托高性能雾节点训练本地数据从而提升本地训练效率,解决客户端间的性能不均衡问题。最后,设计全局梯度聚合策略,增加高精度局部梯度在全局梯度中的权重,剔除恶意客户端,从而减少模型训练次数。

Federated learning optimization algorithm based on incentive mechanism was proposed to address the issues of multiple iterations,long training time and low efficiency in the training process of federated learning.Firstly,the reputation value related to time and model loss was designed.Based on the reputation value,an incentive mechanism was designed to encourage clients with high-quality data to join the training.Secondly,the auction mechanism was designed based on the auction theory.By auctioning local training tasks to the fog node,the client entrusted the high-performance fog node to train local data,so as to improve the efficiency of local training and solve the problem of performance imbalance between clients.Finally,the global gradient aggregation strategy was designed to increase the weight of high-precision local gradient in the global gradient and eliminate malicious clients,so as to reduce the number of model training.