【摘要】 不同的金融机构在进行信用卡欺诈预测时往往采用不同的神经网络模型，使用联邦学习（FL）算法进行联合训练时存在模型异构问题。针对这一问题提出一种改进的联邦学习算法（FedPCD）,该算法基于原型学习的思想，在客户端训练的过程中引入聚散损失来增强模型区分不同类别原型的能力；服务器端通过K-means聚类算法对客户端进行智能分组并计算组内客户端的余弦相似度来实现原型的加权聚合。针对三种模型异构场景进行了实验，结果表明，FedPCD算法能够有效解决联邦学习中的模型异构问题，预测的准确率、召回率、AUC和F₁值均得到提升。更多还原

【Abstract】 Different financial institutions often adopt various neural network models for credit card fraud prediction, which presents a model heterogeneity problem when jointly training using Federated Learning（FL） algorithms. This paper proposes an improved federated learning algorithm, Federated Prototype Clustering Dispersion（FedPCD）, which is based on the concept of prototype learning. During client-side training, it introduces a clustering dispersion loss to enhance the model’s ability to differentiate between different category prototypes. On the server side, intelligent grouping of clients is performed using the K-means clustering algorithm, and prototype weighted aggregation is achieved by calculating the cosine similarity among clients within each group. Experiments conducted in three model heterogeneity scenarios demonstrate that the FedPCD algorithm effectively resolves the model heterogeneity issue in federated learning, improving prediction accuracy, recall, AUC, and F₁ scores.更多还原