This article explores financial data privacy protection and intelligent modeling methods based on federated learning (FL). With the widespread application of artificial intelligence (AI) technology in the financial field, data silos and privacy breaches have become increasingly prominent, seriously restricting model performance and data collaboration efficiency. In response to this challenge, this article proposes a privacy preserving modeling framework that integrates FL and gradient boosting decision tree (GBDT), fully leveraging the advantages of GBDT in fitting ability and prediction accuracy, while utilizing FL's characteristics in breaking data silos to achieve efficient and secure joint modeling. This article delves into the information exchange mechanism during the training process of federated GBDT and identifies the potential privacy leakage risks that label information may pose in gradient exchange. To address this issue, two privacy protection schemes have been designed: in the presence of semi trusted third parties, a basic scheme is constructed using semi homomorphic encryption technology; Introducing threshold semi homomorphic encryption mechanism to enhance privacy protection in scenarios without any trusted third party. The results indicate that the proposed method significantly reduces the loss of model accuracy while effectively ensuring label privacy.
Xinjun Mao (Sun,) studied this question.