FedXHDP: A Federated XGBoost Framework with Hierarchical Differential Privacy for Horizontally Partitioned Data

Federated learning facilitates collaborative model training across decentralized data sources while safeguarding user privacy, making it particularly suitable for sensitive biological applications. However, achieving robust privacy guarantees in decision tree-based models remains a challenge. This paper presents FedXHDP, a federated decision tree-learning framework that combines the predictive capabilities of XGBoost with Hierarchical Differential Privacy (HDP). Our approach utilizes a layered privacy architecture at local, node, and global levels, efficiently safeguarding against data reconstruction, membership inference, and model inversion attacks. Compared to traditional methods that rely on cryptographic techniques, FedXHDP ensures robust privacy with reduced computational demands. To improve generality and mitigate overfitting, we incorporate DART, a dropout-based regularization method, into the boosting process. Experimental evaluations of the datasets indicate that FedXHDP attains a superior privacy-utility trade-off relative to current methodologies, particularly in horizontal partitioning. While our model emphasizes binary classification and multi-class classification in our work, subsequent research will investigate advanced optimization methods.