DFedXGB: An XGB Vertical Federated Learning Framework with Data Desensitization

The emergence of Vertical Federated Learning (VFL) addresses the issue of data isolation and enhances edge intelligence. However, the high computational costs pose a significant challenge when employing Homomorphic Encryption (HE) for privacy protection in VFL. To address this challenge, we present DFedXGB, a secure and reliable framework for Federated XGBoost (FedXGB). DFedXGB incorporates an innovative data desensitization scheme to safeguard raw data and model parameters against privacy disclosure. Furthermore, We also designed a secure aggregation algorithm based on XGBoost to mitigate server collusion. Theoretical analysis confirms the correctness of DFedXGB, and security analysis establishes its provable security. Experimental results on real datasets demonstrate that DFedXGB achieves lossless accuracy comparable to non-privacy-preserving centralized methods. Moreover, DFedXGB reduces computational costs by an average of 85% compared to SecureBoost.