Boosting Privately: Privacy-Preserving Federated Extreme Boosting for Mobile Crowdsensing

The state-of-the-art federated learning brings a new direction for the data privacy protection of mobile crowdsensing machine learning applications. However, besides being vulnerable to GAN based user data construction attack, the existing gradient descent based federate learning schemes are lack of consideration for how to preserve the model privacy. In this paper, we propose a secret sharing based federated extreme boosting learning frame-work (FedXGB) to achieve privacy-preserving model training for mobile crowdsensing. First, a series of protocols are designed to implement privacy-preserving extreme gradient boosting of classification and regression tree. The protocols preserve the user data privacy protection feature of federated learning that XGBoost is trained without revealing plaintext user data. Then, in consideration of the high commercial value of a well-trained model, a secure prediction protocol is developed to protect the model privacy for the crowdsensing sponsor. Additionally, we operate comprehensive theoretical analysis and extensive experiments to evaluate the security, effectiveness and efficiency of FedXGB. The results show that FedXGB is secure in the honest-but-curious model, and attains approximate accuracy and convergence rate with the original model in low runtime.