Privacy-preserving XGBoost inference with homomorphic encryption

There is a plethora of applications that incorporate Machine Learning (ML) in a wide variety of elds. Many sectors, such as healthcare and business, provide ML- based prediction services. However, data privacy concerns may prevent using ma- chine learning services at their full potential. Enabling ML algorithms to perform under highly secure encryption schemes would grant the chance to overcome the data privacy issue. Running ML algorithms on homomorphically encrypted data is a promising way to eliminate data privacy issues. Thus, the compatibility of existing machine learning models with homomorphic encryption is crucial for data privacy. We chose to work on combining widely used XGBoost inference algorithm and homomorphic encryption to provide a post-quantum security level for query data owners. This work proposes and implements a privacy-preserving XGBoost inference method that performs operations on encrypted data. The prediction and time performance of the method is evaluated using various XGBoost models. More- over, security and privacy analysis are provided for both query owners and model owners. In conclusion, a practical and e ective homomorphic encryption solution for XGBoost inference is presented.