Privacy-Preserving Logistic Regression on Vertically Partitioned Data

Logistic regression is the important algorithms of machine learning. Traditional training methods require centralized collection of training data which will cause privacy issues. To solve this problem, this paper proposes privacy-preserving logistic regression. This scheme is suitable for dividing data by feature dimension, and the training data is shared between two parties. The two parties conduct collaborative training and learn a shared model. In this scheme, the two parties train the model locally on private data set while exchanging the intermediate calculation results without directly exposing their private data. Additionally, the additively homomorphic scheme can ensure the calculation security which can be performed on the cipher text. During the training process, the participants can only obtain zero knowledge of each other and cannot get any information about model parameters and training data of another participant. At the same time, a privacy protection prediction method is provided to ensure that the model deployment server cannot obtain the private data of the inquirer. After analysis and experimental verification, within the tolerable loss of precision, the scheme is secure against semi-honest participants and provide privacy protection.