A privacy-preserving decentralized credit scoring method based on multi-party information

With society's wide-scale adoption of information technology, significant information about borrowers is distributed across various parties, information that can be jointly used to improve credit scoring. However, use of such information faces many challenges, such as the problems of preserving privacy and information redundancy. To address these challenges in leveraging multi-source information for credit scoring, we propose a decentralized multi-party method based on logistic regression. Specifically, we formulate a logistic regression model using the vertical federated learning paradigm. To preserve data privacy during multi-party collaborative model training, we use additively homomorphic encryption based on the second-order Taylor series expansion of the loss function and its gradient. To address information redundancy and to improve the performance of the credit scoring model, we use the proposed HE-DPGD algorithm to estimate the distributed coefficients in a privacy-preserving setting. Empirical evaluation shows that the proposed method can leverage information from multiple parties securely and effectively.