An explainable federated learning and blockchain-based secure credit modeling method

Abstract

Federated learning has drawn a lot of interest as a powerful technological solution to the “credit data silo” problem. The interpretability of federated learning is a crucial issue due to the lack of user interaction and the complexity of credit data monitoring. We advocate the importance of a credit data processing-as-a-service model, which completes conventional credit models in local environments, in order to overcome these restrictions. In particular, we describe an explainable federated learning and blockchain-based credit scoring system (EFCS) in this work. First, we propose an explainable federated learning method with controllable machine learning efficiency and controllable credit model decision making, thus having controllable credit model complexity and transparent and traceable credit decision-making mechanism. Then, we suggest an explainable federated learning training mechanism for credit data that prevents leakage of the model gradients trained by individual nodes during the training of the overall model. Neither the credit data provider nor the data user has access to the raw data in the credit model training ecosystem. Therefore, privacy protection, model performance, and algorithm efficiency, the core triangular cornerstones of federated learning, when added with model interpretability, together constitute a more secure and trustworthy federated learning-based methodology, thus providing a more reliable service for credit model training and construction. The EFCS scheme is presented via simulations of different types of federated learning and their resistance to system attack, applying the proposed model to six different credit scoring datasets. Extensive experimental analyses support the efficiency, security, and explainability of the EFCS.