A multi-source credit data fusion approach based on federated distillation learning

Data imbalance and privacy disclosure shortcomings have become the main problems in the process of multi-source credit data fusion, the former causes conflicts during the fusion process, the latter brings huge security risks. While federated learning is used for data privacy protection, communication cost defects and inaccurate fusion results will follow. In order to effectively unify data fusion, the paper proposes an approach based on federated distillation learning, which uses synthetic distillation data instead of traditional parameter transfer models to fuse to reduce time cost and improve accuracy without compromising data privacy,simultaneously utilizing local data to train the model and conducting interactive learning with the server's model. Specifically, it uses a decision tree model to distill knowledge from credit data, replacing the traditional parameter transfer model. At the same time, the Generic Adversarial Network is used to balance data distribution and solve the problem of data imbalance on the server. The experimental results show that the method proposed has improved both utilization performance and unbalanced data processing by at least 3%.