Federated learning financial data management techniques based on data distillation and dynamic tiered scheduling

As the financial industry becomes increasingly reliant on big data and artificial intelligence technologies, the efficient analysis and modeling of data while safeguarding privacy has become one of the key areas of research. A federated learning financial data management system has been developed to improve system security and data management efficiency. It combines homomorphic encryption and multi-party secure computing to provide a foundation of privacy and security, and it introduces knowledge distillation technology to significantly reduce communication data volume. Meanwhile, the study proposes a dynamic hierarchical reservation algorithm to optimize the dynamic allocation efficiency of computing resources in the federated learning process. The results indicated that the total elapsed time of the proposed model was reduced by 48.1% compared to the traditional federated learning model. When the data heterogeneity level was high, the communication epochs of the proposed model in the dataset were reduced by 11.4% compared to the traditional model. It also required the least number of communication epochs, which was only 88. In addition, the distillation method could improve the model's accuracy by 12.43% and 12.19%, respectively, compared to the two baseline models, even with a small training data set. This technology can effectively improve the training efficiency and accuracy of the model. This study provides an efficient, secure, and scalable solution for data management and analysis in the financial industry, while ensuring the privacy and security of financial data.