Adaptive Federated Learning for Privacy-Preserving Modeling in Heterogeneous Financial Environments

This study addresses the conflict between data privacy and modeling performance in financial technology by proposing a federated learning-based privacy-preserving framework. The research first analyzes the sensitivity of user transaction data and the issue of cross-institutional data silos, highlighting the limitations of traditional centralized modeling in terms of privacy and compliance. To overcome these issues, a distributed collaborative modeling mechanism is designed, where participants train models locally and achieve global optimization through parameter uploading and weighted aggregation, thus avoiding centralized storage and transmission of raw data. Differential privacy and secure aggregation are introduced to ensure that individual information is not exposed during parameter exchange, enhancing the overall privacy protection of the system. Furthermore, to address the non-independent and non-identically distributed nature of financial data, a personalized regularization term is incorporated to mitigate the impact of distribution differences across data sources, thereby improving adaptability and robustness in heterogeneous environments. Experiments, including comparisons with mainstream methods and multi-dimensional sensitivity analyses, verify the effectiveness and superiority of the proposed method under privacy-preserving conditions, as shown by improvements in AUC, ACC, F1-Score, and Precision. The results demonstrate that the framework can ensure data security while maintaining strong predictive performance and stability. In summary, this study not only achieves secure modeling of financial data but also provides a feasible direction and reference for further research on privacy-preserving algorithms in financial technology applications.