Federated learning (FL) has emerged as an important paradigm for privacy-preserving machine learning in the financial sector, enabling multiple institutions to collaboratively train models without sharing raw sensitive data. However, in practical cross-institution deployments, federated training still faces two major bottlenecks: (i) parameter transmission requires more fine-grained security protection while keeping communication overhead under control, and (ii) distribution shifts induced by non-independent and identically distributed (non-IID) data can undermine the stability of the global model and degrade detection performance. To address these challenges, we propose a more general and secure federated learning framework. The proposed framework incorporates a selective encryption mechanism that encrypts only sensitive parameters during transmission, thereby balancing communication efficiency and security. In addition, we integrate data augmentation strategies to mitigate inter-client distribution discrepancies and improve training robustness and generalization. Experimental results on a credit card fraud detection datasets demonstrate the feasibility and effectiveness of the proposed approach.
Liu et al. (Wed,) studied this question.