Research on Generative Federated Learning-Driven Multimodal Risk Control Model for Finance

Abstract

The digital transformation of finance demands advanced risk control models capable of multimodal analysis while preserving privacy. Existing federated learning approaches lack synthetic data generation capabilities for comprehensive risk assessment. This study proposes a Generative Federated Learning-Driven Multimodal Risk Control (GFL-MRC) model that integrates federated learning with generative adversarial networks (GANs) for privacy-preserving synthetic data generation and cross-modal feature fusion. The framework enables decentralized institutions to collaboratively train models without sharing raw data, employing a cross-modal attention mechanism to correlate structured (transactions), semi-structured (profiles), and unstructured (text/image) data. Experiments on financial datasets show GFL-MRC achieves $12.7 \%$ higher fraud detection accuracy ($F 1=0.892$) than conventional federated learning, reduces false positives by 23.4 % through synthetic augmentation, and ensures GDPR/CCPA compliance via differential privacy. The model particularly enhances detection of emerging fraud patterns in low-data scenarios by generating privacy-compliant training samples. This work advances collaborative risk management by bridging regulatory constraints with AI innovation, offering financial institutions a scalable solution for multimodal risk analysis without compromising data privacy.