Enhancing AI-Cloud Computing in Healthcare with BERT for Clinical Text Understanding

Abstract

Financial fraud detection has become increasingly complex due to the rising sophistication of fraudulent activities
and the need to preserve user privacy. Traditional fraud detection systems rely on centralized data processing,
posing risks to data security and regulatory compliance. This study proposes a BERT-based Federated Learning
(FL) framework integrated with Graph Neural Networks (GNNs) for privacy-preserving and scalable fraud
detection. The framework enables multiple financial institutions to collaboratively train fraud detection models
without sharing raw data, ensuring compliance with privacy regulations such as GDPR and CCPA. Additionally,
deep reinforcement learning (DRL) optimizes fraud detection strategies by dynamically adjusting classification
thresholds to reduce false positives. The model is trained and evaluated on the Bank Account Fraud Dataset Suite
(NeurIPS 2022), achieving an accuracy of 96.2%, precision of 94.1%, recall of 95.6%, and an AUC-ROC score
of 0.982, significantly outperforming traditional machine learning approaches such as Random Forest (88.2%)
and Centralized CNN models (91.3%). Moreover, the proposed FL framework reduces communication overhead
by 37% compared to traditional centralized approaches, making it highly efficient for real-world cloud
environments. The experimental results demonstrate that the proposed framework enhances fraud detection
performance while maintaining data confidentiality, offering a robust, privacy-preserving AI solution for modern
financial institutions.