Authors - Amelia Santosh, Bhavika Pradeep, Dhanuvarsha S S, Harisurya Reddy S, Shruthi L Abstract - Real-time analysis, high accuracy, and robust privacy protection across several institutions are necessary for financial fraud detection. Restrictions on data sharing and non-IID transaction patterns cause traditional centralized models to fail. Graph Neural Networks (GNNs) for anomaly detection and a structured fraud reporting mechanism are integrated in this paper’s federated learning-based fraud detection framework. While GNNs capture intricate relationships between accounts, devices, and transactions, the system allows institutions to jointly train a global model without exchanging raw data. The feasibility of implementing collaborative fraud detection across financial institutions is demonstrated by the experimental results, which show improved fraud detection performance, enhanced recall on minority fraud cases, and effective privacy preservation.
Open Zoom