Hybrid AI Boosts Financial Fraud Detection Accuracy

Financial institutions face significant challenges in identifying sophisticated money laundering schemes, such as smurfing and layering, primarily due to the extremely low incidence of fraud and the constant evolution of adversarial evasion techniques. This research proposes FraudShield AI, a novel hybrid framework designed to enhance fraud detection capabilities. FraudShield AI integrates Long Short-Term Memory (LSTM) networks to capture temporal sequences of transactions with hand-crafted Graph Topological Features, which provide insights into the structural relationships within transaction networks. By incorporating network-centric features like PageRank Centrality and In-Degree dynamics, the system shifts focus from isolated transactions to a more comprehensive network-level forensic analysis. The framework also employs a Focal Loss objective to mitigate issues arising from class imbalance and introduces a dynamic thresholding mechanism to improve resilience against subtle, low-value smurfing attacks. Experimental results on the PaySim dataset demonstrate that FraudShield AI substantially outperforms traditional baselines in key metrics, particularly for detecting difficult micro-transaction fraud patterns.