DL-RL: Dual-Layer Graph Neural Networks with Curriculum Reinforcement Learning for Financial Fraud Detection

Financial fraud imposes hundreds of billions of dollars in annual losses on the global economy and continuously evolves to outpace static detection systems. This paper proposes DL-RL, a Dual-Layer Graph Neural Network framework with Curriculum Reinforcement Learning for real-time financial fraud detection. The lower GNN layer (a Graph Attention Network with 4 heads and 128 hidden units) captures local transaction-neighborhood embeddings via attention-weighted aggregation, while the upper layer (GraphSAGE, 64 units) propagates inter-account risk signals across the global transaction graph. A PPO-based curriculum scheduler progressively exposes the agent to increasingly complex fraudulent subgraphs, growing difficulty from κmin = 0.1 to 1.0 over 500 warm-up episodes. Experiments on three public benchmarks—IEEE-CIS, Elliptic Bitcoin, and PaySim—yield macro F1 scores of 0.931, 0.947, and 0.963 respectively, improving over the strongest baseline by up to 5.2 percentage points while reducing false positive rates by up to 38%. Ablation studies confirm the independent contribution of each architectural component. A proof-of-concept implementation is available at https://github.com/anonymous/DL-RL-fraud (to be released upon acceptance).