DLG-GNN: Decoupled Local-to-Global Graph Neural Network for Scalable Blockchain Fraud Detection

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Existing hierarchical graph neural networks (GNNs) for blockchain fraud detection often suffer from scalability bottlenecks when processing large-scale transaction graphs. This paper proposes DLG-GNN, a Decoupled Local-to Global Graph Neural Network that separates contract-level local encoding from inter-contract relational reasoning. By combining domain-aware partitioning with a sequential inductive pipeline, DLG-GNN maintains peak GPU memory usage below 200 MB and main memory usage below 6.5 GB. Using GATv2-based local and global encoders, DLG-GNN demonstrates superior performance over representative baselines on Ethereum, BSC, and Polygon.