A Graph Neural Network Platform for Blockchain AntiMoney Laundering: A Literature Review and Proposed System Architecture

Cryptocurrency networks give criminals a fast, borderless and pseudonymous way to launder money, yet the public ledgers that make this possible also produce transaction graphs far too large and too fast-moving for manual or rule-based review. Graph neural networks (GNNs) have become the leading modelling approach for this problem because they work directly on the account-transaction graph and not on flattened, per-transaction feature vectors. This paper reviews the recent literature on GNN-based blockchain anti-money laundering (AML) and organizes eighteen studies into five threads: foundational models built on the Elliptic benchmark, attention-based and heterogeneous architectures, temporal and ensemble methods, explainability and privacy-preserving techniques, and early system-level platforms for real-time use. A comparative analysis and a six-point gap analysis show that few studies deal with end-to-end platforms that combine incremental graph maintenance, low-latency inference and investigator-facing explanation. To fill this gap, the paper proposes ChainSentry-GNN, a six-layer reference architecture covering ledger ingestion, incremental graph construction, model serving, explainability, adaptive thresholding and an analyst feedback loop, together with an evaluation methodology centred on detection quality, latency, throughput and cross-chain generalization.