Abstract
The rapid evolution of decentralized cross-chain bridges and privacy-centric cryptocurrencies has significantly enhanced financial interoperability and user confidentiality. However, these same technological advancements present novel avenues for sophisticated financial illicit activities, notably cross-chain smurfing—a money laundering tactic involving the fragmentation of illicit capital into small, multi-hop transactions across disparate blockchain networks and privacy protocols. Existing anti-money laundering (AML) frameworks predominantly focus on single-chain heuristics or static graph analytics, rendering them ill-equipped to detect dynamic, multi-ledger obfuscation strategies. In this paper, we propose a novel Heterogeneous Temporal Graph Neural Network (HT-GNN) specifically designed to detect cross-chain smurfing and laundering behaviors. By modeling cross-chain transactions as dynamic multi-relational graphs incorporating node heterogeneity (e.g., user wallets, smart contracts, bridge liquidity pools) and temporal edge attributes (e.g., execution latency, fan-in/fan-out timing ratios), our approach effectively captures complex topological laundering signatures without requiring plain-text transaction amounts or compromised cryptographic privacy. Evaluated on a benchmark dataset spanning public block logs and privacy-centric bridge telemetry, our HT-GNN framework achieves an F1-score of 94.2% and an Area Under the Receiver Operating Characteristic curve (ROC-AUC) of 0.978, outperforming traditional machine learning baselines and homogeneous graph architectures. Our findings highlight the viability of topology-aware deep learning for real-time, privacy-preserving regulatory compliance in decentralized finance ecosystems.
