This study proposes a federated graph-learning framework for detecting money-laundering activities across financial institutions without sharing raw customer data. Four banks participated in local training using transaction-graph structures containing 15.7 million nodes and 112.5 million edges. A global aggregation mechanism was used to synchronize encrypted model parameters. Compared with local graph models, the federated framework improved the average AUC from 0.84 to 0.90 and increased precision at 10% recall by 29.7%. Robustness tests showed stable performance even when one institution exhibited severe label imbalance. Differential-privacy noise prevented reconstruction of sensitive records under gradient-inversion tests. The results confirm that cross-institution modeling can significantly enhance detection accuracy while maintaining data confidentiality.
