Application and Effectiveness Evaluation of Federated Learning Methods in Anti-Money Laundering Collaborative Modeling Across Inter-Institutional Transaction Networks

We propose the Graph Foundation Model (GFM): performing self-supervised contrastive pre-training on heterogeneous account-merchant-geo-device graphs locally within each institution. This achieves cross-institutional knowledge transfer and privacy protection through federated learning + secure aggregation + DP-SGD (ε≤3.0). On 95 million transactions across 5 institutions, GFM—deployed as a freezing backbone + lightweight adapter—achieved 23–31% higher PR-AUC and 9–13 percentage points higher Recall (with Precision fixed ≥0.92) compared to independently trained GNNs per institution. For open-set detection, it demonstrated 18–24% higher energy score detection rates for novel typologies. Communication and training overhead were controlled at ≤40MB per round, with total duration reduced by −27%. Grouped SHAP and subgraph attention provided auditable explanations. This demonstrates that federated self-supervised pretraining can significantly enhance AML generalization performance without sharing raw data.