Communication-Efficient Federated Learning for Real-Time Anti-Money-Laundering Monitoring

To support real-time anti-money-laundering (AML) surveillance, this study introduces a communication-efficient federated learning (FL) protocol combining parameter sparsification, quantization, and adaptive client participation. The evaluation uses a dataset representing 28.4 million daily transactions from five commercial institutions. Under a 5-second alert-latency constraint, the proposed method reduced communication volume by 61.1% and update latency by 47.3% compared with standard FL. Detection performance remained stable, with AUC values decreasing only from 0.90 to 0.89 and false-positive rates increasing by 2.0 percentage points at 80% recall. When network congestion occurred, the adaptive mechanism prioritized banks with higher model drift and prevented performance degradation. The system demonstrates the feasibility of deploying FL-based AML models under strict real-time requirements.