Federated learning against money laundering
Federated learning, which trains one model across organizations without pooling their data, is a large topic in the AI journals, yet only 7 of 398 recent papers on it carry a finance, banking, fraud or money-laundering keyword. Money launderers spread transactions across banks precisely because no single bank sees the whole pattern. Can banks catch them together without sharing customer data?
What prompted this
Since October 2025, 398 papers carry a federated-learning keyword (386 in AI journals), and only 7 of them carry a money-laundering, bank, finance, fraud or credit keyword. "Anti-money laundering" appears on 14 papers. In the AI journals, "decentralized learning" is rising (8 papers in July–September 2026 against about 2 expected).
Questions
- On realistic synthetic data in which laundering chains cross banks, how much does a federated model improve detection over models that each bank trains alone?
- How much detection do privacy protections such as secure aggregation and differential privacy cost?
- Who should run and pay for such a system: the banks, a shared utility or the regulator?
Data that could help
Synthetic anti-money-laundering datasets with transactions across banks, public transaction-graph benchmarks, and open federated learning frameworks.
Next step
Split a synthetic money-laundering dataset across five simulated banks and compare federated, bank-by-bank and pooled models on detection and false alarms.