'Federated learning' can supercharge banks' AML detection

Africa Studio - Fotolia

The financial crimes compliance industry has been exploring new machine learning and artificial intelligence technologies for years, but with little adoption. Why? Is it because of the high consequences of our work? Is it because of the high levels of regulatory scrutiny? Why wouldn't we use AI to prevent financial crimes that enable some of the worst offenses including elder abuse, human trafficking and terrorism?

Financial institutions can help prevent financial crime with a new way to train AI called "federated machine learning," or simply federated learning. This innovation has already been applied successfully in highly regulated industries like medicine. Now, it is available and proven in the banking industry. Not only can federated learning reduce costs, but it can also increase the effectiveness of anti-money-laundering (AML) programs, all while preserving privacy.

How does federated learning work? It starts with an AI algorithm. The best AI algorithms are trained on the best data. The more quality data available for training, the more accurate algorithms become.

With federated learning, each institution's data is used to locally train and update an algorithm that is then passed between data sets, getting smarter and more accurate with each use. The algorithm, now trained on the aggregate data, can be distributed back to the participating institutions. Instead of "aggregating" data, which banking, medical and many other industries cannot do, federated learning keeps the data apart, or "federated," and moves the learning instead. All data remains secure and private.

Federated learning transforms traditionally siloed environments — regulated or not — into a collective for learning. This matters for everyone, but federated learning provides the safe path for industries that must prioritize compliance, historically at the cost of accuracy.

Federated learning can also help reduce the costs of AML compliance, which are massively high, and increasing. An estimate published in 2018 in the Journal of Financial Crime showed that the direct cost of AML compliance in the United States has reached $80 billion per year. This study found that large financial institutions spend an average of $48 million on AML compliance per year, with costs increasing annually.

Despite the well-documented large and growing financial and societal costs, there is limited evidence that these regulations effectively combat financial crime. Currently, less than 1% of global illicit financial flows are detected and confiscated by authorities. And the United Nations Office on Drugs and Crime estimates that 90% of financial flows from drug trafficking remain undetected.

Today, costs continue to increase, yet the outcome remains mostly the same. We don't accept that in our revenue-generating business lines, so why would we accept it in the prevention of money laundering and terrorist finance?

Artificial intelligence

How JPMorgan Chase and Capital One became AI leaders

Six North American banks did well in Evident's latest assessment of banks' AI efforts, but JPMorgan and Capital One have a clear lead. 

By Penny Crosman

December 21

Federated learning also increases effectiveness. We've used federated learning to pass an algorithm between datasets within our respective financial institutions, without violating customer privacy. We've seen increases in efficiency of 75% in the discovery of financial crimes at a fraction of the time and cost.

What does that mean? What happens after a possible financial crime is discovered and a Suspicious Activity Report is filed with the government's financial intelligence authority, such as the U.S. Financial Crimes Enforcement Network or Canada's Financial Transactions and Reports Analysis Centre of Canada?

In some cases, it can lead to the discovery and shuttering of large-scale trafficking schemes. In 2021, two dozen defendants were indicted for running a labor trafficking operation that illegally forced Mexican and Central American workers into brutal conditions on farms in southern Georgia. A multi-agency investigation led by the Department of Homeland Security uncovered the trafficking scheme, in part due to the more than $200 million laundered through a casino, cashier's checks and cash purchases of land, homes, vehicles and businesses.

Imagine the impact on people's lives if the financial industry adopted a technology that could make us more effective at finding crimes such as these.

Those working in financial institutions, the creators of financial technology and regulators all share the common goal of innovation. Federated learning can help us operationalize this goal.

For example, let's say 10 banks use their own teams, with their own models, trained on their own data, to identify illicit transactions. If those same 10 banks used federated learning to train one model to find illicit transactions, that model could be used by those 10 banks to help prevent those illicit transactions. Because the model was trained on data from 10 banks, it would be better at identifying true alerts (as opposed to false positives). Not only could this increase effectiveness, but it could also reduce overall program costs and free up capacity for other, more productive tasks.

What if we did that 10 times? What would banking, risk management and financial compliance look like by 2030?

With the arrival of AI and federated learning, banks, regulators and law enforcement professionals will finally have transparent methods of measurement of policy outcomes, a necessary prerequisite to a better system for fighting crime and expanding financial inclusion.