Catching the Bad Guys Using Artificial Intelligence
This first appeared in SIIA’s Digital Discourse Blog.
You probably have gotten a call or email from your credit card issuer asking if you made a particular transaction. Ever wonder what triggered it? Turns out it is a form of artificial intelligence called a neural network. Instead of creating general rules about what transactions are likely to be fraudulent, a neural network just looks at all your transactions and figures out your very own individual pattern of usage. If a new transaction is significantly out of pattern, that’s when you get the call or the email.
This is not new. It was one of the first successful business applications of pattern recognition software, introduced into the financial system over twenty years ago by credit card companies seeking to manage fraud risk.
You’ve seen the same technology at work in voice and image recognition. It’s what makes search engine results so relevant and automatic recommendations for movies, books and music so surprisingly on target. You’ve heard about its ability to defeat world champions at Go, the world’s most computationally intractable game. It’s at the heart of the coming revolution in transportation made possible by autonomous vehicles.
It is also a new way to find the bad guys in international criminal gangs and terrorist groups.
Banks and other financial institutions face legal obligations to detect the use of their systems for terrorist financing and money laundering, and to report these uses to law enforcement.
Money laundering is not a trivial aspect of the world’s financial system. According to the FAQ provided by the Financial Action Task Force, the United Nations estimates money laundering at $1.6 trillion and the International Monetary Fund puts the money laundering figure as high as $1.5 trillion.
The legal obligations against money laundering placed on financial institutions serve the important public purpose of defeating criminal activity and keeping ordinary citizens safe from the terrorist attacks that have become an all-too-common feature of our everyday life.
Financial institutions live up to these urgent obligations by analyzing information about their customers and potential customers, both information provided by the individuals themselves and information that can be combined about these individuals from other public and private sources. They use their own internal data sets for these purposes and often access external commercially available data bases.
Typical compliance systems widely in use today rely on financial institutions developing rules to analyze this information. Such a rules-based approach would “flag cash transactions over a certain currency amount, block transactions to certain countries, use customer data to select accounts for additional monitoring, and categorize merchant accounts based on prior transactions.”
In contrast, under a machine learning approach
“…the system would be trained to identify such transactions over time by analysing a staggering array of factors. These could eventually come to include where a customer opens an account relative to their home address, what time of day an account was opened, duration between transactions, patterns among the merchants where a customer makes transactions, relationships between other customers of those same merchants, whether a customer uses a mobile telephone, what communication channel a customer uses to contact the bank and even changes in a customer’s social media presence. The factors that AI can evaluate are limited only by the available data.”
Regulators want financial institutions to detect terrorist financing and money laundering using whatever techniques seem to be most effective. Traditional techniques of identifying customers and predicting risk based on their identity will not be going away soon. Indeed, the same machine learning approaches that can spot a pattern of bad transactions can be used to assess the risk that a potential customer would engage in these suspicious transactions. For the time being, however, regulatory concerns regarding interpretability are keeping financial institutions from relying solely on the new machine learning techniques to identify money laundering transactions.