Danske Bank is using Artificial Intelligence and Deep Learning to detect and then prevent sophisticated fraud in multiple areas and it’s working better and better every day because the models are learning.
Danske Bank distinguishes between two types of fraud – customer fraud and “fraudsters.” The customer is at the center of the fraud. For example, the customer receives an email from a citizen in a remote country asking the customer to send money to help alleviate hardships or arrange a visit as there is a courtship in the making. And then there is true professional fraud done when a “fraudster” tracks the perfect time to do serious damage. This can include malware that is infected into a bank or where personal ID’s are taken and malware is added to devices.
“Sometimes in some scam cases the fraudsters attack us for ten minutes and then they never return. Their goal is actually to get the maximum value of those 10 – 15 minutes of fame as we call it and then stop up.” Nadeem Gulzar, Head of Global Analytics
Danske Bank is using LIME (Locally Interpretable, Model Explanation.) LIME is a method used to explain the Deep Learning and what features matter. It is an open piece of software that basically helps the team using the model to explain the factors that make them believe that the model is solid. During this step and in one of the test projects, the team had to explain why they wanted to block a credit card transaction. In one example, bank customers buy from eBay, and the payment goes to China. But, today, the customer is using Alibaba. Is that fraud? Don’t know. And in this case, what do we tell the model to do?
In another example, the customer lives in Brazil, but today they are having lunch at a restaurant in Copenhagen. Is that credit card transaction fraudulent or not? This is where behavior data is important. Most customers have a preference, and when that preference isn’t chosen, what is occurring? Shall we execute the credit card transaction or not?
Is it all AI? No, human interaction is required to help train the models. For example, investigative officers were brought in to better understand anomaly detection. Human knowledge was needed to better understand the scene alongside the fraud models. Is this a sophisticated ring of “fraudsters”? Is this only one “fraudster”? Is this a group of 10-15 minute “fraudsters” creating a cluster? Is this a trend?
One of our favorite parts of this story is the ‘champion/challenger’ model strategy. Both champion and challenger models are always being tested using production data. And because there are billions of transactions happening every day, they can constantly improve the models. Danske Bank sets thresholds for the models, and when they go below a threshold, they determine if they are feeding it enough data. For example, do they need to add in geo-location data? Add in ATM data? Model comparison is done live! And when appropriate, challenger models become champion models. That’s so cool!
All of this takes a diverse team. Danske Bank employs platform, technical & data engineers, data scientists, the business and even highly trained criminal investigators – all of them work with experts in AI and Deep Learning to innovate. And they will even hire from local universities!
The results for the business are more than impressive. Before applying AI and Deep Learning, Danske Bank had 1200 false positives a day. Those were cases that had to be analyzed by Danske Bank investigators, sometimes even external agencies like Interpol. Now that number has been reduced by 60%, saving bank investigators significant time and allowing them to investigate real cases of fraud. And that’s not all. Detecting true positives has increased to 50%. Teams at Danske Bank believe this is just the beginning.
Congratulations to Danske Bank for all the success in this renaissance of Artificial Intelligence!