Fighting Fraud by Using Deep Learning

Danske Bank is the largest bank in the Nordics, with more than $200 billion of assets in management. As with all financial institutions, one of its biggest priorities is fighting fraud, an increasingly difficult task in today’s connected, complex, and anonymous financial transaction market.

The bank had been using a rules-based engine for fraud detection, but it was faltering against modern challenges, catching only 40% of fraudulent transactions. Furthermore, this engine suffered from an extremely high rate of false positives—99.5%. This meant an unnecessarily heavy workload for investigators as well as headaches for customers, and the problem was growing worse because of new payment methods and sophisticated fraud techniques. For Danske Bank, lowering the rate of fraud represented a huge financial opportunity for preventing monetary loss, reducing investigator workload, and improving customer satisfaction.

The bank decided to update its fraud detection engine by powering it with machine intelligence, a decision that led to impressive results. Danske Bank saw a 50% increase in its fraud detection rate and a 60% decrease in the false-positive rate, a huge improvement over its rules-based engine. Furthermore, the bank did not use any additional data types to achieve these results. Instead, machine and deep learning algorithms were able to detect fraudulent patterns in the same datasets used in the previous solution.

This project was carried out in three interconnected tracks. As a foundation, the team worked to enhance the analytics infrastructure so that it could support automatic decision-making with machine and deep learning while maintaining latency requirements.

When that was ready, the machine learning track began. During this phase, data had to be gathered, cleaned, and routed to train the models. After the models were proven in a testing environment, they were productionized. This immediately created some lift over the manual rules-based engine. However, the rate of false positives was still quite high (98%), and some instances of fraud were yet undetected.

The team decided to start a deep learning track to capture these remaining cases and improve the engine’s accuracy. Using the same structured data that had been feeding the rules engine, the team discovered a way to turn it into something that looked like an image. The team was then able to use a convolutional neural network—a type of algorithm typically associated with object recognition—to detect visual patterns in the data that predicted certain transactions to be fraudulent. Using this method, and without using any additional data sources, the team saw substantial improvement over the machine learning output.

Danske Bank was able to win with machine intelligence for a number of reasons—it had the right kind of executive support that made key investments in infrastructure and in developing new processes to accommodate the AI; the company built a solid data foundation that could service a high-performance engineering environment; and the team employed rigorous testing and management for the machine and deep learning models, which ensured accuracy and built trust and credibility.

Mining Image Data to Increase Productivity

In addition to gaining insight into traditional datasets, deep learning can help enterprises dig into data that was previously inaccessible. In this next example, one major logistics company in the United States was able to mine images to increase productivity.

Of the millions of packages that this logistics company ships every day, a small number wind up as orphans—both its shipper and its destination are unknown. For years the problem had been solved manually. After the missing package was reported, someone would go to a warehouse where the orphaned packages were stored and try to match up the claim with one of them.

The process was incredibly expensive, costing the company tens of millions of dollars every year. If the company improved the situation, it could likely increase customer satisfaction (helping people get their packages sooner) and also reduce its liability for reimbursing lost property.

The company decided to approach the problem with machine intelligence, using computer vision and image recognition to design and deploy an AI-assisted solution of image-to-image search with a backend of TensorFlow and Keras. People who had sent packages that had then become lost would submit a picture of the contents of the lost package (if it were available). On the other end, the warehouse would upload pictures of the contents of orphaned packages. Using the search, the company was able to match the pictures that people had sent in with their claims for an orphan package with the actual inventory of the orphan warehouse.

The company was able to achieve a 90% match rate between the images, reducing the warehouse search window from weeks to minutes. This represented a huge reduction in resources as well as the opportunity to improve customer satisfaction due to quick resolution of lost-and-found cases.

Deep Learning for Image Recognition

In another example of using deep learning for image recognition in the enterprise, a state-run postal service was able to improve efficiency on the shop floor by building a custom AI solution that could accurately identify plastic bags.

At this postal service, 115 million parcels are shipped every year. Of those, 7.5 million are plastic bags that need a special sorting process to complete shipping. On the shop floor, these bags must be put into designated bins where they can be sorted for delivery. In the system the postal service had been using, only a small percentage of the plastic bags were detected, with most of them remaining on the conveyor belt, where they needed to be manually picked up and delivered to the correct bin. Employees were responsible for both identifying the size and quality of package (whether they were plastic) and loading and unloading the parcels into the correct containers, a time-consuming task that decreased the system’s efficiency significantly.

The postal service needed a way to automate the process, which would reduce labor costs and complete the shipping more effectively. To solve this problem, it trained deep learning algorithms on millions of images from several cameras on the shop floor. After the algorithms could detect plastic bags (not an easy task, as even slight changes in image quality can affect accuracy), the organization was able to build a system to identify the packages in real time and instruct the conveyor belt where to direct the plastic ones.

The solution involved blending traditional data sources with new image data in order to complete the process of package identification and direction. Ultimately, the postal service was able to achieve 80% accuracy on the testing dataset for the identification of plastic bags using an automated pipeline, employing Keras with a Theano backend (because of the more user-friendly interface and limited time for the project).

Natural-Language Processing for Customer Service

Customer communication represented a big challenge for a global communications provider. Not only was the amount of communication significant, but it was also costly to try to maintain high levels of availability for support. The company wanted to find an automated way to answer some of its customers’ most common questions with 24/7 availability, reducing the support workload and increasing customer happiness.

Because of the large repositories of customer interactions, the project was a perfect fit for deep learning, which excels at decoding noise and complex data like natural language.

Using Flask, Python, Jupyter, and Microsoft Azure, the company employed machine and deep learning to create several algorithms that replicated 322 types of queries in English, Roman Urdu, and their mixture. This was a significant technical challenge because there are many approaches to choose from regarding natural-language processing techniques, how to analyze and pair up questions and answers, and how to determine when to use unsupervised versus supervised learning. Dealing with questions in two languages also proved to be a significant challenge.

Ultimately, however, the team was able to deliver a user interface in .NET for interacting with the virtual agent to answer these queries. The virtual agent could also be extended to other response types. This had the dual result of reducing the cost of customer-driven communication and increasing customer intimacy and satisfaction.

Deep Learning for Document Automation

As part of the process for opening an account at a European bank, bank employees were responsible for manually validating the application by examining an applicant’s submitted documents. These could include pay stubs, copies of photo IDs, proof of address, and so on.

The bank employed about 80 people in this process and was interested in using deep learning to make predictions about the validity of the documents for a number of reasons. It wanted to increase the accuracy of the decisions that were made as well as reduce the cost of the process and speed it up.

The bank was able to use deep learning models to read and predict document validity with up to 90% average accuracy, which represented a significant improvement over the human team in terms of speed, scale, and accuracy. The models themselves were built on a sizeable kit of Amazon Web Services (AWS), cognitive APIs, TensorFlow, Keras, Python libraries, libraries for processing PDFs and images, and GPUs.

There were some challenges in creating this solution, given that working with image detection is never easy. Additionally, many of the forms were completely different from one another (especially the pay stubs), and much of the data needed to be manually annotated before the algorithms could be trained on it.

Conclusion

By creating a custom AI solution, these companies were able to utilize their data and expertise to refine their specific business processes. In these same situations, SaaS solutions or cloud APIs might have partially solved the problem, but they would have fallen short of delivering transformative value.

As you have seen, however, deep learning solutions can involve complex engineering problems. In Chapter 6, we take a closer look at the Danske Bank AI case study and how the bank was able to use AI to fight fraud and see a significant improvement in results over its human-based rules engine.