One step ahead of the fraudster
Combining cutting-edge machine learning methods with human expertise
Pull out your card, enter your PIN, get your goods - paying without cash, on the internet or in the store is fast and easy. Yet the consumer is unaware of the numerous processes that run in the background of each transaction. To protect customers, retailers, and the banks involved, it is important to quickly assess the risk of fraud for each transaction. Banks use special fraud prevention software to assess in real-time if the transaction poses a risk. This needs to work reliably for millions of transactions each day, and that is why fraud prevention systems also employ artificial intelligence methods. Banks are supported by algorithms that use machine learning methods to identify fraud patterns.
______
The task is to derive a decision from huge data sets within miliseconds: Fraudelent or not?
Combination of machine lerning with classic fraud prevention
A bank such as the Dutch-based Rabobank manages approximately 12 million customer accounts. It monitors up to eight million transactions each day. At peak times, this can be over 5,000 payments per second. To manage this, the bank uses the intelligent software solution RiskShield from INFORM. It verifies all transactions, internet banking, online purchases, and mobile banking in real time. Along with looking for deviations in typical consumer behavior, the system evaluates information such as address changes, credit lines, or requests for new passwords when looking for fraudulent activities. Banks compare huge datasets consisting of customer profiles, prior financial transactions, as well as login and internet activities with one another to reach a decision within milliseconds. Fraudulent or not? INFORM's intelligent software combines state-of-the-art methods in machine learning with classic fraud prevention based on an expert knowledge system.
Using big data to generate knowledge
The big challenge in fraud prevention is achieving a high level of precise selectivity. On the one hand, fraudulent and suspicious transactions need to be identified and blocked; on the other hand, most of the transactions are legitimate and should be processed as quickly as possible. An erroneously declined payment does not only annoy customers, but also generates more work for the banks. To achieve this level of distinction, fraud prevention systems need to utilize a large amount of data. That is why it is important to asses information above and beyond a single payment transaction in progress. Each transaction is enriched with additional information by the system.
Thousands of variables
Information such as the time and amount of the transaction as well as the typical behavior of the card holder can be added. Deviation from this behavior may indicate fraud. A lot of information is added to the transaction to be assessed. For example, it is possible to determine when a customer has made a purchase at a bookstore. Purchases at a bookstore are typically for small amounts. An unusually high purchase at a bookstore can indicate possible fraud. These examples illustrate the principle but not the complexity. Depending on the payment type, there are thousands of different decision-making variables that determine if a transaction can be flagged with high certainty as potential fraud.
______
An algorithm does not do its job in a type of black box, but functions in a way that experts can understand why a transaction has been flagged as fraudulent.
Identifying fraud patterns automatically
With machine learning methods, banks are able to deduce the probability of a fraudulent transaction by examining the large volume of data. To do this, an algorithm is trained using flagged transactions, meaning transactions historically identified as fraudulent, to identify the fraud pattern in the data sets before it is used in software in real-time. This ensures that fraud patterns are automatically detected. After initial training, the algorithm continues to learn based on current transactions to keep the system up-to-date
Fraud detection with human expertise
Transparency is one of many reasons why banks also use the classic method to detect fraud. This method is a system of rules based on human expertise. This approach also compares typical customer behavior with known fraud patterns. Using technology such as fuzzy logic, a bank's fraud expert can use the information to create rules. Some of these rules are typical if-then rules: If a credit card is used in two different locations within a short period of time, at least one of the purchases is fraudulent. But a rule as simple as this quickly becomes complex in the details. The rule applies only to transactions where the card was physically used. It does not apply to internet shops. And what is the minimum time permitted between purchases made with the same credit card in two different locations that are far apart? This is where the advantage of human expertise comes into play. While machine learning algorithms can identify only one combination of countries as suspicious, the expert sees the bigger picture. Furthermore, a knowledge-based system offers another big advantage: If an expert identifies a new fraud pattern, this pattern can be entered into the system by the expert to immediately stop other fraud attempts of the same nature. In contrast, the self-learning algorithm requires statistical significance. A new fraud pattern must therefore occur more than once to be considered fraudulent.
______
While machine learning algorithms can identify only one combination of countries as suspicious, the expert sees the bigger picture!
The trend is a combined system
The current trend is a combination of machine learning with a knowledge-based system. Rabobank uses INFORM's hybrid approach to monitor its digital transactions, which has enabled the bank to significantly reduce the number of fraud attempts over the last few years
Afraid of a chess computer?
INFORM uses artificial intelligent (AI) methods in many of its software products.
Dr. Ulrich Dorndorf, Chief Technical Officer at INFORM, discusses how INFORM uses artificial intelligence, the roll it will play in the future, and the opportunities and risks it presents.