It’s very clear to me that fraud prevention needs a serious update. For someone like myself, who travels extensively to speak at conferences, meet customers, or getting some well needed vacation, getting intrusive calls from my credit card provider asking questions about my weird travel patterns is among my least favorite things. It is very clear to me that machine learning could assist with fraud detection in a bigger way. I’m still to this day surprised to experience fraud alerts, sometimes even shutting off my credit cards, which then results in long phones calls with customer service despite them having extensive knowledge about my consistent, albeit weird, travel patterns. When studying my payment patterns leading up to the failed transaction, I am often lead to the conclusion that its often a rule based error or trigger that disrupted by transaction.
In addition, I often hear from many new startups that banks or ecommerce companies should completely throw out the rule based authentication because of the frequent false negatives or false rejects that their systems often trigger, causing great inconvenience to end customers who are left unable to pay for goods at a physical retailer or websites. Could that really be true, that all rule based fraud-detection and authentication are bad things?
No of course not, risk-based fraud detection is not a new instrument within financial markets. The FICO score was introduced in 1989 by the Fair Isaac Company, and is one of the best examples of a stable and well adopted risk based score within the financial institutes. It is made up of multiple variables or inputs things like payment history, debt burden, types of credit, and credit inquires made by the borrower. This started the trend of using early forms of neural networks and Bayesian modelling.
Today the financial institutes are phased with an ever-growing threat of global fraudsters which descend on older wealthier clients in remote access attacks, massive bot-attacks of both small and large ecommerce sites, and a much more transient customer base that travel more frequently than ever before.
Since the early 2000s and well into 2010s, we have primarily seen machine learning within supervised systems run by specialized analysts and fraud experts. The problem with these systems is that they are largely based on past observations and rely on human intelligence to design rules to fight the most current fraud trends. As a consumer, you often see these systems fail, when your credit card is shut down due to irregular transactions or, what always seems to be the reason my cards get rejected, frequent travels to far away countries. These new observations make the rule-based systems harder to maintain, since new rules are created by human experts on a regular basis, while old rules remain in the libraries clouding the judgement of the system, and ultimately, causing unnecessary false rejects.
So how do we solve this? Well lets take a look at unsupervised machine learning. These types of systems don’t rely on labeled data by humans, instead they are built from clustering data in real time. In addition to the cluster score and real-time determination of where the score falls within an individual, a smaller specific group or a larger population. It also builds confidence over time, which increases reliability. It’s like having a third eye on the problem.
What I am seeing out in the market when it comes to implementation of unsupervised machine learning (like behavioral biometrics) is that it is added to websites and apps in order for the organization to start building up models of clustering. As more data goes through the system the reliability builds up and it can start to replace supervised machine learning or even older legacy systems in difficult unpredictable scenarios, such as onboarding more services in mobile channel, customer abuse, account sharing, internal collusion and corruption, employee fraud, and even complex omnichannel attacks such as social media engineering.
When implementing unsupervised machine learning is best targeted in areas of your system where the main objectives are anomaly detection but where your system will be able to feed the algorithms large amount of good quality data without explicit end user interaction.
In conclusion, when introducing unsupervised machine learning in parallel to supervised machine learning it’s important to constantly evaluate both scores and design an overall fraud system which allows both types of fraud detection to weigh into final decisions. In the early beginnings, even supervised machine learning may be perceived as successful within the organization without accounting for potential revenue losses due to false positives and false negatives. Since they depend on human intelligence and past observations, they cannot foresee future fraud attacks, since real total cost of fraud is often a grey number. To truly reduce this grey area of fraud you need unsupervised machine learning which can sift through this data and increase your revenue from good customers while also reducing fraud. It also provides future proofing for new unknown fraud behaviors yet to be discovered.
As with any security measurement there is no silver bullet, but having unsupervised machine learning, like behavioral biometrics, added to your fraud fighting arsenal improves both security and user experience, which increases revenue and gives you a much better ability to sort through large amounts of unknown data.