Fraud Protection Through Artificial Intelligence and Machine Learning

Fraud Protection Through Artificial Intelligence and Machine Learning

The payments industry has ridden the coattails of technology through the last 50 years. Methods for purchases, bill payment, and personal fund transfers have evolved with the times, and ecommerce has done the same. Catalog and phone shopping became online shopping. Online shopping morphed into omnichannel. And now you can use digital payments to pay for goods in the real world.

Where payments and shopping go, of course, fraud isn’t far behind. Today, bad actors don’t even need a physical credit card to exploit payment providers, leaving merchants scrambling to fight deception. Fraudsters are using advanced technology tools to fool merchants and collect the information that they need to perpetrate their schemes.

Fraudsters aren’t the only ones who can use technology to their advantage, however. Merchants need ways to improve and scale their fraud operations that are both accurate and reliable. Plus, they need to keep pace with the ever-evolving criminal world looking to exploit payment technology and ecommerce vulnerabilities. Artificial intelligence (AI) and machine learning (ML) provide a significant piece of the puzzle for retailers looking to address fraud in online transactions.

Why Traditional Fraud Protection Is No Longer Adequate

Ecommerce merchants are not unfamiliar with fraud solutions. While there’s a lot that can be done to protect yourself from online scams, you need to be careful with what methods you rely on. To start, many ecommerce platforms come with rudimentary fraud detection included or have plug-ins that can be used in conjunction with their application. There was a time when these traditional fraud prevention solutions provided an adequate level of protection. 

Unfortunately, that time is long past for a number of reasons. To start with, traditional fraud solutions require a fair amount of configuration and setup to be remotely effective. If you’re an online bookseller, your average sale is likely to be much less than that of a luxury watch dealer. A one-size-fits-all filter that flags transactions over $300 will work for the bookseller, but not for the watch retailer.

In fact, the rules used to evaluate a transaction can pose a problem in a linear, static solution, even if the rules can be fine-tuned by the seller. Many traditional systems are binary—a transaction is good, or it’s not. If the transaction violates one of the configured rules, the transaction is declined. A purchase may have several rules that apply to it, including those that automatically approve the transaction and others that trigger a decline. Which order the rules apply will impact whether the purchase is approved or not.

Plus, with the vast array of rules needed in the modern ecommerce marketplace, rule management can be challenging. More complex rules may lead to more approvals, but they are also time-consuming and unruly to keep updated.

All of this can lead to more declines than are warranted. As a merchant, you may think that it’s better to decline a good transaction than risk approving a fraudulent one, but you’d be wrong. A false decline is far more costly to a retailer than a fraudulent one—on average, a false decline can run a merchant four times what a fraudulent purchase would. 

In a joint report from Sapio Research and ClearSale, 2021 Global Ecommerce Consumer Behavior Analysis, 39 percent of respondents indicated that they would never place an order again with a merchant that declines their payment. Furthermore, respondents reported that they would post a comment about the merchant on social media if their payment was declined. In contrast, only 13.6 percent said they would avoid a retailer after having a fraud experience with them.

The alternative, of course, is to have each transaction examined for fraud indicators by a human. That solution is simply not scalable, though. You’ll never be able to review every transaction, and without skilled and experienced fraud professionals looking at the data, it’s too easy to miss a risky purchase.

Clearly rule-based, traditional fraud tools can cause more problems than they solve. 

How Machine Learning Helps Merchants Keep Up With Fraud

There is an alternative, however. Machine learning—a sub-element of AI—excels in the very areas where traditional fraud detection falls short. It can handle complex rules sets and help scale fraud teams while lowering the risks of false declines.

Machine learning handles the complexity and nuance of risk rules in more sophisticated ways. Instead of evaluating a transaction against a binary rule set, ML assigns a score to the transaction’s characteristics that are being evaluated. 

For example, a traditional detection process may identify an order coming from a specific zip code as being risky and decline the purchase. The rules are applied linearly, so it’s a black-and-white decision. A transaction may have characteristics that would cause it to be declined—like our zip code example—but have others that would cause it to be automatically approved. If a rule matches, the action is triggered, and the outcome is dependent on which rule is matched and applied first.

With ML, however, the geographic information can be weighed against other factors, like customer history and purchase amount. This moves fraud detection from black-and-white evaluations into the gray areas in between the two extremes. Plus, these rules are applied simultaneously, not in sequence as they are in a traditional fraud detection system. This is done by computing a risk score.

Machine learning applications for fraud detection assign a value, or score, to each risk factor. Each characteristic is weighted based on its likelihood to indicate a fraudulent transaction and scored accordingly. A baseline score for approvals can then be used to approve or decline a purchase. If your score range is 0 to 100, you can say that any transaction that scores an 80 or above on the weighted scale is automatically approved.

These evaluations simulate the kind of heuristic evaluation performed by a human. The application doesn’t have the intuition that a human has, so the process is a near approximation. But it can be done in milliseconds, so massive amounts of transactions can be evaluated with a sophisticated ruleset very quickly. As we’ll see later, there are ways to improve the process even more.

The ML detection process has another advantage over traditional fraud applications, and it’s right in the name—it can learn. Pattern matching is AI and machine learning’s superpower. They are far more efficient at processing and understanding large volumes of data than humans are. To do it well, though, the algorithm must look at massive amounts of information to spot the patterns. 

As time goes on, the ML application can reevaluate the data it’s given. As new information is provided, the algorithm gets smarter and can even spot new patterns. So, over time, a machine learning fraud detection application will be able to score transactions more accurately and at the same time find new patterns that point to fraud.

This means that, as fraudsters develop new techniques and strategies, a machine learning detection application can learn those strategies and adapt its evaluation process to take those into account. A traditional fraud prevention system would require that the human administering it first identifies the new strategy and its characteristics and then reconfigures the application to account for those. 

This is incredibly important in today’s fraud environment because the criminals are also using AI to advance their techniques. Merchants need to leverage technology to keep up with and ahead of these fraudsters, and machine learning fraud protection applications give retailers that ability.

Part of a Whole: Hybrid Fraud Solutions

Machine learning can accelerate transaction review using a much more fine-grained evaluation system, significantly lowering the risk of declining a legitimate purchase or letting a fraudulent one through. It takes the evaluation from black and white to a much smaller range of gray. There are still some transactions, however, where the AI won’t be able to determine legitimacy with a high degree of accuracy.

Where the real power of AI in fraud prevention comes in is when it’s recognized for its ability to scale human fraud expertise, not replace it. A hybrid model of machine learning detection and expert human evaluation lets the strength of one method cover the weaknesses of the other.

Let’s examine this in the context of the previous example. If you have a system that provides a weighted score for transactions on a scale from 0 to 100, you can define approved transactions as those that exceed a score of 80. However, with a human fraud team available to review questionable transactions, you could change that. 

Instead, you could define the system as approving every transaction that scores a 90 and decline those with a score below 50. Any transaction that falls in between those values could then be flagged for human review. You’d be automatically approving transactions with a high level of confidence in their legitimacy, but not automatically declining those that may require further review. At the same time, you’re scaling your human review process because the majority of transactions won’t need intervention by experts.

With these combined, you would approve an immense number of transactions quickly while avoiding false declines. You’d also be lowering your fraud risk and minimizing losses caused by a wide range of fraud techniques, automatically evolving your evaluation criteria over time.

Fraud Protection Is Within Your Reach

AI and machine learning are taking automated fraud detection to a whole new level, which works in conjunction with the tools businesses and consumers use to protect themselves. The examination of transactions for risk factors can now be more nuanced and more accurate and can approve far more transactions at volume. Moreover, AI and ML can minimize costly false declines.

Weighted risk scoring can speed along the evaluation of all but the most complex transactions. When ML is combined with experienced fraud analysts, merchants can be confident that nearly every approved transaction is a legitimate one. The exceptional news for online retailers is that fraud protection that uses machine learning is available and integrates with the majority of ecommerce platforms. Scalable fraud protection, reduced losses, and increased revenue are accessible and available for online retailers of all sizes.

 

By Rafael Lourenco is EVP and Partner at ClearSale

Rafael Lourenco is Executive Vice President and Partner at ClearSale, a global card-not-present fraud protection operation that helps retailers increase sales and eliminate chargebacks before they happen. The company’s proprietary technology and in-house staff of seasoned analysts provide an end-to-end outsourced fraud detection solution for online retailers to achieve industry-high approval rates while virtually eliminating false positives. Follow on LinkedIn, Facebook, Instagram Twitter @ClearSaleUS, or visit ClearSale.

Join the 7,000+ Companies Using Logical Position

Sign up for email updates

Spam? No. Cool articles. Yes! Get notified on the latest news and tips from Logical Position.