Want to read more on cybersecurity? Check out our comprehensive guide analyzing the trends and themes impacting cybersecurity in 2017 and beyond.
Payment companies and retailers have been obsessing over the transition to EMV chip cards and payment terminals for the last couple of years, and with good reason: EMV has been seen as a significant weapon in the fight against transaction fraud. But even as the sector is finally making progress on that transition, it’s a good time to remind ourselves that EMV alone — and even in combination with more manual fraud screening processes — can’t banish fraud entirely.
Furthermore, fraud concerns are no longer limited to transactions that occur in stores. Several studies have suggested in the last year e-commerce transaction fraud — a problem which may have gone under-acknowledged as retailers and payments players were more focused on EMV — is spreading rapidly.
Artificial intelligence, and specifically a type of AI technology called machine learning, can bring a new weapon to this fight. While in the movies, intelligent machines often wear black hats (cases in point: HAL 9000, The Terminator and Transformers), in reality artificial intelligence can be deployed in many positive and productive applications, including fighting the guys in the black hats.
Many companies in the sector may think of AI as a tool for customer interaction, as demonstrated by chatbots or Amazon’s Alexa virtual assistant. But in payments and retail specifically, AI can be applied to bring new levels of constantly-improving analytical capabilities to the ongoing fight against payment and transaction fraud, both in stores and online.
Michael Manapat, head of machine learning at payments firm Stripe, believes AI offers significant advantages over current strategies. “Machine learning allows computers to train themselves on data, and apply learnings to new situations, without intervention or instructions from human programmers," Manapat told Retail Dive.
What machine learning brings
Fraud detection and prevention strategies traditionally have relied heavily on large-scale manual screening programs in which teams of people trained to recognizes patterns likely indicative of fraud have been assigned to review transactions. But with hundreds, thousands or even millions of transactions to review each day, manual screening is an expensive and massively time-consuming approach to detecting fraud.
It also has the potential to be ineffective, as human screeners may be likely to impose hard and fast rules on all transaction patterns when the reality of transactions and the reasons and conditions around them can be much more complex. In truth, there may be many more signals and more subtle patterns at play than humans can (or are willing) to evaluate.
Not so for AI. “The ability to quickly glean insights from massive amounts of data helps to reduce costs relative to leveraging human analysts, while also reducing the time to respond to threats,” Al Pascual, senior vice president, research director and head of fraud and security at research firm Javelin Strategy & Marketing, told Retail Dive. Also, as a result of learning from each new situation and data set, such solutions may help lower potential losses from new fraud threats, Pascual added.
New tools in the fraud fight
Machine-based analytical approaches to fighting fraud are not necessarily a new idea, of course, and some amount of machine-based analysis has supplemented human fraud screening in recent years. However, Pascual suggested that an AI renaissance has occurred over the last couple of years, during which time machine learning capabilities have greatly improved their ability to analyze vast troves of data to render smarter decisions more quickly. Those improvements have been significant to developing market-ready tools, and payment companies have jumped aggressively at the opportunity to do just that.
Stripe's Radar fraud monitoring solution, which leverages machine learning to improve screening for e-commerce fraud activities, launched in October 2016 after years of in-house development. Manapat said much of that work involved training and refining Radar’s machine learning algorithms to analyze a multitude of transaction signals and patterns.
“For the few months prior to launch, we were focused on building out the breadth and depth of our models, ensuring that the signals we were measuring were actually indicative of the fraud we were trying to prevent,” he said. Stripe also had to perfect Radar's ability to check fraud probability on transactions in real time, while also doing complex background computations and constantly updating its learned analytical methods.
Creating a real-time analytical tool wasn’t Stripe’s only consideration, however. It also wanted to build a solution that was designed for transactions occurring in a globally distributed e-commerce environment where card-not-present purchases can come from anywhere.
“The decision [to develop Radar] really came from a recognition that old ways of combating fraud weren’t designed for modern internet businesses, which are often global from the outset,” Manapat said.
Stripe isn't the only company to recently launch an anti-fraud tool that leverages machine learning: Mastercard’s Decision Intelligence rolled out in late November. Mark Wiesman, senior vice president for security and decision products at Mastercard, said the company is aiming the solution particularly at the problem of false declines at in-store or online checkout, which result from inadequate fraud scoring processes sometimes leading to payment cards being declined in cases when they should have been accepted.
“Decision Intelligence turns thousands of data points about a given account into one decision score to help issuing banks approve more genuine declines,” Wiesman told Retail Dive. “More genuine transactions approved means less false declines at point-of-sale, and overall better shopping experience.”
Like Stripe, Mastercard made the move to a machine learning-driven solution as a response to dated decision-scoring products that imposed the same rules on every transaction. "Current decision-scoring products are focused primarily on risk assessment, working within predefined rules," Wiesman said. "Decision Intelligence takes a broader view in assessing, scoring and learning from each transaction. That score then enables the card issuer to apply the intelligence to the next transaction."
The new tools from Stripe and Mastercard likely only represent the beginning of an eventual industry-wide adoption of AI capabilities to fight fraud an improve overall security. Both companies said the goal with their machine learning-based tools is not to replace other payment security tools, but to work in concert with technologies, like tokenization, to provide multiple layers of protection to consumers.
Ultimately, AI capabilities won't just be used in anti-fraud applications, either. There already is some evidence that they can be used to improve detection of malicious security attacks, including cybersecurity hacks that sometimes go on for months unnoticed. Last year, MIT's Computer Science and Artificial Intelligence Lab and AI technology firm PatternEx demonstrated a AI2, a platform they said proved to be significantly better at predicting cyberattacks than existing systems. It did so by leveraging both machine learning techniques and by continuously incorporating input from human analysts. MIT and PatternEx declined to say if they have discussed these findings with retailers, but told Retail Dive that many companies across a variety of industries have shown interest.
Overall, it seems like the white hat fits AI well — and at least in retail, it's starting to build a legacy as hero instead of villain.