The first area most people think of with fraud is finance. That extends past scammers and includes a wide range of attacks including banking and trades. There has been much discussion on how artificial intelligence (AI) is being used to address wider areas of fraud, such as in pharmaceutical prescription fraud. Last year saw a phenomenal growth in the use of online marketplaces and delivery services. The growth of fraud in those areas also increased. AI is being applied to the problems for the same reason it has been used in other areas.
One of the simplest methods of fraud in an online marketplace is to register a number of accounts and to publish fake listings. Standard procedural techniques often fail for the same reason they didn’t work well with email spam – fraud is indicated by the text in the post, and AI is needed to understand the text and to categorize the posts for potential fraud.
The gig economy tightly linked to ecommerce and is another area where modern fraud needs to be addressed. Multiple platforms have been created for customers, businesses, and delivery drivers, with restaurant services being a perfect example.
The biggest challenge with identifying financial and related fraud is that it is rare. The technical term
is that is it sparse. Another critical aspect of fraud is that fraudsters are constantly inventing new methods. Rule based systems and supervised learning AI fail when something outside of the rules or labeled events happen. This is why an unsupervised model is always a critical component in fraud, as it can detect anomalous transactions without knowing what they are.
An interesting area to analyze is that of ecommerce and physical delivery. It was a growing sector before the pandemic, but the last year has seen a massive increase that is attractive to inimical actors. “Our customers who are ecommerce platforms have seen an increase in fraud,” said Yinglian Xie, CEO of DataVisor. “Organized groups have, for instance, coordinated between buyers, sellers and delivery companies, to make fraudulent purchases that are never delivered, which drive payments from the platforms.”
It is important to identify fraudulent transactions as quickly as possible, but there is the other side – false positives. Part of good customer service is not halting a customer’s transaction because it falls outside her normal purchase pattern. One way of doing that is to realize that fraud rarely happens from a single account. AI can look for patterns such as recognizing a number of accounts being created in a short time from the same or similar IP addresses. The result can be a closer focus on transactions by those accounts than by isolated or existing accounts.
That methodology, however, isn’t perfect (as if there is a perfect methodology…). The type of platform or industry will impact how the AI can respond. While the goal is to have auto-actions whenever possible. “Social media platforms have a very high volume of communications,” said Ms. Xie. “It is not realistic to have significant human review of those, so the industry primarily wants auto-action. For financial institutions, given transaction importance, they are more open
to manual review to limit false positives. The trend and customer experience requirements, however, are also increasingly pushing all transactions to take place in real time without manual review.” eCommerce falls in between the two other examples.
The core of the system is neural networks trained with unsupervised learning. However, as is oft mentioned, AI is a tool, or even a tool belt, and that isn’t the only AI used. Rules do exist, and can simplify the understanding of patterns while also using lower compute overhead. Some standard patterns of fraud can also be used in some supervised learning. This isn’t a choice made purely for technology. The operators of existing systems are used to rules. As feature engineering is a challenging aspect of AI often overlooked by people beginning to add AI systems, leveraging existing features and rules can speed time to ROI for any system.
The DataVisor system uses a mix of technologies to analyze account creation, account usage, and order detail information to recognize potentially fraudulent behavior. While management team admits that this is of more value when entering a new market, client companies are subsidizing more players with higher rates in order to grow the markets, even mature markets see value from fraud analysis. After all, in a larger market, a smaller percentage of fraud can add up to significant money. It’s been a few years since I described similar fraud in the healthcare/pharmacy market, but the same issue exists in multiple industries.
There’s an old saying the history may not repeat, but it does rhyme. Many other sectors have dealt with fraud. Cloud companies have rushed to expand their landscape, and they are seeing some of the repeat benefits and challenges seen in more mature industries. Last year, the push for more online ordering and delivery created a massive expansion to ecommerce models, and added an increased risk. Artificial intelligence is being applied to this sector to address risk in similar ways as has already begun in other industries.