Custom Modeling for Risk Management Requires Human Element, Not Just AI
While there is value in leveraging Artificial Intelligence for modeling and analytics to detect cyber threats and fraud, security professionals are still more likely to indicate that a human touch is more valuable. A recent survey found that half of organizations are making use AI or machine learning but 60 percent put more trust in findings verified by humans. Meanwhile, changing consumer patterns and the rush to work from home in response to the pandemic has likely led to higher rates of false positives.
RSA interviewed over 100 security professionals at their annual conference in February where half use AI or machine learning (ML) but 60 percent put more trust in findings and decisions verified by a person. When asked why the human element was preferred 30 percent cited human intuition, 21 percent said because of creativity and nearly 20 percent said this preference comes from people with expertise and previous experience in the arena.
This isn’t to say AI and ML are not important – they are. In the same survey, 65% said these tools allow them to focus more on preventing cyber attacks than before and 40% reported feeling less stress. Even so, the human factor is a critical component as experts are needed to setup, continually monitor, retrain and feed the AI, as well as introduce new risk factors and vulnerabilities before the AI detects them. In short, machine learning needs to be supervised, checked for quality control and supplemented with user guided training to identify threats recognized or expected by humans.
While the RSA survey focused more on AI for cybersecurity, a recent analysis from Kount on their client data underscored the importance of model supervision along with human intuition and experience to monitor and train AI and ML modeling systems for payment transaction fraud detection.
In April, Kount’s eCommerce and Fraud Trend Tracker identified a 183% increase in same day or next day shipping along with a sharp increase in buy online pick-up in store orders. Fraudsters often prefer expedited shipping, meaning rules and modeling systems often assign higher risk to these orders. With more legitimate customers paying for rush shipping, merchants need to ensure rules and model features aren’t interfering with too many good customers. This is something that might take weeks to months for ML to adapt, whereas an astute risk expert could recognize and implement this change in matter of days.
Kount also found that online sales of home office furniture and electronics increased 54 percent, gaming and streaming entertainment was up 61 percent, wine increased over 110 percent and hand sanitizer skyrocketed over 1,200 percent. Modeling features based on velocity of use of a payment card, shipping address or device would likely flag a high number of false positives related to these sudden spikes in volume. This highlights the need for fraud detection professionals to identify the industries and SKUs seeing a rise in volume and instructing model-based risk systems to account for quickly changing patterns.
For more information: