Capitalizing AI for Safer Banking

Fraud detection solutions powered by artificial intelligence (AI) and machine learning (ML)-based systems can greatly enhance security across the banking industry. By increasing the accuracy of fraudulent activity alerts, banks significantly decrease the time and costs associated with security breaches—and secure brand credibility.

As worldwide banking transactions continue to increase, so does the threat of online fraud. According to UK Finance, in 2018 advanced security systems for the UK banking industry prevented more than £1.6B ($2.04B) of unauthorized fraud. Despite this, £1.2B ($1.53B) was stolen through fraudulent activity. Online fraud accounts for a large part of the $600B cybercriminals take from the global economy each year.

AI/ML-based fraud detection systems work best when they are trained by a human and in a similar manner as a human would be trained—with information and demonstration of what to and what not to do. However, AI/ML solutions learn in a more scalable way than human counterparts. By ingesting continuous volumes of data, AI/ML-enabled fraud detection systems adopt and learn to identify new patterns and types of fraud. The more automated a fraud detection system is, the higher the quality of data needed because the system would recognize and be familiar with different types of data funneling through accurately designed data pipelines.

Let’s take a look at some ways that banks can utilize AI/ML-based technologies to prevent fraud.

Anomaly detection

Anomaly-based fraud detection is the most common and base layer of fraud detection. This type of intelligent model requires an ML-based algorithm to track the continuous stream of incoming transactional data. It is programmed to operate exactly how a human would when conducting bank transactions, mortgage loans, and more. Many bank applications require the use of anomaly detection. AI-based algorithms can detect any unusual activity by monitoring the customer’s usage in real time, looking for any deviation from prior usage patterns.

The ML-based algorithm model is programmed with an alarm system to instantly notify the bank of any deviations from the ‘normal’ pattern. If fraud is detected, the system automatically rejects the application and notifies the bank staff via an app in real time. The ML-based model is trained through observations on purchasing patterns to ’understand’ if the deviation found is either fraud or not, and at scale. Such AI-enabled applications helped Visa Inc. prevent an estimated $25B in annual fraud, the company announced in June 2019.

Predictive analytics

Predictive analytics provide an added level of sophistication from anomaly detection. ML-based fraud detection models can be used to develop predictive analytics software to analyze data with a pre-trained ML-based algorithm.

Data experts in banks label large volumes of transactions as either fraudulent or genuine, and then run those transactions to train the ML-model. The model uses this information to quickly recognize fraudulent transactions.

For example, a fraudulent transaction alert may be issued for a product or service a customer has bought online, based on the customer’s purchasing patterns and location data. The system may flag an item that has never been purchased before which, in itself, may not be indicative of fraud. However, the AI/ML solution will also check a customer’s location—provided through geo-location data—with the site of purchase to determine the likelihood of fraud.

Predictive analytics can also be used to prevent activity from incorrectly flagged transactions as fraudulent. Through its Decision Intelligence and AI Express platforms, MasterCard used predictive analytics powered by ML to cut the rate of transactions being incorrectly flagged as fraudulent by 50 percent and prevent more fraud. Decision Intelligence uses sophisticated AI/ML algorithms to measure, score, and learn from each transaction, and provide a predictive score to the card issuer. The scoring is based on intelligent analysis of large-scale behavioral patterns, combined with specific account usage (location, time, and type of purchase) to truly identify fraudulent activity.

Predictive analytics-based fraud detection software is used to detect fraud across multiple interactive devices. These can involve payment processes for online purchases and for detecting anomalous user behavior within banking apps, using geo-locational data.

Accurate data analysis

This risk-based analytics approach enhances fraud detection by detecting complex and hidden patterns. AI/ML algorithmic models are most effective with increased availability and accessible data. The sophistication of fraudulent activity detected by AI/ML algorithms improves with the quantity of data—the more information the ML model has as sources, the better the ability to detect the differences, patterns, and similarities between multiple data behaviors.

AI-based fraud mitigation technologies take many data points into account to not only pinpoint a fraudulent transaction, but also explain why a customer’s account may have been fraudulently compromised. Factors accommodated by AI/ML algorithms include customer location, and key contextual data points of each transaction, to name just a few.

ML/AI algorithms can work through real-time and historical bank transactions to block or separate genuine transactions from fraudulent ones, without impacting the customer experience. This automated process can become more scalable for banks because AI/ML algorithms can make a more granular analysis of available data to predict the future state as information is captured from new transactions.

Anton Vaisburd is a Data Scientist at SoftServe