Artificial Intelligence for Real-Time Financial Fraud Detection

Real-time analysis of data in-database and often in-stream, rather than by pulling it out into a data warehouse or data lake and then applying model inference, is emerging in several application domains, primarily driven by the low-latency constraints of use cases such as fraud detection in financial systems and cyber-security. When transaction processing systems detect patterns of user behavior that appear suspicious, they typically take a do-not-honor strategy and decline the transaction. Such patterns, unless they are simple signature patterns that can be identified with very small false-positive rates, are hard to model during training because fraudulent activity is highly imbalanced compared to non-fraudulent activity. Therefore, anomaly-detection solutions that look toward the tail of the distribution space have a better chance of success. A hallmark feature of modern financial transaction systems is the ability to analyze a transaction before honoring it and thereby also potentially preventing a fraud. Fraud detection in that real-time operational manner differs from cold-start fraud detection, where fraud detection is a batch analysis with limited decision support for the fraudsters. In operational fraud detection, the decision support provided makes the transaction failure prediction a mark of caution and the user behavior requires examination, since a genuine user behavior may be wrongly identified as a fraud attack.