RFM RiskRadar

Any successful approach to risk detection must be adaptive as new threats are continually emerging, not only from malevolent outsiders, but also from employees, customers, and even partners and suppliers. RFM has integrated the latest techniques in machine learning and applied its world-leading expertise to develop a multi-layered risk detection framework that not only detects existing threats with precision, but continuously adapts to detect new ones.

A multi-layered approach to risk detection

Click to enlarge

Being able to deliver both precision as well as self-adaptability requires an integrated mixture of technologies that not only work together to drive event detection, but also dynamically learn from one another as new potential threat scenarios are encountered.

Rules-based event prediction and detection.
Many events such as fraud, attempts at money laundering, or defaults on credit facilities are well understood and respond well to rules-based approaches such as Bayesian networks and decision trees. The first layer of protection provided by RFM analyzes data using the integrated, high-performance World Modeler engine that executes a variety of rules-based models and allows them to be easily updated as new event signatures are discovered.
Supervised deep learning pattern detection.
While offering a high degree of precision (i.e. low false-positive rates), rules based approaches do not automatically adapt to new threat scenarios. As the second layer of protection, RFM also process data through a deep learning model that is trained using data from threat detected by the rules-based systems. Once trained, the model can detect a broader range of threats than the rules-based approach.
Unsupervised deep learning auto-encoder for anomaly detection.
Completely new scenarios will not be recognized as potential threats by either rule-based or supervised learning approaches. To detect previously unseen potential significant events, a third layer of neural networks is employed. This network is trained using only data that contains no known threat events and provide a baseline encoding for “normal” activity. When trained and deployed as a detector, any events that the network fails to encode as “normal” are treated as suspect and analyzed further. Those that are found to be genuine threats are labeled, and used as training data for the supervised learning detector.

Access the world's best minds in risk, data, and machine learning

Click to enlarge

RFM has a top-shelf team of professional consultants who can assist you over the entire development, deployment, and operational stages of the solution lifecycle. RFM’s senior data scientists, risk analysts and software engineers each have careers in the commercial sector that average well over a decade. Our team combines the highest levels of technical sophistication with a focus on applying the technology that generates the optimal return to the business.

Right: KNN medioid cluster analysis of social media connections between potential fraud actors

Username:
Password: