Business Case: Real-time fraud detection platform

Task

Fraud scams are doubly painful for any company, because not only does it cost them money to fall for fraudsters, but in the worst-case scenario, they also lose the relationship of trust with their customers. At the same time, cases of fraud often only account for an extremely small proportion of transactions, making it easy for them to get lost in the crowd. So how can you protect yourself? Our customer has set up a department to manually analyze all transactions in order to identify cases of fraud. This approach was time-consuming and proved error-prone because it is difficult for people to examine and compare each one in detail in the context of such a large volume of transactions.

This is where we came into play. The challenge was to develop a model that could automatically detect higher-dimensional fraud patterns in transactions in real time. As soon as a critical transaction occurs, an alert should be sent to the customer in real time, enabling them to take immediate action. In addition, the model should be able to make a statement about the relevance of the alerts in order to prioritize the critical transactions.

Data

The fraud detection model is based on two different data sources. On the one hand, static data from the customer's data warehouse, which serves as a history to better understand patterns and trends in connection with transactions. And secondly, real-time data that is extracted from Kafka streams and contains the current transaction data. The integration of real-time data and static data through ETL processes (Extract, Transform, Load) and Kafka consumer was a crucial aspect of the project. Only by integrating the real-time data was it possible to implement this project and its requirement for a reaction (alarm) within a few seconds of a transaction.

Analytics

The fraud detection platform uses a model based on a combination of heuristics and machine learning to identify suspicious transactions.

The heuristics consist of predefined rules that have been developed to map the manual process of fraud detection into an automated process. These rules are based on known patterns, indicators and customer experience.

As the machine learning model based on this could not perfectly assess the relevance of alarms at the start of the project due to a lack of data on existing alarms, feedback loops were introduced. A dashboard enables the customer to evaluate the incoming alarms and provide feedback. This feedback is fed back into the machine learning model and is used to improve the model's assessment of the relevance of alerts.

The use of Redis as a cache database and regular backups ensure fast processing and comparability of transactions.

Solution

Our fraud detection platform is revolutionizing the way the customer deals with fraudulent transactions. By integrating real-time data and machine learning, we offer a solution that not only enables immediate responses, but also continuously improves. The use of our hybrid model of proven heuristics and a machine learning algorithm has led to a significant reduction in losses and costs. In addition, our approach enables rapid adaptation to new fraud patterns and trends while optimizing the use of resources by only having to manually review relevant alerts. The platform is designed to handle a high frequency of transactions and scales with increasing data volumes.