Smells Like Fraud? Applying AI to FDS Systems

The existing Rule-based FDS relied on predefined rules, making it difficult to respond to new fraud patterns, and had limitations such as the operational burden of continuously updating rules and the possibility of fraudsters evading the rules.

To address this, ML(Machine Learning)-based FDS was introduced to increase flexibility and adaptability. ML models learn from diverse data to detect unpredicted patterns, and can improve precision and recall to lower false positive and false negative rates.

To implement the system, the following 3-stage process was built.

• Real-time data pipeline: Bank data flows in real time via Kafka, and events are processed using Flink.
• Feature Store: Preprocessed data is stored and managed so it can be reused for model training and serving.
• ML model serving: The model performs inference using data from the Feature Store and provides results to customers.

In particular, to improve collaboration efficiency between data scientists (DS) and engineers, a structure was adopted where DS focus on event processing logic while engineers focus on event engineering and serving server development.