AI BriefingA Case Study on Applying a Continuously Growing ML Model to FDS
Key point
Kakao Pay introduced an adaptive ML model to its FDS that continuously learns and adapts itself to respond to changing fraud patterns.
Details
Kakao Pay operates FDS (Fraud Detection System) to protect a massive volume of transactions—13 million per day and over 400 million per month. Existing Rule-based detection struggles to block increasingly sophisticated fraud patterns, and typical Static ML has the limitation of degrading performance when data distribution changes.
To address this, Kakao Pay introduced Adaptive ML. Adaptive ML reflects the latest data through Continuous Train, and responds to data changes (Data Drift) on its own through Self Adaptation of features, detecting new types of fraud.
An automated MLOps system is used for model operation. Airflow periodically trains the model on the latest data, utilizing data from Hadoop and the Feature Store (HBase, Redis). The trained model's performance is compared against the existing model via Mlflow, and when a meaningful improvement is confirmed, it is immediately deployed to the Triton Inference Server.
Through this structure, even new payment fraud patterns that the existing model has not learned—such as those related to newly released game items—can be quickly detected and responded to.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
How DS and MLE Work Together