AI Reads the Sequence of Customer Behavior: KakaoBank's Sequence-Based FDS Model Development Story

KakaoBank introduced a sequence-based detection model that analyzes customers' app usage history and transaction records over time, in order to overcome the limitations of the existing single-transaction (point) based detection method.

The existing method analyzed only the pattern at a specific transaction point, making it difficult to respond to new types of fraud schemes. The new model approaches the problem as a Time Series Classification task, proactively capturing potential risk signals.

For the model architecture, Vision Transformer (ViT) from the image recognition field was adapted for financial time series data. This choice was made to address the limitations in capturing nonlinear patterns and the long-term dependency problem found in existing statistical models or RNN/LSTM-based models.

The data consists of four key elements: event type, time information, amount, and text information, and the model is designed to learn patterns directly from the data without separate complex Feature Engineering.

This system processes an average of over 18 million inference requests per day at an average speed of under 20ms, operating stably even in a large-scale real-time serving environment, minimizing blind spots in voice phishing detection.