AI-based 'sequence detection model' developed; nearly half of Q1 fraud cases caught independently
Kakao Bank said Tuesday it has developed an AI-based fraud detection model called the "sequence detection model" that predicts financial fraud risk by comprehensively analyzing customer behavior before and after transactions, and has applied it to its fraud detection system (FDS).
Unlike conventional methods that examined only the outcome of individual transactions such as transfers or withdrawals, the sequence model analyzes the broader behavioral context surrounding each transaction. It applies an "attention mechanism" that allows the AI to understand correlations and patterns across data points. This enables the model to connect and assess three types of behavioral cues — the order in which transactions occur, the time intervals between actions, and device-switching behavior — to more precisely detect fraud attempts disguised as normal transactions.
The model can pick up on subtle patterns, such as a brief pause in activity after a customer logs into the app and begins transacting, followed by a resumption. In many cases, such "activity interruption periods" coincide with moments when voice phishing criminals are persuading victims or prompting additional transfers. The sequence model analyzes these behavioral contexts in aggregate to assess risk.
After Kakao Bank piloted the sequence model in November last year, the monthly average number of fraud cases prevented through FDS monitoring rose 4.4-fold compared to before its introduction. In the first quarter of this year, when full operations began, the model independently detected 49.8 percent of all suspected fraud cases blocked by Kakao Bank.
Real-world detections also highlighted the model's effectiveness against new fraud tactics designed to evade existing FDS systems. A notable example is the surge in detection of so-called "voice phishing collection accounts," or ghost accounts. When funds are repeatedly deposited from multiple unrelated parties but no withdrawals follow, legacy systems struggled to flag the accounts unless rapid post-deposit withdrawals occurred. The sequence model accurately classified such accounts as ghost accounts by analyzing deposit patterns alongside time-of-use data, enabling preemptive action before the funds could be funneled to criminal networks.
The model also caught suspected cases of "device handover," in which a user changes their phone and passes the device to a criminal organization. The sequence model flagged anomalies by analyzing the device change that occurred just before a transfer, followed by subsequent app usage and transaction flow. Under the previous rule-based system, such cases would likely have passed undetected due to their resemblance to past normal transaction patterns, but the new model allowed Kakao Bank to protect the funds in advance.
Kakao Bank said it plans to continue refining the sequence model to make its FDS framework more intelligent and to stay ahead of increasingly sophisticated fraud schemes. "We are now able to detect elaborate fraud tactics in advance — ones that were difficult to catch from a single transaction record alone — by analyzing the continuous flow of behavior before and after transactions," a Kakao Bank official said. "We will continue to further enhance our detection capabilities to proactively protect our customers' valuable assets from ever more cunning financial fraud."
won@heraldcorp.com
