AI-DRIVEN REAL-TIME FRAUD DETECTION USING KAFKA STREAMS IN FINTECH

Abstract

The volume and complexity of the financial transactions are continuously go- ing up, calling for a stronger case toward real-time fraud-detection systems in FinTech. Existing fraud detection systems used in FinTech often struggle to find constantly evolving pattern, leading to missed fraudulent activities or false positives. This work presents an implementation of a real-time fraud detection system meant to be used in FinTech settings. The stream data system efficiency comes from Kafka Streams, and the machine learning mod- els evaluate the transactions for scoring and verification. Transactions are pushed into the system via Kafka producers and then get processed inside Kafka Streams, which enriches the transactions with various behavioral and temporal features such as transaction frequency of a user, and the average amount of claims made. Lightweight models such as Isolation Forest and XGBoost, trained on historical fraud data, then generate fraud risk scores. Depending on the scores, transactions are categorized as low risk (permitted to continue), medium risk (flagged for further review), or high risk (blocked outright from proceeding). To evaluate the system, we replayed a publicly available credit card fraud dataset as a real-time transaction stream. The system processed over 500 transactions per second while maintaining an av- erage latency under 250 milliseconds. The models achieved a precision of 94% and recall of 92%, accurately identifying most fraud cases while mini- mizing false alarms. These results demonstrate that the system is both fast and reliable enough for deployment in real-world FinTech applications. Un-like other batch or micro-batch frameworks, such as Apache Spark or Flink, our system uses Apache Kafka Streams to perform fraud detection in real time with high scalability and low latency. The system is integrated with lightweight machine-learning models deployed as microservices.