Detecting Fraud in Food Delivery: Leveraging Amazon Web Services' Random Cut Forest and XGBoost Models for Enhanced Security

This paper introduces a fraud detection model tailored for identifying fraudulent cases and anomalies within business systems. Choosing the multivariate Random Cut Forest over the univariate Isolation Forest, the model demonstrated effectiveness in detecting various anomalies and ensuring system safety. Evaluated with a limited set of confirmed fraud cases, it achieved significant results with a cutoff score of 1.4, effectively identifying instances resembling known fraud cases and flagging anomalies for further investigation. These findings refined engine rules, enhancing system security and stability. The model, leveraging XGBoost for its proficiency with complex datasets and multivariate features, integrates parameters like price, delivery fee, discounts, cash back value, and distance metrics. Notably, the models achieved strong performance, with the food model scoring 0.816 and the on-demand model 0.790 in test F1 scores. Deployment involves storing models in an S3 bucket, deploying them in SageMaker endpoints using XGBoost version 1.0-1 and ml.t2.medium instance type, with weekly re-training and deployment schedules and daily dataset updates. An AWS Lambda function ensures efficient inference processing via the API. Overall, this study contributes a versatile and efficient fraud detection model for enhancing business system security and stability.