Real Time UPI Fraud Detection in Evolving Streams

Abstract— The UPI has done so, where instant payments arenow made through different banking systems. The rapidintroduction of UPI has increased the risk of fraudulentactivities and that needs to be countered with efficient real-timedetection systems. This study proposes a ML-based frauddetection framework on UPI transaction, based on differentclassification models, including LR, DT, XGBoost, SVM, KNNand Gaussian Naive Bayes. A Voting Classifier is a type ofclassifier that makes use of an ensemble approach that bringsthe benefits of Decision Tree, KNN, XGBoost, and ExtraTreeclassifiers to enhance prediction effectiveness. Preprocessing ofthe dataset includes label encoding, missing and infinite valuemanagement, the solution of the imbalance of classes usingSMOTE, and scaling the features using StandardScaler.Exploratory data analysis is conducted in order to understandinterrelations of the features and patterns of frauds. Accordingto the accuracy, precision, recall, F1-score and ROC-AUC, theanalysis of the model shows that the Voting Classifier achieveshigh results, as the accuracy, precision, recall and F1-score are99.6, and the ROC-AUC is a perfect 1.0. Both local and globalinterpretation of model predictions are made using explainableAI methods including LIME and SHAP. An SQLite-basedfrontend with Flask used to verify users helps to process allinputs to the real-time, pre-classifying and visualizing frauddetection results. The unified system provides a flexible andunderstandable way of providing security to digitaltransactions