SecurePay Shield: An Ensemble Machine Learning Architecture for Real-Time Fraud Detection and Probabilistic Risk Scoring in the Indian Digital Payment Ecosystem

The exponential growth of digital financial transactions in India has precipitated a commensurate escalation in sophisticated payment fraud, necessitating intelligent, adaptive detection systems capable of operating at scale. This paper presents SecurePay Shield, a comprehensive, end-to-end machine learning pipeline engineered for real-time identification of fraudulent transactions within the Indian digital payment ecosystem. The proposed system addresses three principal challenges endemic to fraud detection: severe class imbalance, high-dimensional heterogeneous feature spaces, and the requirement for probabilistic, interpretable risk assessments amenable to regulatory scrutiny. The architecture employs an ensemble learning strategy integrating three complementary algorithms: a Random Forest Classifier (200 estimators), a Gradient Boosting Classifier (150 estimators), and an Isolation Forest anomaly detection model. A domain-specific feature engineering pipeline transforms 23 raw transaction attributes into a 35-dimensional feature space, computing composite risk indicators including the Risk_Composite score, IP_Risk_Score, and Velocity_Score, which collectively emerge as the most discriminative predictors of fraudulent activity. Class imbalance is mitigated through the application of the Synthetic Minority Over-sampling Technique (SMOTE), yielding a balanced training corpus of 22,560 instances. The Random Forest model, selected as the production deployment candidate, achieves an Area Under the Receiver Operating Characteristic Curve (AUC-ROC) of 1.0000 and an F1-Score of 1.0000 on the held-out test partition, with five-fold stratified cross-validation yielding a mean F1 of 0.9978 (0.0021), confirming model robustness and generalizability. Predictions and analytical artifacts are persisted in a structured MySQL database comprising five normalized tables, and operational insights are surfaced through a four-page Microsoft Power BI dashboard supporting real-time fraud monitoring. The system is demonstrated on a synthetic dataset of 15,000 Indian financial transactions and evaluated against 2,000 prospectively generated records, achieving a holistic prediction corpus of 17,000 transactions with four-tier risk stratification (Critical, High, Medium, Low).