A Context-Driven Artificial Intelligence Framework for Detecting Fraudulent Activities in Payment Systems

Abstract

The expansion of digital payment platforms has created a growing need for robust and reliable fraud detection systems to safeguard financial transactions. Earlier approaches to fraud detection were primarily based on manual verification, rule-driven mechanisms, and simple statistical models. These methods required substantial human involvement, were time-intensive, and struggled to identify sophisticated fraud patterns within large and continuously evolving transaction datasets. As the volume and complexity of digital payments increased, these traditional techniques became inadequate, particularly in supporting real-time decision-making. To address these limitations, this work proposes an intelligent web-based framework for comprehensive fraud detection in Unified Payments Interface (UPI) transactions by leveraging advanced machine learning and deep learning methodologies. The system is implemented using the Flask framework and incorporates natural language processing techniques to process textual transaction data. For feature representation, Sentence Bidirectional Encoder Representations from Transformers (SBERT) is employed to generate contextual embeddings that capture semantic relationships within the data. These embeddings are then used to train and evaluate multiple classification models, including Gaussian Naive Bayes (GNB), Bernoulli Naive Bayes (BNB), Multinomial Naive Bayes (MNB), and a Histogram-based Gradient Boosting (HGB) classifier. The framework supports both binary classification for fraud identification and multi-class classification for categorizing transaction types. Experimental results demonstrate that the HGB model achieves superior performance, delivering high accuracy along with strong Precision, Recall, and F1-score compared to conventional approaches. Additionally, the system provides a secure web interface that allows users to upload datasets and obtain real-time predictions, with options to download results for further analysis. Overall, the proposed solution enhances automation, scalability, and predictive accuracy, contributing to more effective and efficient fraud detection in digital financial systems.