A Novel Deep Learning Approach for Online Payment Fraud Detection

With rapid growth of digital transaction online payment systems have become increasingly vulnerable to fraudulent activities. This paper present novel deep learning based framework for detecting online payment fraud using hybrid architecture that combines Bidirectional Encoder Representations from Transformers (BERT) based feature extraction with a Transformer Enhanced Convolutional Neural Network (TECNN). The model developed using real world dataset from Kaggle which include user profile data, transaction history, metadata such as IP addresses, geolocation. The preprocessing pipeline involve normalization of numerical attribute, encoding of categorical data, and imputation of missing values along with feature engineering to derive temporal behavioral pattern. Textual metadata is processed through fine tuned BERT model to generate contextual embedding which help to enhance the representational power of the feature set. These embeddings along with numerical and categorical feature are passed through convolutional layers to detect local patterns and transformer module to capture long-term dependencies across transaction sequences. The model leverage multi-head attention to prioritize anomalous feature followed by fully connected layers for binary classification into fraudulent and authentic transactions. The aim of integrated approach to improve robustness, accuracy of fraud detection systems while addressing challenges like data imbalance, feature correlation, concept drift in dynamic financial environment. Preliminary results shows potential for significant performance gain over traditional models, establishing foundation for secure, intelligent payment processing in modern e-commerce systems.