TransSecure: Transformer-Based Anomaly Detection with Self-Supervised Learning

Abstract

Financial fraud detection continues to be an important challenge as a result of changing fraud schemes and high-dimensional transactional information. This work introduces TransSecure, a Transformer-based anomaly model with self-supervised learning incorporated for financial fraud detection. The model employs a Masked Transaction Model (MTM) for pretraining with masked financial data to enhance its capacity to detect fraudulent activities. Self-attention mechanisms allow for the identification of short-term and long-term fraud patterns by modeling intricate dependencies in transaction sequences. The approach is tested on a large-scale Fraudulent Transactions Dataset with 99.31% accuracy, 99.54% precision, 98.08% recall, and an AUC-ROC of 0.9934. Experimental results show that TransSecure effectively minimizes false positives and negatives compared to conventional machine learning and deep learning models. This research demonstrates the power of self-supervised Transformers in detecting financial fraud and offers insights into actual fraud prevention methods.