Research Article
Enhanced Detection of UPI Frauds Using Advanced Machine Learning Techniques for Secure Digital Transactions
@INPROCEEDINGS{10.4108/eai.28-4-2025.2357830, author={Mangali Vijaya Sree and Panyam Bhuvaneshwari and Karne Nandini and Koppula Ratna Phoebe Amulya and Palla Swetha and B. Chandrakala}, title={Enhanced Detection of UPI Frauds Using Advanced Machine Learning Techniques for Secure Digital Transactions}, proceedings={Proceedings of the 4th International Conference on Information Technology, Civil Innovation, Science, and Management, ICITSM 2025, 28-29 April 2025, Tiruchengode, Tamil Nadu, India, Part I}, publisher={EAI}, proceedings_a={ICITSM PART I}, year={2025}, month={10}, keywords={upi fraud detection causal inference federated learning transformer networks secure digital transactions explainable ai}, doi={10.4108/eai.28-4-2025.2357830} }- Mangali Vijaya Sree
Panyam Bhuvaneshwari
Karne Nandini
Koppula Ratna Phoebe Amulya
Palla Swetha
B. Chandrakala
Year: 2025
Enhanced Detection of UPI Frauds Using Advanced Machine Learning Techniques for Secure Digital Transactions
ICITSM PART I
EAI
DOI: 10.4108/eai.28-4-2025.2357830
Abstract
The exponential rise of digital transactions through Unified Payments Interface (UPI) kind of platforms, fraud detection has become very important but challenging. Most of the existing systems need to rely on correlation-based models and depend on centralized data aggregation that are limited with their scalability, interpretability and privacy. We propose CauFedFormer, a hybrid approach that combines sequential transformer models, UPI causal inference techniques, and federated learning for UPI fraud detection. Temporal behavior anomaly capture, causally relevant features recovery, and distributed training with privacy preservation are present in the model. Experimental results show that CauFedFormer can reach precision of 86%, recall of 82%, F1-score of 84%, ROC-AUC of 93% and outperforms traditional baselines, including logistic regression and standalone transformer models. In addition, CauFedFormer proposes interpretable fraud risk score with confidentiality constraint. It thus makes for a promising candidate to be deployed in secure and scalable digital transaction ecosystems.
