A Machine Learning-Based UPI Security and Fraud Detection

Abstract

Unified Payments Interface (UPI) has revolutionized digital payments in India, enabling seamless, real-time money transfers between accounts. However, its growing popularity has made it increasingly susceptible to fraudulent activities such as phishing, account takeovers, and transaction manipulation. This study introduces a hybrid fraud detection system integrating Convolutional Neural Networks (CNN) with Autoencoder, Local Outlier Factor (LOF), and K-Means Clustering to detect anomalous UPI transactions efficiently. The system utilizes anonymized UPI transaction data including transaction amount, time, device identifiers, and geolocation. Preprocessing involved encoding categorical features, normalizing numerical variables, and addressing missing data. The proposed hybrid approach was evaluated using accuracy, precision, recall, F1-score, and AUC-ROC metrics, achieving higher accuracy and fewer false positives compared to traditional methods. The findings highlight that deep learning combined with unsupervised techniques offers a robust solution for ensuring secure and reliable UPI payment operations.

Introduction

The text presents a hybrid deep learning-based system for detecting fraudulent UPI transactions in real time. The rapid growth of UPI has increased the risk of fraud, while traditional rule-based systems struggle with changing fraud patterns and often produce missed detections or false alarms.

The proposed system combines Convolutional Neural Networks (CNN), Autoencoders, Local Outlier Factor (LOF), and K-Means Clustering. CNN learns complex relationships between transaction features, the Autoencoder detects anomalies through reconstruction error, LOF identifies unusual local behavior, and K-Means groups similar transactions to isolate suspicious activity. SMOTE is used to address the highly imbalanced nature of fraud datasets.

The methodology involves collecting anonymized transaction data, cleaning and encoding the data, normalizing features, selecting important attributes, reshaping data for CNN processing, training the hybrid model, and evaluating it using Accuracy, Precision, Recall, F1-Score, and ROC-AUC. The system was also deployed through a Flask web application for single and batch transaction prediction.

According to the reported results, the CNN-based model achieved 98.85% accuracy and an ROC-AUC of 0.99, outperforming models such as Random Forest (96.29%), Logistic Regression, and KNN. Important fraud indicators included transaction amount, merchant category, transaction time, device ID, location, and transaction frequency. The confusion matrix showed relatively few false positives, while the system achieved high fraud recall.

The study concludes that the proposed hybrid approach is accurate, scalable, adaptable, and suitable for real-time UPI fraud detection, with an average response time of less than 0.8 seconds per transaction. By combining deep learning with anomaly detection and clustering, the system can identify both known and emerging fraud patterns while reducing false alarms and potentially improving the security and reliability of digital payments.

Conclusion

This work developed a hybrid Convolutional Neural Network (CNN)-based framework for detecting fraudulent activities in Unified Payments Interface (UPI) transactions. The model integrates CNN with Autoencoder, Local Outlier Factor (LOF), and K-Means Clustering, combining the strengths of deep and unsupervised learning to enhance detection accuracy. The proposed system successfully recognized complex, non-linear transaction behaviors that traditional rule-based approaches often overlook. After applying rigorous preprocessing, data balancing with SMOTE, and parameter tuning, the CNN achieved an overall accuracy of 98.85 %, outperforming classical algorithms such as Random Forest and Logistic Regression. The results confirm the system’s ability to maintain high recall while minimizing false positives, ensuring that genuine users are not wrongly flagged during real-time payments. The model was deployed through a Flask-based web application that enables live transaction analysis and batch processing. This validated its feasibility for real-time use in digital banking environments. The system’s design emphasizes adaptability, allowing it to evolve as fraud patterns change, and scalability, making it capable of managing high-volume transaction streams efficiently. Future work can explore the inclusion of Recurrent Neural Networks (RNNs) or LSTM architectures to capture time-based dependencies and improve sequential pattern detection. Additionally, incorporating Explainable AI (XAI) could increase transparency and regulatory trust. Overall, the research demonstrates that deep-learning-driven hybrid models offer a reliable and intelligent approach to strengthening the security of digital payments and enhancing user confidence in the UPI framework.

References

[1] S. Maes, K. Tuyls, B. Manderick, and B. Vanschoenwinkel, “Credit card fraud detection using Bayesian and neural networks,” Proc. 1st Int. NAISO Congr. Neuro Fuzzy Technologies, 2002. [2] P. Bentley, “Fuzzy-ID3: A novel approach for fraud detection in financial transactions,” Int. J. Comput. Intell., vol. 2, no. 1, pp. 15–24, 2005. [3] A. Kundu, P. Kumar, and S. Rani, “A hybrid sequence alignment method for financial fraud detection,” Expert Syst. Appl., vol. 42, no. 12, pp. 5483–5493, 2015. [4] S. Rani, P. Gupta, and V. Jain, “Secure-UPI: A machine learning approach for real-time fraud detection in digital payments,” J. Inf. Secur. Sci. Int., vol. 8, no. 2, pp. 90–99, 2019. [5] A. Patil, S. Kulkarni, and P. Deshmukh, “Machine learning framework for fraud detection in UPI transactions,” Int. J. Comput. Appl., vol. 182, no. 20, pp. 25–31, 2020. [6] N. Krithiga Lakshmi and S. Deepa, “Security framework for UPI-based mobile banking applications,” Banking and Financial Technology Journal, vol. 5, no. 1, pp. 45–56, 2021. [7] R. Kolekar, M. Patil, and A. Jadhav, “Security challenges and prevention techniques in digital banking with reference to UPI,” J. Adv. Comput. Sci. Appl. Int., vol. 12, no. 8, pp. 233–240, 2021. [8] V. Patil and R. Sharma, “Hybrid machine learning model for UPI fraud detection,” Procedia Comput. Sci., vol. 218, pp. 1125–1133, 2023. [9] R. Rani, A. Alam, and A. Javed, “Secure UPI: Machine learning-driven fraud detection system for digital payments,” Proc. 2nd Int. Conf. Disruptive Technologies (ICDT), pp. 104–110, 2024. [10] K. Krithiga Lakshmi, H. Gupta, and J. Ranjan, “UPI-based mobile banking applications – security analysis and enhancements,” Proc. Amity Int. Conf. Artificial Intelligence (AICAI), pp. 212–218, 2024. [11] S. Kolekar, S. Panhale, D. Rengade, D. Pawar, and P. Kothawale, “UPI fraud detection using machine learning,” Int. J. Sci. Res. Eng. Manag., vol. 8, no. 4, pp. 102–107, 2024. [12] Y. Patil, A. Shinde, Y. Parthe, and S. Sayyad, “Detection of fraudulent digital payments using deep learning,” Int. Res. J. Modernization Eng. Technol. Sci., vol. 6, no. 9, pp. 554–560, 2024. [13] I. M. Adekunle and P. Ozoh, “Fraud detection model for illegitimate financial transactions,” Kabale Univ. Interdiscip. Res. J., vol. 2, no. 2, pp. 21–37, 2023. [14] P. Boulieris, J. Pavlopoulos, A. Xenos, and V. Vassalos, “Financial fraud detection using natural language processing,” Future Gener. Comput. Syst., vol. 108, pp. 785–797, 2023. [15] B. Mytnyk, O. Tkachyk, N. Shakhovska, S. Fedushko, and Y. Syerov, “Application of artificial intelligence for detection of fraudulent banking operations,” Big Data and Cognitive Computing, vol. 7, no. 2, p. 93, 2023. [16] E. Alhassan, M. Gholami, and R. Ahmad, “Enhancing digital payment security using deep learning and blockchain integration,” IEEE Access, vol. 11, pp. 84522–84535, 2023. [17] D. Banerjee and P. Narayanan, “A convolutional neural network-based approach for online payment fraud detection,” J. King Saud Univ. Comput. Inf. Sci., vol. 36, no. 4, pp. 1221–1230, 2024.

Copyright

Copyright © 2026 Meghana L , Dr. Amutha S, Soumya Patil. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.