FedUpi: Federated Learning for Fraudulent UPI Transaction Detection

Abstract

The rapid adoption of the Unified Payments Interface (UPI) in India has revolutionized digital transactions while simultaneously catalyzing a surge in sophisticated financial fraud. Traditional detection systems, often reliant on rule-based engines or basic machine learning models like Random Forest, frequently fail to capture the complex, non-linear dependencies inherent in high-cardinality data such as Merchant IDs, VPA handles, and Device Fingerprints. This research proposes a deep learning-based approach using the TabTransformer architecture to enhance detection accuracy. By utilizing a multi-head self attention mechanism, the model maps categorical features into robust contextual embeddings, enabling it to learn intricate relationships between user behavior and transaction attributes. The framework encompasses rigorous data preprocessing with velocity feature engineering, class imbalance mitigation via SMOTE, and real-time inference simulation. Evaluated using Precision-Recall curves and F1-scores, the results demonstrate that this attention-based approach significantly outperforms traditional gradient-boosted trees, providing a scalable, robust, and highly accurate solution for securing the modern fintech ecosystem against evolving fraudulent patterns.

Introduction

The widespread adoption of digital payment systems, particularly UPI (Unified Payments Interface), has transformed financial transactions by making them fast, convenient, and cost-effective. However, this growth has also led to an increase in fraud-related activities such as phishing, account takeovers, and transaction manipulation. Traditional fraud detection systems primarily rely on centralized databases and supervised machine learning models trained on historical fraud patterns. While these systems can effectively detect known fraud types, they often struggle to identify new and evolving attack methods and raise concerns regarding user privacy due to centralized data collection.

To address these challenges, the proposed FedUPI system introduces a federated learning-based fraud detection framework. Instead of sharing sensitive transaction data, multiple banks collaboratively train machine learning models while keeping customer data within their own infrastructure. This approach preserves privacy while enabling institutions to benefit from collective learning. FedUPI combines federated learning with anomaly detection techniques to identify both known and previously unseen fraudulent transactions in real time.

The literature review highlights the effectiveness of various machine learning techniques in fraud detection. Supervised models such as Random Forest, XGBoost, and Deep Neural Networks perform well for detecting known fraud patterns, while unsupervised methods like autoencoders and anomaly detection algorithms are useful for identifying unusual transaction behavior and zero-day fraud. Federated learning has emerged as a privacy-preserving alternative to centralized learning, allowing organizations to collaboratively train models without exposing confidential data. Additionally, ensemble learning techniques improve detection accuracy and reduce false positives by combining predictions from multiple models.

FedUPI is designed as a distributed fraud detection ecosystem with several key layers:

1. Data Acquisition Layer – Collects transaction records from each bank's UPI logs while keeping data on-premise.
2. Preprocessing and Feature Engineering Layer – Cleans and transforms raw transaction data into structured features suitable for machine learning.
3. Federated Training Layer – Trains local fraud detection models at each bank and shares only model updates with a central server.
4. Ensemble Decision Layer – Combines outputs from multiple models to improve fraud detection accuracy.
5. Real-Time Visualization Layer – Provides analysts with live alerts, risk scores, and transaction monitoring dashboards.

The system follows a hub-and-spoke federated architecture, where a central coordinating server distributes the latest global model to participating banks. Each bank trains the model locally using its own transaction data and returns only the updated model parameters. The central server aggregates these updates to create an improved global model, which is redistributed for further training. This iterative process continues until the model converges.

Conclusion

This paper presents FedUPI, a federated learning-based fraud detection framework for UPI transactions. The system successfully combines privacy preservation with effective fraud detection. In conclusion, this project successfully presents an intelligent fraud detection system using the TabTransformer model to identify fraudulent transactions with improved accuracy and efficiency. By leveraging advanced machine learning techniques, the system is capable of analyzing complex relationships between transaction features such as amount, location, and user behavior. The model shows clear learning through decreasing loss values during training, indicating effective pattern recognition. To enhance real-world applicability, rule-based logic is integrated with model predictions, significantly reducing false alarms and improving reliability. The system also addresses the challenge of imbalanced datasets, which is a critical issue in fraud detection scenarios. Furthermore, the use of probability-based outputs allows for better decision-making rather than simple binary classification. Overall, the proposed solution is scalable, adaptable, and suitable for real-time applications in digital payment systems. With further improvements such as larger datasets, continuous learning, and deployment in live environments, this system has strong potential to contribute to more secure and trustworthy financial transactions.

References

[1] K. D. Hartomo et al., \"A Novel Weighted Loss TabTransformer Integrating Explainable AI for Imbalanced Credit Risk Datasets, \" IEEE Access, vol. 13, 2025. [2] R. Rethisha et al., \"Leveraging Machine Learning Techniques of Real Time Detection of UPI Fraud, \" 2025 7th Int. Conf. on Intelligent Sustainable Systems (ICISS), IEEE, 2025. [3] R. Rani et al., \"Secure UPI: Machine Learning-Driven Fraud Detection System, \" 2nd Int. Conf. on Device Intelligence (DICCT), IEEE, 2024. [4] X. Huang et al., \"TabTransformer: Tabular Data Modelling Using Contextual Embeddings, \" arXiv preprint, 2020

Copyright

Copyright © 2026 Abhishek Sreeji, Akshay G, Amal Shajahan Rawther , Aswin Krishna, Subina S B, Mrs. Asha Liza John . 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.