Sentinel AI: An AI-Powered UPI Fraud Detection System

Existing UPI controls, such as PIN and device checks, rule-based monitoring, mule-account detection, and complaint-based reporting, have made the network safer. Static checks, fragmented views, and after-the-fact recovery still leave gaps in judging each payment as it happens, explaining decisions to reviewers, and adapting to new fraud tactics. Sentinel AI addresses these gaps with an explainable, behavior- and network-aware system that scores each payment before it completes. An XGBoost model estimates the risk, and a rules and threshold engine turns it into one of three outcomes, so honest customers face little friction while riskier payments receive extra checks. The key contribution is the integration of explainable scoring with administrator control and continuous monitoring. Every feature uses only information available at the moment of payment, each feature group is compared with a baseline on a time-based test set, administrators can change policy without retraining, analyst outcomes become labels for future models, and drift monitoring tracks changes in data and performance. The prototype shows how a simulated UPI app, a scoring backend, and an admin console work together as one workflow. Validation on real, authorized data, a full literature and patent review, and independent testing are needed before any claim of real-world effectiveness or novelty. Future work includes integration with real-time payment processing, banking systems, and fraud-investigation platforms through secure APIs; stronger explainability that highlights the transaction features behind each score; graph analytics, including graph neural networks, for mule accounts and coordinated fraud; adaptive learning and concept-drift monitoring; multimodal risk analysis combining transaction history, device, geography, and account relationships; advanced anomaly detection for unseen patterns; a dedicated investigation dashboard with analyst feedback; privacy-preserving collaboration through federated learning; optimization for high-volume transaction streams; and continuous evaluation with precision, recall, F1-score, false-positive rate, and latency, including monitoring across customer groups.

[1]   A. Dal Pozzolo, G. Boracchi, O. Caelen, C. Alippi, and G. Bontempi, “Credit card fraud detection: A realistic modeling and a novel learning strategy,” IEEE Trans. Neural Netw. Learn. Syst., vol. 29, no. 8, pp. 3784–3797, 2018. [2]   A. Dal Pozzolo, G. Boracchi, O. Caelen, C. Alippi, and G. Bontempi, “Credit card fraud detection and concept-drift adaptation with delayed supervised information,” in Proc. Int. Joint Conf. Neural Networks (IJCNN), 2015. [3]  C. Whitrow, D. J. Hand, P. Juszczak, D. Weston, and N. M. Adams, “Transaction aggregation as a strategy for credit card fraud detection,” Data Min. Knowl. Discov., vol. 18, no. 1, pp. 30–55, 2009. [4]  D. J. Weston, D. J. Hand, N. M. Adams, C. Whitrow, and P. Juszczak, “Plastic card fraud detection using peer group analysis,” Adv. Data Anal. Classif., vol. 2, no. 1, pp. 45–62, 2008. [5]  A. Correa Bahnsen, D. Aouada, A. Stojanovic, and B. Ottersten, “Feature engineering strategies for credit card fraud detection,” Expert Syst. Appl., vol. 51, pp. 134–142, 2016. [6]   U.S. Patent 8,458,069 and U.S. Patent 11,875,355, real-time fraud detection systems based on behavior profiles and velocity counters. [7] J. Jurgovsky et al., “Sequence classification for credit-card fraud detection,” Expert Syst. Appl., 2018. [8]   J. Jurgovsky, Context-Aware Credit Card Fraud Detection, Ph.D. dissertation, Univ. of Passau / INSA Lyon, 2019. [9]   Y. Lucas et al., “Multiple perspectives HMM-based feature engineering for credit card fraud detection,” in Proc. ACM Symp. Applied Computing (SAC), 2019, arXiv:1905.06247. [10]  V. B. Nguyen, K. G. Dastidar, M. Granitzer, and W. Siblini, “The importance of future information in credit card fraud detection,” in Proc. 25th Int. Conf. Artificial Intelligence and Statistics (AISTATS), vol. 151, pp. 10067–10077, 2022.