Copyright © 2026 Authors retain the copyright of this article. This article 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.

@article{193446,
        author = {Kaushal Mahajan and Allauddin Ansari and Kartik Dhar and Sachin Narkhede},
        title = {FinSentinel: A Hybrid Machine Learning and Graph Neural Network Framework for Real-Time Financial Fraud Detection},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {12},
        number = {10},
        pages = {352-358},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=193446},
        abstract = {The growing sophistication of financial Digital payments are susceptible to fraud that requires the implementation of fraud detection mechanisms that are not limited to traditional rule-based approaches. This paper presents FinSentinel: a hybrid system of machine learning (ML) and Graph Neural Networks (GNN) for real-time fraud detection in banking transactions. FinSentinel's architecture employs three independent fraud detection modules running concurrently: (1) Random Forest Classifier for supervised classification of transactions, (2) Isolation Forest for unsupervised anomaly detection, (3) Graph Topology Analysis Engine for the identification of relational fraud patterns (e.g., money mule networks, circular trading loops, and device farm clusters). These three detection modules provide input to the Fraud Detection Framework via a Graph override mechanism at the weighted ensemble aggregation layer, which prioritizes structural indications of fraud. The implementation of FinSentinel is achieved using a FastAPI backend, a PostgreSQL Feature Store, and a Stream lit-based administration dashboard. An analysis of a synthetic database of 105,164 transactions, which was modelled on India’s Unified Payments Interface (UPI) and includes six separate types of fraud, resulted in a weighted ensemble framework accuracy of 96.1%, weighted ensemble precision of 92.8%, weighted ensemble recalls of 95.4%, and an average end-to-end latency of 187 ms. Each of the six types of fraud has a test data detection rate of over 85%. The results indicate that combining supervised, unsupervised, and graph-based methods into a single framework significantly outperforms any of the respective individual models, thereby providing a viable and deployable solution to detecting financial crimes.},
        keywords = {Anomaly Detection, Financial Fraud Detection, Graph Neural Networks, Isolation Forest, Machine Learning, Random Forest, Real-Time Transaction Monitoring, UPI Fraud},
        month = {March},
        }