AUTOMATED GST INVOICE FRAUD DETECTION SYSTEM

About the Project

Inspiration

GST invoice fraud can create financial losses and compliance difficulties, especially for Micro, Small, and Medium Enterprises (MSMEs). We were inspired to develop a system that can automatically identify unusual invoice patterns and help businesses recognize potentially fraudulent transactions at an early stage.

What We Learned

Through this project, we learned how Machine Learning and anomaly detection can be applied to financial data. We gained practical knowledge in data preprocessing, feature selection, anomaly detection, risk scoring, and explainable fraud identification. We also learned how to convert a theoretical algorithm into a practical software solution.

How We Built It

The system starts by collecting GST invoice data and preprocessing it by handling missing values, removing duplicates, correcting inconsistent data, and preparing the required fields. Important features such as invoice amount, GST amount, tax rate, vendor details, discount, payment information, transaction frequency, and GSTIN validity are considered.

We use the Isolation Forest algorithm to identify unusual invoice patterns without completely depending on pre-labeled fraud data. The detected anomalies are then evaluated using our Weighted Fraud Risk Scoring (WFRS) approach.

The Fraud Risk Score is calculated as:

$$ FRS = (0.4 \times Amount\ Anomaly) + (0.3 \times Tax\ Anomaly) + (0.2 \times Frequency\ Anomaly) + (0.1 \times GSTIN\ Risk) $$

The resulting score is used to classify invoices into different risk levels. The system also provides the factors contributing to a high-risk classification, making the results easier to understand.

Challenges We Faced

One of the major challenges was determining which invoice features were useful for identifying suspicious transactions. Handling inconsistent and duplicate invoice data was another challenge during preprocessing. We also needed to design a risk-scoring approach that could combine multiple fraud indicators while remaining understandable to users. Integrating anomaly detection with the WFRS method and presenting meaningful reasons for suspicious invoices required several refinements and testing.

Overall, the project helped us gain hands-on experience in applying AI, Machine Learning, anomaly detection, and risk scoring to a real-world GST compliance problem for MSMEs.