FraudGuard: AI-Powered Fraud Transaction Detection

💡 Inspiration

Every year, billions of dollars are lost to payment fraud worldwide. According to recent reports, global fraud losses exceeded $485 billion in 2023 alone — affecting individuals, businesses, and financial institutions at an unprecedented scale.

As Computer Science students, we asked ourselves:

"Can Machine Learning detect fraud before the damage is done?"

This question became the foundation of FraudGuard — a real-time, AI-powered fraud detection system that analyzes transactions instantly and flags threats with high accuracy.

---

🔨 How We Built It

We built FraudGuard in three layers:

1. Machine Learning Core

We trained and evaluated multiple ML models on the Kaggle Credit Card Fraud Dataset — a highly imbalanced dataset containing 284,807 transactions, of which only 0.172% are fraudulent.

Models evaluated:

• Random Forest
• XGBoost ⭐ (best performer)
• Isolation Forest
• MLP Neural Network

To handle class imbalance, we applied SMOTE (Synthetic Minority Oversampling Technique).

Our top model — XGBoost — achieved:

$$\text{AUC-ROC} = 0.98+$$

$$\text{Precision} = 94\%+, \quad \text{Recall} = 91\%+$$

2. Web Platform

We developed and deployed a live web application at fraudguard.live where users can:

• Input transaction details
• Get instant fraud/legitimate prediction
• View confidence score of the ML model

3. Android Application

We published FraudGuard on the Google Play Store — making fraud detection accessible directly from mobile devices.

---

📚 What We Learned

• How to handle severely imbalanced datasets using SMOTE
• Practical differences between supervised vs unsupervised anomaly detection
• End-to-end ML pipeline: data → training → deployment → production
• Building and publishing a real Android application
• Deploying a live web service with ML model integration

---

⚡ Challenges We Faced

Class Imbalance

With only 0.172% fraud cases in the dataset, naive models predicted everything as legitimate. Solving this with SMOTE and careful threshold tuning was our biggest technical challenge.

Real-Time Prediction

Integrating the trained ML model into a live web and mobile environment required careful optimization to ensure instant predictions without latency.

Model Selection

Every model had trade-offs between precision and recall. In fraud detection, a false negative (missed fraud) is far more costly than a false positive — so we tuned accordingly.

---

🚀 What's Next for FraudGuard

• Real-time bank API integration
• Deep Learning models for sequential transaction patterns
• Multi-language support for global accessibility
• Dashboard for financial institutions to monitor transactions at scale

---

Built with ❤️ by Team TriGuard — The Islamia University of Bahawalpur

Built With

• android
• firebase
• flask
• github
• imbalanced-learn
• isolationforest
• keras
• kotlin
• matplotlib
• numpy
• pandas
• python
• randomforest
• scikit-learn
• smote
• tensorflow
• xgboost