Real-time UPI fraud detection engine with per-card behavioral profiling and explainable AI.
Built for Datathon 2026 by Team PARADIGM.
- Live Demo: https://trustsentinel-paradigm.streamlit.app
- Prototype Screenshots: Available in this repository under assets/
Fraud detection in UPI / digital payments (PS-07)
Most fraud detection systems compare every transaction against a global average. TrustSentinel takes a different approach — it builds a personal behavioral profile for every single card and asks:
"Is this transaction suspicious for THIS specific person?"
A 90,000 rupee transaction is an emergency for one person and completely normal for another. Our system knows the difference.
Two-Tier Hybrid Engine:
- Tier 1 — Rule Engine: Catches obvious attacks instantly (rapid-fire transactions, extreme amount spikes, suspicious timing) with zero ML overhead
- Tier 2 — ML Engine: Random Forest trained with SMOTE on balanced data catches sophisticated fraud patterns that rules cant detect
On top of that:
- SHAP generates plain English explanations for every fraud flag
- Per-card risk scoring based on individual behavioral deviation
- Model drift detection monitors when fraud patterns shift
- Adjustable threshold lets banks control precision vs recall tradeoff
Evaluated on 118,000 unseen real transactions:
- AUC-ROC: 0.93
- AUC-PR: 0.72
- F1 Score: 0.68
- Precision: 81%
- Recall: 58.6%
- Frauds Caught: 2423 out of 4133
Streamlit dashboard with role-based access:
Bank View:
- Live transaction feed with real-time metrics
- Alert terminal with fraud history
- Transaction network graph
- Model health monitoring with drift detection
- Adjustable fraud detection threshold
Customer View:
- Personal transaction history
- Risk score gauge
- Plain English explanation when a transaction is blocked
- Clear instructions on what to do next
- IEEE-CIS Fraud Detection (Kaggle)
- 590,000 real transactions from an actual company
- 13,553 unique card profiles built
- NOT synthetic data — real-world transaction patterns
- Language: Python
- ML: scikit-learn (Random Forest)
- Balancing: SMOTE (imbalanced-learn)
- Explainability: SHAP
- Dashboard: Streamlit + Plotly
- Graph Analysis: NetworkX
- Data: pandas, numpy
- engine.py — Backend fraud detection engine
- app.py — Streamlit dashboard frontend
- requirements.txt — Python dependencies
- .gitignore — Files excluded from repo
- README.md — This file
- Clone the repo
- Run: pip install -r requirements.txt
- Download IEEE-CIS dataset from Kaggle and place train_transaction.csv in the project folder
- Run: streamlit run app.py
Login credentials:
- Bank: bank_admin / bank123
- Customer: priya / priya123 or arjun / arjun123 or sneha / sneha123
- Manaswi Mishra
- Rushda Jagtap
Built for GirlScript Datathon 26'