ShieldNet — Real-Time Financial Fraud Detection
Catching fraud before the transaction clears
Project Overview
A payment gateway serving 180,000 merchants was seeing a sharp rise in card-not-present fraud and synthetic identity attacks. Their legacy rule engine produced 40% false positives (blocking good customers) while still missing coordinated fraud rings. ShieldNet replaced it with a layered ML system that scores every transaction in under 80 ms.
The Challenges
- 1
The existing rule engine blocked 40% of legitimate transactions — any new system had to drastically cut false positives without opening the fraud door wider.
- 2
Fraud rings operate across hundreds of accounts; individual transaction models miss the network — graph-level signals were essential but expensive to compute.
- 3
New account fraud uses synthetic identities with clean histories — the model had to generalise to zero-history actors.
- 4
Regulatory requirements mandated that every block decision be explainable to both the merchant and the card issuer.
Our Approach
We built a three-layer scoring stack. Layer 1 is a sub-10 ms XGBoost model on device + transaction features that handles 95% of clear cases. Layer 2 is a graph neural network (GraphSAGE) trained on the historical transaction graph — it scores the neighbourhood risk of any account: if your new card shares a device fingerprint with 12 previously flagged accounts, the ring is visible. Layer 3 is a behavioural sequence model (LSTM) that spots velocity anomalies even on clean accounts. All three scores are combined via a meta-learner, and every decision comes with a SHAP explanation report for compliance.
Key Features & Metrics
Sub-80ms end-to-end scoring latency at 4,000 transactions per second peak load
Graph Neural Network surfaces coordinated fraud rings invisible to per-account models
SHAP-based explainability report auto-generated for every block decision
Adaptive thresholds per merchant category — jewellers get tighter thresholds than grocery stores
False positive rate reduced from 40% to 2.7% vs. the legacy rule engine
Model retraining pipeline runs nightly on the last 30 days of labelled outcomes
Results & Business Outcome
$2.1 M in fraudulent transactions blocked in Q1. False positives dropped from 40% to 2.7%, recovering an estimated $380 K per month in previously declined legitimate revenue. Chargeback rate fell from 1.8% to 0.3%, keeping the gateway out of Visa's high-risk monitoring programme.
Fraud is a cat-and-mouse game that never ends. The only sustainable edge is a system that learns faster than the attackers adapt.
Let's Build Something Intelligent Together
Tell us about your project. We'll respond within 24 hours with a custom plan and transparent pricing.