Graph Neural Network Fraud Detection

Overview

A Graph Convolutional Network (GCN) for fraud detection trained on a public dataset (80/20 split). The model was attacked with adversarial samples to test robustness, then fine-tuned on those samples to harden it against future attacks — improving overall accuracy in the process. Outcome: Produced a fraud detection model that is both more accurate and noticeably more robust to adversarial input.

Architecture & Pipeline

flowchart LR
    n0["
Transaction Graph
Public dataset
"] n1["
Train/Test Split
80 / 20
"] n2["
GCN Training
PyTorch
"] n3["
Adversarial Attack
Robustness probe
"] n4["
Adversarial Fine-tune
Hardened model
"] n5["
Production Classifier
Higher accuracy + robustness
"] n0 --> n1 n1 --> n2 n2 --> n3 n3 --> n4 n4 --> n5 classDef step0 fill:#f1f5f9,stroke:#64748b,color:#1e293b,stroke-width:2px,rx:10,ry:10; classDef step1 fill:#ecfeff,stroke:#06b6d4,color:#1e293b,stroke-width:2px,rx:10,ry:10; classDef step2 fill:#f0fdfa,stroke:#0d9488,color:#1e293b,stroke-width:2px,rx:10,ry:10; classDef step3 fill:#ecfdf5,stroke:#10b981,color:#1e293b,stroke-width:2px,rx:10,ry:10; classDef step4 fill:#fffbeb,stroke:#f59e0b,color:#1e293b,stroke-width:2px,rx:10,ry:10; class n0 step0; class n1 step1; class n2 step2; class n3 step2; class n4 step3; class n5 step4;

End-to-end flow derived from this project's scope and tech stack. Tap View Fullscreen for a larger view, or scroll horizontally on small screens.

Key Features

  • GCN architecture for graph-structured fraud data
  • Adversarial attack evaluation
  • Adversarial fine-tuning for hardened robustness
  • Tech Stack:** Python, PyTorch