How sohu9/zero-day-fraud-detection works
This repository is a zero-day payment-fraud detection application: it learns normal transaction behavior with an autoencoder, flags anomalous transactions, and presents risk decisions with explanations. The main workflow runs from transaction submission through the Flask application’s preprocessing and inference to a decision, persistence, and dashboard/API access. Separate training and evaluation workflows prepare the model and measure its performance. The README also describes SHAP/LIME interpretability and broader behavioral features; sampled code shows feature-based explanation text, but does not establish SHAP/LIME execution. Some application wiring is omitted from the excerpts.
Main entry points: app.py Flask transaction service · evaluate.py Offline evaluation · train.py Model training · data_pipeline.py Feature preparation
Components 17
Data and modeling
- Feature preparation
data_pipeline.py - Fitted preprocessor
data_pipeline.py - Model training
train.py - Autoencoder builder
model_builder.py - Trained autoencoder
train.py - Anomaly threshold
train.py
Fraud decision service
- Flask transaction service
app.py - Live preprocessing
app.py - Anomaly scoring
app.py - Decision explanation
app.py - Web dashboard
templates/index.html
Transaction records
- Transaction store
database.py
Model assessment
- Offline evaluation
evaluate.py - Evaluation results
evaluate.py
Other components
- Transaction submitter
- Training transactions
- Test transactions
Connections 17
- Training transactions → Feature preparation: loads and transforms
data_pipeline.py - Model training → Feature preparation: runs preprocessing
train.py - Feature preparation → Fitted preprocessor: saves fitted state
data_pipeline.py - Model training → Autoencoder builder: builds model
train.py - Model training → Trained autoencoder: trains and saves
train.py - Model training → Anomaly threshold: calibrates and saves
train.py - Transaction submitter → Flask transaction service: submits transaction
README.md - Flask transaction service → Live preprocessing: prepares input
app.py - Flask transaction service → Anomaly scoring: analyzes transaction
app.py - Flask transaction service → Decision explanation: builds explanation
app.py - Flask transaction service → Transaction store: queries records
app.py - Flask transaction service → Web dashboard: serves web interface
README.md - Test transactions → Offline evaluation: evaluates against
evaluate.py - Offline evaluation → Feature preparation: uses time features
evaluate.py - Offline evaluation → Trained autoencoder: loads for scoring
evaluate.py - Offline evaluation → Anomaly threshold: loads threshold
evaluate.py - Offline evaluation → Evaluation results: writes results
evaluate.py