IDNet-Fraud-Detection-System-An-Integrated-Approach-with-PyTorch-and-PostgreSQL

This repository is an end to end identity card fradud detection system that uses inference querying for generating insights. This project was completed as part of CSE 598- Data Intensive Systems for Machine Learning under Prof. Jia Zou

Fraud Detection CNN Pipeline

This project implements a CNN-based pipeline to detect visual fraud indicators in identity documents. It automates the process from data preparation to model training and deployment on Hugging Face Hub.

Project Overview

Addresses the challenge of identity document fraud by providing an automated detection system using Convolutional Neural Networks (CNNs). Aims for high accuracy (>95%) and low latency (<2 seconds per document).

(Note: The full project scope includes PostgresML integration, which is not covered by this specific code pipeline.)

Features

• Data Preparation: Extracts images from base64 encoded data and splits them into stratified train/test/out-of-sample sets.
• Hyperparameter Tuning: Uses Optuna to find optimal hyperparameters for the ResNet50 model.
• Model Training: Trains a ResNet50 model for binary classification (Fraud/Non-Fraud) with class weighting, learning rate scheduling, and early stopping.
• Evaluation: Generates classification reports, confusion matrices, and training history plots.
• Hugging Face Integration: Packages the trained model, configuration, and processor for easy upload and use via the Hugging Face Hub.

Installation

1. Clone the repository:

   git clone https://github.com/IrishMehta/IDNet-Fraud-Detection-System-An-Integrated-Approach-with-PyTorch-and-PostgreSQL.git
   cd "https://github.com/IrishMehta/IDNet-Fraud-Detection-System-An-Integrated-Approach-with-PyTorch-and-PostgreSQL.git"

2. Create and activate a virtual environment (recommended):

   python -m venv venv
   source venv/bin/activate  # On Windows use `venv\Scripts\activate`

3. Install dependencies:

   pip install torch torchvision pandas scikit-learn optuna matplotlib Pillow huggingface_hub transformers requests
   

4. Prepare Data: Place your idimage.csv, idmeta.csv, and idlabel.csv files inside a data_files directory (or update paths in config.json).

Usage

The pipeline consists of several steps, executed via Python scripts. Ensure config.json is correctly configured before running.

1. Prepare Dataset: Extracts images and creates train/test/OOS splits.

   python dataset_creation.py

   (Outputs images to extracted_images/ and splits to dataset/)

2. Hyperparameter Tuning: Runs Optuna to find the best hyperparameters.

   python hyperparameter_tuning.py

   (Outputs best parameters to hpo_output/best_hyperparams.json) (Note: Current training.py uses hardcoded hyperparameters; update it if you want to use tuned ones.)

3. Train Model: Trains the final model using parameters defined in the script or config.json.

   python training.py

   (Outputs model (best_model.pth), history, and plots to training_output/)

4. Upload to Hugging Face Hub: Packages the model and uploads it to the specified Hugging Face repository.

   python hf_upload.py

   (Creates fraud_detection_hf/ folder locally before upload. Requires Hugging Face Hub credentials set in config.json or environment variables.)

Configuration

Pipeline behavior is controlled by config.json. Key sections:

• data_prep: Paths for input CSVs, output directories for extracted images and dataset splits, split ratios.
• hpo: Paths for hyperparameter optimization, validation split ratio, number of classes/epochs/trials, output directory.
• training: Paths for data directories, validation split ratio, best model output path, output directory for training artifacts.
• huggingface: Paths for Hugging Face model packaging, local model weights, HF tokens, repository name, and an optional sample image for inference testing.

Note: Ensure file paths and Hugging Face credentials (HF_WRITE_TOKEN, HF_READ_TOKEN, HF_REPO_NAME) are correctly set either in the config or as environment variables.

Project Structure

.
├── config.json                 # Configuration file
├── dataset_creation.py         # Script for data preparation and splitting
├── hyperparameter_tuning.py    # Script for HPO using Optuna
├── training.py                 # Script for model training and evaluation
├── hf_upload.py                # Script for packaging and uploading to Hugging Face Hub
├── data_files/                 # Input CSV files (idimage.csv, idmeta.csv, idlabel.csv)
├── extracted_images/           # Output directory for decoded images
├── dataset/                    # Output directory for train/test/oos splits
│   ├── train/
│   ├── test/
│   └── out_of_sample_test/
├── hpo_output/                 # Output directory for HPO results
├── training_output/            # Output directory for training results (model, plots, history)
├── fraud_detection_hf/         # Staging directory for Hugging Face upload
└── requirements.txt            # Python dependencies

Integration with PostgreSQL