Signature Forgery Detection Using GANs

This project implements an end-to-end system for signature forgery detection by combining Generative Adversarial Networks (GANs) for synthetic signature generation and a Convolutional Neural Network (CNN) for classification.

---

Overview

Signature verification is a key task in biometric authentication and fraud prevention. Collecting large datasets of genuine and forged signatures is often challenging. This project addresses this by:

1. Training a GAN to learn the visual distribution of genuine signatures.
2. Generating synthetic (fake) signature samples to augment limited forged data.
3. Training a CNN classifier to distinguish between genuine and forged signatures.

The system demonstrates the potential of combining generative and discriminative models for enhanced forgery detection.

The system demonstrates the potential of combining generative and discriminative models for enhanced forgery detection.

A full detailed report documenting the methodology, experiments, and results is included in this repository.

---

Features

• GAN Training: Generates realistic synthetic signatures using a Wasserstein GAN with Gradient Penalty (WGAN-GP).
• CNN Classification: Trains a convolutional neural network to classify genuine vs forged signatures.
• Evaluation on GAN Samples: Assesses the realism of generated signatures using the trained CNN.

---

Dataset

• 2,500 genuine signature images and 2,500 forged signature images.
• Split: 80% training / 20% validation.
• Preprocessed to grayscale, resized to 128x128 for GAN, 64x64 for CNN, and normalized appropriately.

---

Project Structure

src/
├── app.py                     # Main script to run different stages
├── data_prep.py               # Preprocessing pipeline for the dataset
├── train_gan.py               # GAN training script
├── generate_fakes.py          # Generate synthetic signatures from trained GAN
├── train_signature_classifier.py  # CNN classifier training
└── evaluate_gan_with_cnn.py   # Evaluate generated images with trained CNN
data/
└── generated_fakes/           # Stores generated fake signatures

---

Usage

Run the main pipeline using:

python src/app.py

In app.py, each step can be uncommented depending on which stage you want to execute:

• run_data_prep() – Preprocesses and organizes the dataset.
• run_gan_training() – Trains the GAN on genuine signatures.
• run_generate_fakes() – Generates synthetic forged signatures.
• run_cnn_training() – Trains the CNN classifier.
• run_gan_evaluation() – Evaluates GAN-generated images with the trained CNN.

---

Results

• CNN Accuracy: ~75% on the test dataset.
• GAN Evaluation: ~55% of generated samples predicted as forged by the CNN, indicating the GAN produces partially realistic features.
• Training Observations: Generator gradually improves during training, producing visually coherent signatures, though some instability occurs in later epochs.