Composite Data Augmentations for Synthetic Image Detection Against Real-World Perturbations

Official implementation of "Composite Data Augmentations for Synthetic Image Detection Against Real-World Perturbations"
Accepted at EUSIPCO 2025 (33rd European Signal Processing Conference)

Overview

This repository implements a robust deep learning framework for detecting GAN-generated synthetic images with robustness to real-world perturbations. The approach combines multiple data augmentation techniques with a dual-criteria training pipeline to enhance model robustness against common image distortions encountered in practice. It also explores the utilization of a genetic algorithm for finding the optimal augmentaions to use during training.

Key Features

• Composite Data Augmentations: Multiple augmentation strategies including JPEG compression, Gaussian blur, noise, sharpening, color transformations, and more
• Dual-Criteria Training Pipeline: Novel training approach combining two losses
• Genetic Algorithm Optimization: Automated augmentation strategy selection using PyGAD
• Comprehensive Evaluation: Testing across 13 different synthetic image datasets
• ResNet-50 Backbone: Pre-trained on ImageNet with custom binary classification head

Installation

1. Clone the repository

2. Create and activate a virtual environment:

   # Create virtual environment
   python -m venv venv
   
   # Activate virtual environment
   # On Windows:
   venv\Scripts\activate
   
   # On macOS/Linux:
   source venv/bin/activate

3. Install dependencies:

   pip install -r requirements.txt

4. Set up your dataset directory structure:

   dataset/
   ├── train/
   │   ├── class1/
   │   └── class2/
   ├── val/
   │   ├── class1/
   │   └── class2/
   └── test/
       ├── progan/
       ├── stylegan/
       ├── biggan/
       └── ...
   

Dataset Structure

Organize your datasets in the following structure:

dataset/
├── train/
│   ├── class1/
|   |   ├──0_real
|   |   ├──1_fake
│   ├── class2/
|   |   ├──0_real
|   |   ├──1_fake
├── val/
│   ├── class1/
|   |   ├──0_real
|   |   ├──1_fake
│   ├── class2/
|   |   ├──0_real
|   |   ├──1_fake
└── test/
    ├── progan/
    |   ├──0_real
    |   ├──1_fake
    ├── stylegan/
    ├── biggan/
    ├── cyclegan/
    ├── stargan/
    ├── gaugan/
    ├── crn/
    ├── imle/
    ├── seeingdark/
    ├── san/
    ├── deepfake/
    ├── stylegan2/
    └── whichfaceisreal/

Usage

Basic Training

python train.py --name experiment_name --lr 0.001 --optim adam --num_epochs 25

Dual-Criteria Training

python double_criteria_pipeline.py --name dual_experiment --lr 0.001 --optim adam

Genetic Algorithm Optimization

python sid_ga.py --name ga_experiment --num_generations 10 --solutions_per_population 8

Evaluation

python evaluate.py --model_path best_model --runner local  # For local testing
python evaluate.py --model_path best_model --runner hpc    # For full evaluation on HPC environments

Results

The model outputs comprehensive evaluation metrics:

• Accuracy: Overall classification accuracy
• Average Precision (AP): Area under precision-recall curve
• Real Accuracy: Accuracy on real images
• Fake Accuracy: Accuracy on synthetic images

Results are automatically logged to TensorBoard and saved as formatted tables.