DeepFeatureX Net: Deep Features eXtractors based Network for discriminating synthetic from real images

Abstract

Deepfakes, synthetic images generated by deep learning algorithms, represent one of the biggest challenges in the field of Digital Forensics. The scientific community is working to develop approaches that can discriminate the origin of digital images (real or AI-generated). However, these methodologies face the challenge of generalization, that is, the ability to discern the nature of an image even if it is generated by an architecture not seen during training. This usually leads to a drop in performance. In this context, we propose a novel approach based on three blocks called Base Models, each of which is responsible for extracting the discriminative features of a specific image class (Diffusion Model-generated, GAN-generated, or real) as it is trained by exploiting deliberately unbalanced datasets. The features extracted from each block are then concatenated and processed to discriminate the origin of the input image. Experimental results showed that this approach not only demonstrates good robust capabilities to JPEG compression but also outperforms state-of-the-art methods in several generalization tests.

For more information, here the link to the paper:

• Springer: link
• Arxiv: link

Project Structure

This section outlines the recommended folder structure for the project. The working directory is where all the images for training and testing the models, as well as the model weights, are stored. The project is organized as follows:

working_dir/
├── datasets/                        # Training and testing datasets
├── testing_robustness/              # Data for testing model robustness
├── testing_generalization/          # Data for testing model generalization
├── guidance.csv                     # CSV file for data processing guidance
└── models/                          # Models and related files for the unbalancing approach

This structure is recommended but not mandatory. You can specify alternative paths as needed. To use the suggested structure:

1. Create a file named wd.py inside src/dfx/
2. Add the following line to wd.py:

working_dir = '<your_working_directory_path>'

3. Replace <your_working_directory_path> with your actual working directory path.

Installation

1. Clone the repository:

   git clone git@github.com:opontorno/block-based_deepfake-detection.git
   cd <repository_name>
   

2. Install the dfx package and dependencies:

   pip install -e .
   pip install -r requirements.txt
   

Train the models

Data preparation

1. Organize your data following this structure:

  Datasets/
  ├── gan_generated/
  │   ├── AttGAN/
  │   ├── BigGAN/
  │   ├── CycleGAN/
  │   ├── GauGAN/
  │   └── .../
  ├── dm_generated/
  │   ├── DALL-E 2/
  │   ├── DALL-E MINI/
  │   ├── Glide/
  │   ├── Latent Diffusion/
  │   └── .../
  └── real/
      └── real/

2. Create a guidance CSV file:

python scripts/prep/make_guidance.csv --datasets_dir <your_data_dir> --saving_dir <your_save_path> --guidance_dir <your_guidance_path>

Train Base Models

Train the three base models:

python scripts/training/training_base-model.py --datasets_dir <your_data_dir> --main_class <main> --saving_path <your_save_path>

Replace <main> with one of: 'dm_generated', 'gan_generated', or 'real'.

Train Complete Model

Train the complete model:

python scripts/training/training_complete-model.py --datasets_dir <your_data_dir> --main_class <main> --saving_path <your_save_path>

Inference

Test the complete model specifying the backbone type and, if any, the model_path:

python scripts/testing/testing_complete-models.py --backbone <backbone_type> --model_path <your_model_path> --test_raw True --datasets_dir <your_data_dir> --main_class <main> --saving_path <your_save_path>

Dataset

The dataset comprises a total of $72,334$ images, distributed as shown in the Table. The image sizes vary considerably, ranging from 216x216 pixels up to 1024x1024 pixels, thus offering a wide spectrum of resolutions for analysis. For each generative architecture, special attention was paid to the internal balancing of the corresponding subset of images. This balancing was pursued in terms of both semantic content and size in order to minimise potential bias and ensure a fair representation of the different types of visual input. All images are in PNG format.

Nature	Architecture	Type	# Images	Total	Different Sizes
GAN	AttGAN	FO	6005		256 × 256
	BigGAN	O	2600		256 × 256
	CycleGAN	FO	1047		256 × 256; 512 × 512
	GauGAN	O	4000		256 × 256; 512 × 512
	GDWCT	O	3367	37.572	216 × 216
	ProGAN	O	1000		256 × 256; 512 × 512
	StarGAN	F	6848		256 × 256
	StyleGAN	O	4705		256 × 256; 512 × 512
	StyleGAN2	FO	7000		256 × 256; 512 × 512; 1024 × 1024
	StyleGAN3	F	1000		256 × 256; 512 × 512; 1024 × 1024
DM	DALL-E 2	FO	3421		512 × 512; 1024 × 1024
	DALL-E MINI	O	1000		256 × 256
	Glide	O	2000	15.421	256 × 256; 512 × 512
	Latent Diffusion	FO	4000		256 × 256; 512 × 512
	Stable Diffusion	FO	5000		256 × 256; 512 × 512

Nature	Sources	Type	# Images	Total	Different Sizes
REAL	CelebA	F	5135		178 × 218
	FFHQ	F	4981	19341	1024 × 1024
	Others	O	9225		256 × 256; 512 × 512; 1024 × 1024

The dataset is available at the following link: link [DATASET AVAILABLE AS SOON AS THE PAPER IS PUBLISHED]

Citation

@inproceedings{pontorno2025deepfeaturex,
  title={DeepFeatureX Net: Deep Features eXtractors based Network for discriminating synthetic from real images},
  author={Pontorno, Orazio and Guarnera, Luca and Battiato, Sebastiano},
  booktitle={International Conference on Pattern Recognition},
  pages={177--193},
  year={2025},
  organization={Springer}
}