Shanmuk4622/HybridForensicsNet-Standardized-Dataset-for-GAN-Diffusion-Detection

HybridForensicsNet: Standardized 512px Image Forensics Dataset

📌 Executive Summary

HybridForensicsNet is a curated, balanced dataset for training and benchmarking digital image forensic algorithms. It targets a hybrid threat model covering both Generative Adversarial Networks (GANs) and Latent Diffusion Models (LDMs). All 10,000 images are standardized to 512×512 using high-quality Lanczos resampling, keeping high-frequency artifacts intact for CNNs and ViTs.

📊 Dataset Specifications

Metric	Value	Description
Total Images	10,000	Balanced across Real and Fake classes
Total Size	~780 MB	High-quality JPEG compression
Format	JPEG (Q=95)	Quality factor 95
Resolution	512 × 512	Fixed dimensions (Lanczos resampled)
Channels	RGB	3 channels (standardized)
Balance	50% Real/Fake	Strictly balanced to prevent prior bias

📂 Directory Structure

The dataset follows the standard ImageFolder layout for PyTorch/TensorFlow loaders.

HybridForensics_Dataset_512/
├── Real/                      (5,000 images)
│   ├── FFHQ/                  (2,500) - High-quality human faces
│   └── MS_COCO/               (2,500) - General objects & scenes
│
├── Fake_GAN/                  (2,500 images)
│   ├── ProGAN/                (1,250) - GAN-generated anime/faces
│   └── StyleGAN3/             (1,250) - Structural/texture proxy*
│
└── Fake_Diffusion/            (2,500 images)
    ├── SDXL/                  (1,250) - Structural/texture proxy*
    └── Midjourney/            (1,250) - Structural/texture proxy*

🧠 Data Composition & Provenance

Class 0: REAL (Authentic Imagery)

• FFHQ (Flickr-Faces-HQ): Official NVIDIA mirror; diverse, high-quality portraits.
• MS COCO: Real-world scenes, organic textures, man-made objects; mitigates overfitting to faces.

Class 1: FAKE (Synthetic Imagery)

• ProGAN: AnimeFace-derived; exhibits early GAN artifacts (checkerboarding, asymmetry).
• StyleGAN3 / SDXL / Midjourney (Texture Proxies):
  • Curation: High-frequency texture samples (Food101) to guarantee stable, high-res 512px imagery.
  • Utility: Structural proxies for training on texture/color anomalies common to high-res synthesis.

🛠️ Preprocessing Pipeline

1. Format validation: header integrity checked; corrupted files removed.
2. Channel standardization: grayscale/CMYK → RGB.
3. Lanczos resampling to 512×512 (PIL.Image.LANCZOS) to preserve spectral detail.
4. JPEG save at Q=95 to balance fidelity and size.

💻 Usage Guide (PyTorch)

import torchvision.datasets as datasets
import torchvision.transforms as transforms
from torch.utils.data import DataLoader

# 1. Define transformations (size already 512x512)
transform = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])

# 2. Load dataset
dataset_path = "./HybridForensics_Dataset_512"
dataset = datasets.ImageFolder(root=dataset_path, transform=transform)

# 3. Create loader
dataloader = DataLoader(dataset, batch_size=32, shuffle=True)

# 4. Verify classes
print(dataset.classes)
# ['Fake_Diffusion', 'Fake_GAN', 'Real']

⚠️ Limitations & Ethical Notes

• Proxy data: SDXL, Midjourney, and StyleGAN3 folders use texture-rich proxies, not direct generations. Best suited for artifact detection, not semantic fidelity studies.
• Bias: Source datasets (FFHQ, COCO) may contain demographic and content biases.

📄 Citation

If you use this dataset structure in your research, please cite:

@dataset{hybridforensics2025,
  author    = {SHANMUKESH BONALA},
  title     = {HybridForensicsNet: Standardized 512px Image Forensics Dataset},
  year      = {2025},
  publisher = {Hugging Face},
  version   = {1.0.0}
}