MD-SynReal12K: A Multi-Domain Benchmark Dataset for AI-Generated and Authentic Image Forensics

Description

Dataset Overview: MD-SynReal12K is a rigorously curated, balanced multi-domain benchmark dataset comprising 12,000 high-resolution images designed to evaluate, benchmark, and enhance deep learning forensic models for detecting AI-generated (synthetic) media. With the recent proliferation of diffusion models and generative adversarial networks (GANs), synthetic imagery poses severe challenges to digital security and biometric integrity. This dataset provides a systematic benchmark spanning multiple visual categories to study both spatial-domain textures and frequency-domain (Fourier spectral) artifacts. Composition and Categories: The dataset contains a strictly balanced distribution of 12,000 images divided equally into two primary forensic classes: - Authentic / Real (6,000 images): Genuine photographic captures acquired from verified real-world sources. - Synthetic / AI-Generated (6,000 images): High-fidelity synthetic images generated via state-of-the-art generative paradigms. To prevent domain bias and evaluate cross-semantic generalization, both Real and Synthetic classes are stratified across three diverse visual domains: 1. Human Faces (Portraiture, facial biometrics, facial textures). 2. Fauna and Wildlife (Animals, birds, biological organisms). 3. Structured Objects (Vehicles, furniture, tools, architectural entities). Structure and Specifications: - Total Images: 12,000 instances. - Image Resolution: Standardized high-resolution (512x512 pixels). - Color Space: RGB (3 channels). - File Formats: Uncompressed/lossless PNG and high-quality JPEG. - Directory Hierarchy: Organized into stratified Train, Validation, and Test splits with dedicated sub-folders for each category and class label. Intended Applications: This benchmark is tailored for training and validating computer vision architectures, including Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), Swin Transformers, and Hybrid Spatial–Frequency dual-stream networks. Researchers can utilize MD-SynReal12K for benchmarking spatial anomaly detection, 2D Discrete Fast Fourier Transform (2D FFT) spectral forensics, and domain generalization studies.

Institutions

• University of Halabja

  Sulaymaniyah, Halabja

Categories

Licence