Revisiting DIRE: towards universal AI-generated image detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Digital Forensics

Background:

• Generative models like GANs and diffusion models have advanced rapidly, improving image synthesis quality and accessibility.
• This advancement raises significant concerns regarding the credibility and authenticity of digital content.
• Existing methods struggle to reliably detect sophisticated synthetic images.

Purpose of the Study:

• To propose a novel method for detecting synthetic images generated by various generative models.
• To establish a universal approach for identifying image manipulation and deepfakes.
• To improve the accuracy and reliability of synthetic image detection.

Main Methods:

• Developed Universal Reconstruction Residual Analysis (UR²EA) utilizing reconstruction errors from pre-trained diffusion models.
• Introduced a Multi-scale Channel and Window Attention (MCWA) module for extracting fine-grained features from residual maps.
• Constructed the UniversalForensics dataset with diverse synthetic images from 30 different generative models.

Main Results:

• UR²EA effectively differentiates between real, GAN-generated, and diffusion-generated images based on reconstruction error patterns.
• GAN-generated images exhibit lower reconstruction quality, while diffusion-generated images show higher fidelity compared to real images.
• The proposed method achieved state-of-the-art results, improving average accuracy by 3.3% and precision by 1.6% over existing baselines.

Conclusions:

• UR²EA provides a robust and universal prior for detecting synthetic images across different generation methods.
• The MCWA module enhances feature extraction from residual maps, capturing crucial local and global details.
• The UniversalForensics dataset facilitates further research and development in synthetic image detection.