ADRD: Detecting diffusion-generated images via adversarial perturbation induced reconstruction discrepancy.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Diffusion models rapidly advance, raising concerns about misuse for deceptive visual content generation.
• Existing AI-generated image detection methods often rely on semantic features, which are becoming less effective as diffusion models improve.
• Current reconstruction-based detection methods treat reconstruction error statically, limiting robustness against generator variations and post-processing.

Purpose of the Study:

• To develop a robust framework for detecting AI-generated images from diffusion models.
• To address the limitations of existing semantic-based and static reconstruction-based detection methods.
• To propose a novel approach that models diffusion reconstruction as a dynamic process.

Main Methods:

• Propose Adversarial Diffusion Reconstruction Distance (ADRD), a detection framework.
• Model diffusion reconstruction as a dynamic response process, not a fixed descriptor.
• Probe reconstruction behavior by introducing latent space perturbations and measuring deviation responses.

Main Results:

• Real images show larger and more variable reconstruction responses to perturbations.
• Diffusion-generated images exhibit more stable reconstruction behavior under identical perturbations.
• Reconstruction sensitivity, rather than absolute error, proves to be a meaningful signal for detection.

Conclusions:

• ADRD offers a complementary perspective to existing reconstruction-based detectors by characterizing reconstruction sensitivity.
• The dynamic response of reconstruction under controlled perturbations is an effective signal for identifying diffusion-generated images.
• The proposed method enhances the robustness and accuracy of AI-generated image detection frameworks.