Reproduce “Manifold Induced Biases for Zero-shot and Few-shot Detection of Generated Images” on OpenResearch

About this paper

Jonathan Brokman, Amit Giloni, Omer Hofman, Roman Vainshtein, Hisashi Kojima, Guy Gilboa

Researchers at Fujitsu Research of Europe and Technion - Israel Institute of Technology developed a method for detecting AI-generated images that leverages manifold-induced biases from pre-trained diffusion models. This approach quantifies geometric properties of the implicit probability manifold to distinguish real from synthetic images, achieving an average AUC of 0.835 for zero-shot detection across 20 diverse generative models, significantly outperforming prior state-of-the-art methods.

Key insights

• Generated images inherently exhibit quantifiable biases, characterized by high positive curvature and low gradient magnitude on the learned log-probability manifold of diffusion models.
• These manifold-induced biases can be effectively approximated using the score function (noise predictions) of pre-trained diffusion models, even when the exact probability distribution is implicit.
• Mapping image features and score function approximations into a common embedding space (e.g., CLIP) facilitates robust generalization of these bias-driven criteria across various generative techniques.

Summary by alphaXiv.