Presentation Information
[N-1-37]Investigation of the Origin of Fine-Grained Features Contributing to Diffusion-Generated Image Detection
◎Takumi Owada1, Kenya Jin'no1 (1. Tokyo City University)
Keywords:
Diffusion Models,DDPM,Fake Image Detection,CNN
As diffusion-model image generation advances, detecting generated images is increasingly important. Conventional methods apply heavy neural networks to entire high-resolution images, whereas our previous work showed that partitioning an image into small local regions enables lightweight, accurate detection, with regions of low image brightness entropy yielding further gains. This suggests that the features aiding detection are not only the subject's global structure but fine-grained elements distributed across the whole image.The DDPM studied here generates images by gradually adding noise and learning to remove it. Training typically minimizes the simplified objective L_simple to predict the noise ε, an approximation that simplifies the per-timestep weighting of the original variational lower bound (VLB). Since fine-grained, high-frequency components are most strongly affected by the early, low-noise stage, we examine how the objective's timestep weighting affects the generated images' properties and detection recall. We found that suppressing the early-stage weighting increases the high-frequency components of generated images and lowers the detector's recall.