Tracing the Roots: Leveraging Temporal Dynamics in Diffusion Trajectories for Origin Attribution

  • 2024-12-19 18:51:28
  • Andreas Floros, Seyed-Mohsen Moosavi-Dezfooli, Pier Luigi Dragotti
  • 0

Abstract

Diffusion models have revolutionized image synthesis, garnering significantresearch interest in recent years. Diffusion is an iterative algorithm in whichsamples are generated step-by-step, starting from pure noise. This processintroduces the notion of diffusion trajectories, i.e., paths from the standardGaussian distribution to the target image distribution. In this context, westudy discriminative algorithms operating on these trajectories. Specifically,given a pre-trained diffusion model, we consider the problem of classifyingimages as part of the training dataset, generated by the model or originatingfrom an external source. Our approach demonstrates the presence of patternsacross steps that can be leveraged for classification. We also conduct ablationstudies, which reveal that using higher-order gradient features to characterizethe trajectories leads to significant performance gains and more robustalgorithms.

 

Quick Read (beta)

loading the full paper ...