Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Andreas Floros, Seyed-Mohsen Moosavi-Dezfooli, Pier Luigi Dragotti

TL;DR

This work tackles provenance for diffusion-based image synthesis by leveraging temporal dynamics along complete diffusion trajectories to infer whether a sample is part of training, a novel generation, or external. It shows that traditional membership inference approaches fail under realistic conditions and that rich trajectory features, including gradient signals, yield robust provenance cues, challenging the Goldilocks zone notion. The authors extend the framework to origin attribution with white-box model attribution for diffusion and demonstrate data-extraction-style risks, arguing for a unified, provenance-centric approach tailored to modern generative systems. The results highlight practical implications for accountability and privacy, while outlining methodological directions and threat-model considerations for future work.

Abstract

Diffusion models have transformed image synthesis through iterative denoising, by defining trajectories from noise to coherent data. While their capabilities are widely celebrated, a critical challenge remains unaddressed: ensuring responsible use by verifying whether an image originates from a model's training set, its novel generations or external sources. We introduce a framework that analyzes diffusion trajectories for this purpose. Specifically, we demonstrate that temporal dynamics across the entire trajectory allow for more robust classification and challenge the widely-adopted "Goldilocks zone" conjecture, which posits that membership inference is effective only within narrow denoising stages. More fundamentally, we expose critical flaws in current membership inference practices by showing that representative methods fail under distribution shifts or when model-generated data is present. For model attribution, we demonstrate a first white-box approach directly applicable to diffusion. Ultimately, we propose the unification of data provenance into a single, cohesive framework tailored to modern generative systems.

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

TL;DR

This work tackles provenance for diffusion-based image synthesis by leveraging temporal dynamics along complete diffusion trajectories to infer whether a sample is part of training, a novel generation, or external. It shows that traditional membership inference approaches fail under realistic conditions and that rich trajectory features, including gradient signals, yield robust provenance cues, challenging the Goldilocks zone notion. The authors extend the framework to origin attribution with white-box model attribution for diffusion and demonstrate data-extraction-style risks, arguing for a unified, provenance-centric approach tailored to modern generative systems. The results highlight practical implications for accountability and privacy, while outlining methodological directions and threat-model considerations for future work.

Abstract

Diffusion models have transformed image synthesis through iterative denoising, by defining trajectories from noise to coherent data. While their capabilities are widely celebrated, a critical challenge remains unaddressed: ensuring responsible use by verifying whether an image originates from a model's training set, its novel generations or external sources. We introduce a framework that analyzes diffusion trajectories for this purpose. Specifically, we demonstrate that temporal dynamics across the entire trajectory allow for more robust classification and challenge the widely-adopted "Goldilocks zone" conjecture, which posits that membership inference is effective only within narrow denoising stages. More fundamentally, we expose critical flaws in current membership inference practices by showing that representative methods fail under distribution shifts or when model-generated data is present. For model attribution, we demonstrate a first white-box approach directly applicable to diffusion. Ultimately, we propose the unification of data provenance into a single, cohesive framework tailored to modern generative systems.

Paper Structure

This paper contains 25 sections, 8 equations, 10 figures, 9 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Background
  3. Diffusion-based generative models
  4. Continuous-time formulation
  5. Denoising diffusion probabilistic models
  6. Data provenance
  7. Membership inference
  8. Model attribution
  9. Threat model
  10. Do membership inference attacks work?
  11. Representative attacks cannot audit synthetic data
  12. Blind baselines may beat most membership inference attacks
  13. Rethinking the Goldilocks zone conjecture
  14. Toward origin attribution
  15. Stronger membership inference attacks
  16. ...and 10 more sections

Figures (10)

  • Figure 1: Member, model-generated and external data loss (averages with solid lines) as a function of $t$ for CelebA-HQ DDPM.
  • Figure 2: Visualization of CelebA-HQ DDPM features.
  • Figure 3: Parameters from $\boldsymbol{W}$ in \ref{['eq:linclass']} corresponding to different features against $t$ for our MIAs.
  • Figure 4: MIA performance as a function of the data budget (% of member set) on CIFAR-10.
  • Figure 5: MIA TPRs @ 0.1% FPR as the data budget varies. See Appendix \ref{['appendix:setup']} for details.
  • ...and 5 more figures

Theorems & Definitions (1)

  • Conjecture 1