Data Attribution for Diffusion Models: Timestep-induced Bias in Influence Estimation
Abstract: Data attribution methods trace model behavior back to its training dataset, offering an effective approach to better understand ``black-box'' neural networks. While prior research established quantifiable links between model output and training data in diverse settings, interpreting diffusion model outputs in relation to training samples remains underexplored. In particular, diffusion models operate over a sequence of timesteps instead of instantaneous input-output relationships in previous contexts, posing a significant challenge to extend existing frameworks to diffusion models directly. Notably, we present Diffusion-TracIn that incorporates this temporal dynamics and observe that samples' loss gradient norms are highly dependent on timestep. This trend leads to a prominent bias in influence estimation, and is particularly severe for samples trained on large-norm-inducing timesteps, causing them to be generally influential. To mitigate this bias, we introduce Diffusion-ReTrac as a re-normalized adaptation that retrieves training samples targeted to the test sample of interest, enabling a localized measurement of influence and considerably more intuitive visualization. We demonstrate the efficacy of our approach through various evaluation metrics and auxiliary tasks, outperforming in terms of specificity of attribution by over $60\%$.
Submission Length: Regular submission (no more than 12 pages of main content)
Changes Since Last Submission: All feedback from reviewers and the Action Editor is addressed & incorporated into the main text or appendix.
**Main text**
- Added results on additional datasets in Targeted Attribution (6.2)
- Discussed trade-offs of ReTrac, and potential works mentioned by reviewers (Conclusion)
- Fixed typos & clarified questions raised
**Appendix**
- Additional norm distributions (Appendix A.2)
- Added baseline on normalization: Different levels of normalization intensity (Appendix B)
- Added baseline: Influence Functions (Appendix C)
- Additional analysis: Accuracy & Efficiency of timestep sampling (Appendix E)
Code: https://github.com/txie1/diffusion-ReTrac
Supplementary Material: zip
Assigned Action Editor: ~Changyou_Chen1
Submission Number: 2161
Loading