Paper calibrates diffusion membership inference across two time axes
Original title: Difficulty Calibration Across Time for Membership Inference in Diffusion Models
Reference-Model Calibration uses an earlier checkpoint; Temporal Self-Calibration uses one model's neighboring diffusion timesteps.
The paper says standard membership inference on diffusion models flags low reconstruction loss, which confuses easy images with memorized ones. It calibrates difficulty across training time with Reference-Model Calibration, comparing a fine-tuned model to an earlier checkpoint, and across diffusion time with Temporal Self-Calibration, comparing a target timestep to neighboring timesteps from the same model. The authors report near-perfect auditing with RMC even without captions, wide gains over prior attacks, and generalization across x-, v-, and epsilon-prediction models. Code is released.
Unconfirmed: Near-perfect auditing and the margin over prior attacks are author-reported, not independently confirmed.
- RMC is for cases where an earlier checkpoint exists.
- TSC needs only a single model and its own trajectory.
- Authors say memorization is localized in diffusion time while difficulty persists.
- Methods are reported to generalize across x-, v-, and epsilon-prediction models.
From the original article
1 min read at arxiv.orgModern diffusion models are trained on massive datasets that often include copyrighted or private images. Membership inference attacks (MIAs) address this by testing whether a given image was in a model's training set. The standard approach scores an image by its reconstruction loss and flags low-loss images as members.
By Michael Joseph · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.