SpectralCanary uses low-frequency traces to audit T2I dataset use
Original title: SpectralCanary: Embarrassingly Simple Low-Frequency Traces for T2I Dataset-Use Auditing
With SD3.5 across five datasets it averages 92.4% per-generation TPR at 1% FPR and mean ΔFID of 0.05.
SpectralCanary blends one fixed canary image into every image in a collection with a smooth low-pass mask, leaving captions unchanged and skipping per-image optimization. A lightweight verifier looks for the transferred signal in a suspect text-to-image model's outputs. On Pokemon and three transformations it reports the highest average per-generation TPR at 1% FPR, 91.1%, and 94.8% model-level accuracy from 30 generations among methods tested. Matched-energy ablations find low-frequency placement transfers more reliably than mid- or high-frequency placement.
Unconfirmed: Robustness claims cover only the adaptive attacks the authors evaluated.
- Released-image fidelity is SSIM 0.989 and LPIPS 0.022.
- Domains tested include artistic, medical, satellite, and fashion data.
- The strongest evaluated removal attack needs more than 25 ΔFID points.
- The method needs neither auxiliary generation nor per-image optimization.
From the original article
1 min read at arxiv.orgAuditing whether a text-to-image (T2I) model was fine-tuned on a particular image collection is difficult: a model can absorb collection-level visual characteristics without reproducing training images. Proactive dataset-use auditing plants a shared signal in a dataset before release, then later checks whether that signal appears in a suspect model's generated outputs.
By Sy-Tuyen Ho · 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.