A shared image trace helps reveal when models learn a collection's style

If a model learns an image collection’s visual style without reproducing its pictures, how can the collection’s owner find evidence that the model used it?

Checks for copied training images may miss that kind of learning, so the question is whether a shared signal can be detected in later generations.

SpectralCanary adds a shared low-frequency trace to images before release, then uses a verifier to check for it in a model’s generations.

The researchers test whether the trace remains detectable when training data or model family changes. On Pokemon, they compare eight settings across two model families, using each method’s own protection and scoring procedure.

At a threshold that falsely flags at most 1% of clean-model generations, SpectralCanary detects 91.1% of protected-model generations on average, versus 73.2% for EnTruth. This mean supports detection across the tested settings, though no method leads in every one.

The audit provides statistical evidence under the tested conditions, not proof that a collection was used.

The evaluation covers two model families, five datasets, one main-study canary, and three removal attacks; it does not establish transfer to every model or to mixtures with larger unprotected datasets.