SpectralCanary: Embarrassingly Simple Low-Frequency Traces for T2I Dataset-Use Auditing
Auditing 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. Existing
Reading in Plain English mode — toggle below to switch
Auditing 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. Existing
This paper introduces a novel, simple, and effective method for auditing the use of datasets in text-to-image models, addressing a critical challenge in model transparency and intellectual property. For ML engineers, understanding such proactive auditing techniques is crucial for developing responsible AI systems and ensuring compliance in model development and deployment.