SpectralCanary
An active watermarking method for auditing the usage of text-to-image (T2I) datasets, embedding canary images with low-frequency signals through low-pass filtering to achieve dataset provenance with high detection rates and minimal visual impact.
Timeline (last 90 days)
SpectralCanary is a proactive dataset auditing method that blends a fixed canary image into every dataset image using a smooth low-pass mask, requiring no additional generation or optimization process.
On the Pokemon dataset, across three real data transformations and two architecture families, SpectralCanary achieves the highest average TPR@1% FPR of 91.1% per generation.
On the Pokemon dataset, with 30 generations, SpectralCanary achieves a model-level accuracy (fixed threshold) of 94.8%.
On five datasets spanning the art, medical, satellite, and fashion domains, using the SD3.5 model, SpectralCanary achieves an average TPR@1% FPR of 92.4%.
On five datasets using the SD3.5 model, SpectralCanary preserves the visual quality of released images (SSIM 0.989, LPIPS 0.022) with a near-zero FID change (average ΔFID 0.05).
Matched-energy ablation experiments show that on the Pokemon and ROCOv2 datasets, low-frequency-placed signals are detected more reliably after fine-tuning than mid- and high-frequency-placed signals.
The most effective adaptive attack that removes the SpectralCanary signal requires aggressive image modifications and causes an increase in ΔFID of more than 25 points.
SpectralCanary processes each image with a low-pass filter (e.g., Gaussian blur), then superimposes the canary image by weight, making the blend nearly imperceptible visually.
All Facts (8)
SpectralCanary is a proactive dataset auditing method that blends a fixed canary image into every dataset image using a smooth low-pass mask, requiring no additional generation or optimization process.
50%UnverifiedOn the Pokemon dataset, across three real data transformations and two architecture families, SpectralCanary achieves the highest average TPR@1% FPR of 91.1% per generation.
50%UnverifiedOn the Pokemon dataset, with 30 generations, SpectralCanary achieves a model-level accuracy (fixed threshold) of 94.8%.
50%UnverifiedOn five datasets spanning the art, medical, satellite, and fashion domains, using the SD3.5 model, SpectralCanary achieves an average TPR@1% FPR of 92.4%.
50%UnverifiedOn five datasets using the SD3.5 model, SpectralCanary preserves the visual quality of released images (SSIM 0.989, LPIPS 0.022) with a near-zero FID change (average ΔFID 0.05).
50%UnverifiedMatched-energy ablation experiments show that on the Pokemon and ROCOv2 datasets, low-frequency-placed signals are detected more reliably after fine-tuning than mid- and high-frequency-placed signals.
50%UnverifiedThe most effective adaptive attack that removes the SpectralCanary signal requires aggressive image modifications and causes an increase in ΔFID of more than 25 points.
50%UnverifiedSpectralCanary processes each image with a low-pass filter (e.g., Gaussian blur), then superimposes the canary image by weight, making the blend nearly imperceptible visually.
50%