IC-1195Model substitution adversarial attack reduces TreeRing AUROC to 0.14 at ε=2/255 and StegaStamp AUROC to 0.492 at ε=12/255

Mehrdad Saberi, Vinu Sankar Sadasivan, Keivan Rezaei, Aounon Kumar, Atoosa Chegini, Wenxiao Wang, Soheil Feizi

SourceRobustness of AI-Image Detectors: Fundamental Limits and Practical Attacks

For high-perturbation watermarks where diffusion purification is ineffective, the paper trains a substitute ResNet-18 classifier to distinguish watermarked from non-watermarked images, then applies PGD adversarial perturbations using the substitute. The resulting adversarial images transfer to the original black-box watermark detector. TreeRing is highly vulnerable: a perturbation budget as low as ε=2/255 reduces its AUROC to 0.14. StegaStamp is more resilient, requiring ε=12/255 to reach AUROC 0.492, though this level of perturbation may leave perceptible artifacts.

Evidence
correlational
Key metric
TreeRing AUROC 0.14 at ε=2/255; StegaStamp AUROC 0.492 at ε=12/255 (Table 4, base column)
Caveat
Transferability depends on substitute classifier architecture; a more suitable configuration may increase attack success. StegaStamp requires ε=12/255 which may leave perceptible artifacts on images.
Model
TreeRing, StegaStamp
Concepts
Failure mode
Datasets
ImageNet-1k / ImageNet / ImageNet-1k-val / ImageNet-Val [eval], LAION-Captions [eval]
Methods
PGD (Projected Gradient Descent) [primary]
Related findings
IC-1196
Extraction
automatic-extraction