Diffusion watermarks get a spider-web pattern: top fixed-threshold accuracy in eight of nine transforms, while Tree-Ring still wins on clean images
Synopsis
Images from diffusion models need traceable provenance, yet existing watermarks spread energy uniformly by radius, so they weaken under directional distortions and their detection thresholds drift across transforms. SpiderMark writes a spider-web pattern of radial spokes and concentric polygons into the Fourier representation of the initial noise without touching pretrained weights, then a ResNet-18 verifier fusing PSNR and L1 similarity cues outputs a continuous confidence score. Across nine image transforms it reaches the highest fixed-threshold accuracy in eight, Tree-Ring leads on clean images, detection holds on unseen COCO prompts, and CLIP and FID show only small shifts in prompt alignment and image quality.
Interpretation
The watermark no longer spreads energy only by radius but carries directional information, and detection moves from a handcrafted threshold to a learned continuous confidence score. Earlier diffusion watermarks mostly used isotropic structures such as concentric rings, uniform grids or random masks, distributing energy uniformly by radius and leaving them vulnerable to motion blur and anisotropic resampling; detection relied on frequency correlation or distance thresholds whose optimum shifts markedly across JPEG compression, geometric warping and cropping. Evaluated on Stable Diffusion 2.1 under nine conditions, SpiderMark has the highest fixed-threshold accuracy in eight, while Tree-Ring leads on clean images at 0.9759 versus 0.9467; these are observed detection differences, not per-attack significance claims.
The change happens in the spectrum of the initial noise before generation starts, not in the model itself. The spider-web pattern consists of evenly spaced radial spokes joined by polygonal rings, normalized to a zero-mean, unit-variance carrier and blended with the original spectrum only inside a predefined frequency support, leaving coefficients outside it unchanged, so the generator needs no retraining. The verifier uses an ImageNet-pretrained ResNet-18 backbone on the magnitude spectrum from a 2D FFT, concatenated with PSNR and L1 scalars computed against the known pattern; training data are 500 watermarked and 500 clean images generated from identical prompts and noise seeds, split 85%/15%.
The learned part of detection can be measured on its own, and with the injected pattern held fixed the learned verifier clearly beats handcrafted correlation. Under a controlled retraining protocol with the spider-web pattern and support fixed, the full verifier reaches 72.86% macro accuracy, versus 67.88% for a ResNet-18 on spectral features alone and 47.33% for classical correlation. The PSNR and L1 cues together add 4.98 percentage points, a joint contribution rather than separate effects; the protocol uses a 150-image validation split, nine conditions and five seeded evaluation runs, and its absolute scores are not directly comparable with the main benchmark.
Perspective
Evaluation centers on images generated with Stable Diffusion 2.1 Base across nine conditions: clean, strong JPEG, messaging-app-style resampling and recompression, down-up sampling, Gaussian blur, random crop, occlusion, geometric warping and mixed augmentation, with verification relying on deterministic DDIM inversion using a diffusion model configured as at generation. Cross-dataset conclusions are limited to COCO prompts unseen in training, and cross-architecture conclusions to SDXL after retraining the lightweight verifier. The authors propose augmentation with compounded distortions as an extension.
The authors state the method is provenance verification rather than cryptographic authentication: it makes a binary presence decision and does not encode model versions or issuer identifiers, nor provide non-repudiation. In informed removal experiments, full suppression raises the watermarked-image false-negative rate to 97.47% but drops SSIM to 0.053 and raises FID from 83.14 to 261.68; a DC-preserving variant still detects 75 of 79 watermarked images under stronger noise. Forgery, key leakage, collusion across marked outputs and fully adaptive attacks are not covered by the current experiments.
