TL;DR
Digital watermarks are widely viewed as the frontline defense for regulating AI-generated content (AIGC), enabling copyright tracking and attribution. This paper introduces WMaGi, a unified, black-box framework that breaks these protections by both removing watermarks from protected images and forging specific user watermarks onto arbitrary images—all without requiring access to clean training pairs or the underlying watermarking models, and running up to 11,000× faster than pure diffusion-based attacks.
What Problem Does This Paper Address?
As generative models (e.g., Stable Diffusion, Midjourney) become ubiquitous, service providers embed invisible digital watermarks into outputs to enforce terms of service, track copyright, and attribute harmful generations (e.g., deepfakes or misinformation) to specific users. A common industry assumption—reiterated in recent industry consensus workshops—has been that invisible watermarks are secure against attackers who lack access to the secret key or detection model, and that watermark forging is especially impractical.
This paper challenges that assumption by examining two critical black-box attack vectors:
- Watermark Removal: Stripping embedded ownership tags so generated content can be commercialized or abused without attribution.
- Watermark Forging: Imprinting a victim’s unique watermark onto malicious or third-party content to maliciously frame them.
Methods
WMaGi achieves both removal and forging through a three-stage pipeline under a strict black-box threat model (the attacker only collects watermarked samples from a user/service):
Watermarked Images (X) ──> [Pre-trained Diffusion Denoising] ──> Purified Images (X̂)
│
▼
[Wasserstein GAN Mapping: X ⟷ X̂]
│
┌────────────────────┴────────────────────┐
▼ ▼
Removal: Generator(X) ──> X̂ Forging: Generator(X̂) ──> X
- Data Collection: The adversary gathers watermarked images sharing the target secret watermark bitstring (e.g., scraped from a user’s social feed or API outputs).
- Data Pre-processing: Rather than requiring paired clean originals, WMaGi passes the collected images through an off-the-shelf, pre-trained unconditional diffusion model (DiffPure/DDPM). The denoising process perturbs the imperceptible watermark signal enough to render it unrecoverable, producing a pseudo-clean reference set $\hat{\mathcal{X}}$.
- Adversarial Mapping via GAN: Using the collected set $\mathcal{X}$ and the purified set $\hat{\mathcal{X}}$, the attacker trains a Wasserstein GAN (WGAN) augmented with perceptual ($L_1$, MSE, and LPIPS) loss terms:
- For Removal: The generator maps $\mathcal{X} \to \hat{\mathcal{X}}$ to erase watermark patterns while preserving high-frequency image structure.
- For Forging: The generator maps $\hat{\mathcal{X}} \to \mathcal{X}$ to inject the victim's signature watermark residual into unwatermarked images.
Datasets
The authors evaluated WMaGi across multiple datasets and generative watermarking setups:
- CIFAR-10 ($32 \times 32$): Baseline image benchmark.
- CelebA ($64 \times 64$): Celebrity face dataset used as the primary benchmark for post-hoc watermarking (StegaStamp).
- FFHQ ($64 \times 64$ / higher): Evaluated against model-integrated watermarks in diffusion models (EDM) and GANs (WGAN-div).
- LSUN-bedroom ($256 \times 256$): Used to evaluate higher-resolution, complex scenes and explore diffusion removal properties.
No new dataset benchmarks were introduced; existing standard benchmarks were modified with watermarking encoders.
Key Insights and Ideas
- Watermarks behave like learnable noise residuals: Even when an attacker has no clean images and no knowledge of the encoder/decoder, a GAN can learn the distribution shift between watermarked images and diffusion-purified images.
- Removal and forging are duals: The exact same architectural framework can forge or remove watermarks simply by reversing the generator's source and target distributions.
- Few-shot adaptability: Once a base GAN model is trained, it can be fine-tuned with as few as 10 to 50 samples of an unseen watermark to successfully remove or forge it.
- Inference efficiency matters: While pure diffusion denoising can disrupt watermarks, it alters image semantics and requires hundreds of iterative sampling steps. Distilling this behavior into a feed-forward generator produces comparable or better image fidelity while slashing inference time by orders of magnitude.
Key Results
- Watermark Removal: Reduced watermark decoding accuracy from 100% to near-random chance (~47–54% bit accuracy for 4-bit to 48-bit payloads) across CIFAR-10, CelebA, and prior watermarking methods (WGAN-div, EDM), while maintaining high fidelity (CelebA PSNR $\approx 25\text{--}26\text{ dB}$, SSIM $\ge 0.88$).
- Watermark Forging: Forged unseen watermarks onto clean images with 90.14% to 99.11% bit accuracy on CelebA (32-bit and 48-bit payloads), successfully passing verification checks without substantial image degradation.
- Speedup: On an NVIDIA A100 GPU processing 1,000 images, WMaGi took 0.46 seconds, compared to 2,325 seconds ($5,050\times$ slower) for small-noise diffusion (DMs) and 5,231 seconds ($11,373\times$ slower) for large-noise diffusion (DMl).
- Few-Shot Performance: With only 100 reference images containing an unseen 32-bit watermark, WMaGi achieved a 93.47% forging accuracy and reduced removal bit accuracy to 53.27%.
Limitations and Future Work
- Higher Resolutions and Complex Scenes: When scaling up to $256 \times 256$ resolution on diverse indoor scenes (LSUN-bedroom), forging bit accuracy plateaued around 75–80%, falling short of reliable real-world framing attacks.
- Dependency on Denoising Heuristics: WMaGi relies on the initial diffusion model being able to sufficiently weaken the watermark during preprocessing without completely destroying source image fidelity.
- Domain Scope: The current framework focuses exclusively on image watermarks; the authors leave text and audio watermarking vulnerabilities to future exploration.
Summary updated