Hidden in the Noise: Two-Stage Robust Watermarking for Images

Kasra Arabi, Benjamin Feuer, R. Teal Witter, Chinmay Hegde, Niv Cohen

Abstract

As the quality of image generators continues to improve, deepfakes become a topic of considerable societal debate. Image watermarking allows responsible model owners to detect and label their AI-generated content, which can mitigate the harm. Yet, current state-of-the-art methods in image watermarking remain vulnerable to forgery and removal attacks. This vulnerability occurs in part because watermarks distort the distribution of generated images, unintentionally revealing information about the watermarking techniques. In this work, we first demonstrate a distortion-free watermarking method for images, based on a diffusion model's initial noise. However, detecting the watermark requires comparing the initial noise reconstructed for an image to all previously used initial noises. To mitigate these issues, we propose a two-stage watermarking framework for efficient detection. During generation, we augment the initial noise with generated Fourier patterns to embed information about the group of initial noises we used. For detection, we (i) retrieve the relevant group of noises, and (ii) search within the given group for an initial noise that might match our image. This watermarking approach achieves state-of-the-art robustness to forgery and removal against a large battery of attacks.

What Problem Does This Paper Address?

As generative AI produces increasingly realistic "deepfakes," the need for reliable image provenance—watermarking—has become a societal necessity. However, most current watermarking techniques face a "catch-22": to be detectable, they must slightly alter the image's pixel distribution. These subtle distortions act as a beacon for attackers, who can use "regeneration attacks" (running the image back through a different diffusion model) to strip the watermark or even forge it onto unauthorized content.

Comparison of WIND with existing watermarking methods like TreeRing and Gaussian Shading.

The challenge is creating a watermark that is distortion-free (indistinguishable from a standard AI-generated image) yet robust enough to survive heavy editing, and efficient enough to be detected among millions of possible keys.

Methods

The authors introduce WIND (Watermarking with Indistinguishable and Robust Noise), which leverages the mathematical property that diffusion models (specifically using DDIM inversion) can map an image back to the exact noise pattern used to create it.

Generation: The Two-Stage Embedding

  1. Stage 1 (Group Identification): The system partitions a massive set of $N$ possible initial noises into $M$ groups. It embeds a specific Fourier-domain pattern into the noise to act as a "group identifier." This is similar to prior methods like Tree-Ring but used here only for coarse sorting.
  2. Stage 2 (The Secret Key): The actual watermark is the specific pseudo-random noise vector itself, generated using a secret cryptographic salt and a hash function. Because this noise is drawn from a standard Gaussian distribution, it doesn't distort the image's statistical properties.
The WIND watermark generation pipeline, showing the integration of noise and Fourier patterns into the diffusion process.

Detection

To verify a watermark, the suspected image is "inverted" to reconstruct its initial noise. The detector first extracts the Fourier pattern to identify which of the $M$ groups the noise belongs to. It then performs a targeted search only within that group to find a high-correlation match with the original secret noise. This reduces the search space from $O(N)$ to $O(N/M)$, making the process feasible for large-scale deployment.

Key Insights and Ideas

  • Noise as the Watermark: The most elegant insight is that you don't need to add a watermark to a diffusion-generated image; the initial noise is the watermark. If you can reconstruct the noise, and you have a record of the noise you used, you have proof of origin.
  • Distortion-Free is Attack-Proof: Because the initial noise is exactly what the model expects, there is no "distribution shift." Attackers cannot easily detect the presence of the watermark to remove it, nor can they easily approximate the "watermark signal" to forge it onto other images.
  • Inversion Robustness: Surprisingly, even after an image is attacked (rotated, blurred, or compressed), the reconstructed noise remains significantly more correlated with the original noise than with any random noise pattern.

Key Results

  • Superior Robustness: In a "battery of attacks" (including 75° rotation, JPEG compression, and Gaussian blur), WIND maintained near 100% detection accuracy in scenarios where previous state-of-the-art methods like Tree-Ring dropped to near 0%.
  • Resistance to Regeneration: Against the "Zhao et al. regeneration attack"—a common "watermark killer"—WIND maintained a 100% detection rate even after 50 iterations of the attack.
  • Image Quality: WIND achieved an FID score of 24.33, the lowest (best) among compared methods, indicating that it preserves image realism better than techniques that more aggressively perturb the latent space.
  • Speed: Using a single NVIDIA RTX 3090, the detection phase for 100,000 noise samples takes approximately 22 seconds, proving the efficiency of the two-stage search.

Limitations and Future Work

Detection currently requires access to the private model weights used for generation to perform the DDIM inversion. This means the watermark is "private"—only the model owner can verify it. Additionally, while the watermark is robust to global transformations, a "slight edit" (like adding a small object to a watermarked image) might be an easier path for an attacker than trying to forge the noise itself. Future work may explore scaling to billions of images using advanced vector database search techniques like HNSW.

Summary updated

Abstract

As the quality of image generators continues to improve, deepfakes become a topic of considerable societal debate. Image watermarking allows responsible model owners to detect and label their AI-generated content, which can mitigate the harm. Yet, current state-of-the-art methods in image watermarking remain vulnerable to forgery and removal attacks. This vulnerability occurs in part because watermarks distort the distribution of generated images, unintentionally revealing information about the watermarking techniques. In this work, we first demonstrate a distortion-free watermarking method for images, based on a diffusion model's initial noise. However, detecting the watermark requires comparing the initial noise reconstructed for an image to all previously used initial noises. To mitigate these issues, we propose a two-stage watermarking framework for efficient detection. During generation, we augment the initial noise with generated Fourier patterns to embed information about the group of initial noises we used. For detection, we (i) retrieve the relevant group of noises, and (ii) search within the given group for an initial noise that might match our image. This watermarking approach achieves state-of-the-art robustness to forgery and removal against a large battery of attacks.
BibTeX
@inproceedings{arabi2025hidden,
  author = {Kasra Arabi and Benjamin Feuer and R. Teal Witter and Chinmay Hegde and Niv Cohen},
  title = {{Hidden in the Noise: Two-Stage Robust Watermarking for Images}},
  booktitle = {International Conference on Learning Representations},
  year = {2025},
  url = {https://openreview.net/forum?id=ll2nz6qwRG},
}
Figure