AEON: Adaptive Embedding Optimized Noise for Robust Watermarking in Diffusion Models

Muhammad Shahid Muneer, Simon S. Woo

TL;DR

This paper introduces AEON, a novel watermarking technique for AI-generated images that embeds invisible watermarks directly into the diffusion model's latent space. AEON's adaptive approach ensures watermarks are semantically aligned with the image content, making them robust against various removal attacks while preserving image quality.

What Problem Does This Paper Address?

The rapid advancement of AI image generation models like Midjourney and Stable Diffusion has created a pressing need for robust methods to protect the copyright and authenticity of synthetic images. Existing watermarking techniques often struggle to balance watermark invisibility with robustness against sophisticated removal attacks, such as reconstruction and adversarial methods. Some methods compromise image quality by embedding visible artifacts or are easily defeated by these attacks. Furthermore, traditional post-hoc watermarking methods applied after image generation are often less effective and can introduce noticeable distortions. This paper addresses the challenge of creating an invisible, high-quality watermark that is resilient to a wide range of attacks, particularly those targeting diffusion-generated content.

Methods

AEON employs an adaptive, in-diffusion watermarking strategy that integrates the watermark into the intermediate latent noise of the diffusion process. The core of AEON involves three key stages:

  1. Watermark Hash Generation: A learnable hashing network is trained to generate a "hash noise" that is semantically similar to the intermediate latent representation of the image. This is achieved using a combination of cosine similarity loss (to ensure semantic alignment) and reconstruction loss (for adversarial robustness). This network adapts the watermark to the specific content being generated.
  2. Adaptive Watermark Embedding: The generated hash noise is then blended into the frequency domain of the intermediate latent noise during the diffusion pipeline. This adaptive embedding, rather than injecting a static watermark, allows the watermark to be infused seamlessly into the image's semantic structure. The embedding can occur at later diffusion stages, minimizing visual artifacts.
  3. Watermark Verification: The watermark is verified by reconstructing the image from the watermarked latent space and then extracting the watermark using a reverse diffusion process. The extracted watermark is then compared against a stored hash to determine its presence.

A key innovation is the use of a semantic hashing network that generates adaptive noise, ensuring the watermark aligns with the image's content and prompt, unlike static watermarks that can lead to quality degradation or artifacts.

Datasets

  • MS COCO [22]: Used for training the hash network.
  • Gustavosta [16] dataset: Used for text-to-image generation prompts.

The paper does not introduce new benchmarks but evaluates against existing SOTA methods on these standard datasets.

Key Insights and Ideas

  • Semantic Alignment is Crucial: The paper highlights that embedding watermarks that are semantically aligned with the image content, rather than just being statistically embedded, significantly improves robustness against attacks. AEON's adaptive hashing network achieves this by generating noise that mirrors the image's semantic features.
  • Late Embedding for Quality: Injecting the watermark at later diffusion stages, as AEON allows, significantly improves the visual quality of the watermarked image and reduces artifacts, a common issue with earlier in-diffusion methods.
  • Adaptive Noise as Watermark: The core idea of using "optimized noise" as the watermark, which is dynamically generated based on the image content, is a novel approach to achieving both invisibility and robustness.
  • Reconstruction Attack Resilience: AEON's design, particularly its reconstruction loss during training, specifically targets resilience against attacks that attempt to reconstruct the original image by removing the watermark.

Key Results

  • Superior Robustness: AEON consistently achieves high Area Under the Curve (AUC) scores across various attacks, outperforming existing SOTA methods like ROBIN and Tree-Ring, especially under cumulative attacks. For instance, in Table 1, AEON achieves an average AUC of 0.994 for Stable Diffusion and 0.992 for ImageNet Diffusion, often reaching 1.000 for specific attacks.
  • High Image Quality: The method maintains excellent image quality, as evidenced by high PSNR, SSIM, and MSSIM scores (Table 3), with minimal degradation compared to the original images. For example, AEON achieves a PSNR of 25.93 on Stable Diffusion and 25.18 on ImageNet, comparable to or better than other methods.
  • Invisibility: Qualitative comparisons (Figure 5) show that AEON's watermarks are virtually indistinguishable from the original image, unlike some other methods that leave noticeable artifacts.
  • Adaptive Embedding Benefits: Table 4 demonstrates that embedding the watermark at later diffusion steps (e.g., step 40) leads to higher PSNR and AUC values, indicating improved quality and robustness.

Limitations and Future Work

  • Computational Cost: While AEON achieves strong results, the diffusion process itself is computationally intensive, and watermarking adds to this overhead. The paper mentions verification time, but the overall training and generation time might be a consideration.
  • Trade-offs in Embedding Steps: Table 4 shows a trade-off: later embedding steps improve PSNR and AUC, but the paper notes that existing research [20] can leave artifacts when embedding too late. AEON aims to mitigate this, but a perfect balance might still be an area for further exploration.
  • Future Work: The authors suggest exploring further improvements in watermark strength and resilience against more complex adversarial scenarios. They also implicitly point towards optimizing the balance between embedding time, image quality, and watermark robustness.

Summary updated

BibTeX
@inproceedings{muneer2026aeon,
  author = {Muhammad Shahid Muneer and Simon S. Woo},
  title = {{AEON: Adaptive Embedding Optimized Noise for Robust Watermarking in Diffusion Models}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2026},
}
Figure