acceptodds
Withdrawn from ICLR 2027

SAEMark: Architecture-Agnostic Image Watermarking via Sparse Autoencoder Representations

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Watermarking diffusion models is essential for tracing the provenance and enabling the trustworthy authentication of generated content. However, existing methods typically rely on architecture-specific generation or inversion procedures, limiting their applicability to emerging generative paradigms such as flow matching. We observe that Sparse Autoencoders (SAEs) can independently learn internal representations for different generative architectures, yielding structured and semantically rich feature spaces that provide a new substrate for decoupling watermarking from the underlying generation process. Based on this insight, we propose SAEMark, the first SAE-based watermarking framework for diffusion models, which directly embeds and extracts watermarks in the SAE feature space, thereby decoupling watermarking from the specific generation process. This representation-level design enables a unified watermarking mechanism across heterogeneous generative architectures while substantially improving robustness against watermark removal and image manipulation attacks. We conduct experiments across multiple architectures, covering 8 representative watermarking methods and 9 watermark removal attacks. The results show that SAEMark achieves leading watermark robustness while preserving high generation quality, demonstrating the potential of SAE representations as a unified substrate for robust and architecture-agnostic watermarking across generative architectures. Code is available at https://anonymous.4open.science/r/SAEMark-exp-C3E0/.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
watermarking, image generative AI

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_FSb8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SAEMark, a watermarking method for text-to-image models that works in the sparse autoencoder (SAE) feature space. It selects one SAE feature and boosts it during generation to embed the watermark. Detection measures that feature’s activation with a single forward pass, without inversion. Experiments on SD-v2.1 and FLUX.1-schnell show strong robustness against nine attacks compared with eight baselines.

Critical strengths
  • The problem is important, and the proposed idea is new and simple.
  • Detection is fast and needs no inversion.
  • Average robustness is strong (0.97 on SD-v2.1 and 0.90 on FLUX).
Critical weaknesses
  • The “architecture-agnostic” claim is overstated. Each model needs its own SAE, hookpoint, feature search and tuning.
  • 1% FPR is lenient for real deployment. Results at 0.1% FPR would be more convincing.
  • It is a single-bit method. It can only detect a watermark, not identify users. A discussion of extending it to the multi-bit setting is missing.
  • Only two models are tested (SD-v2.1 and FLUX.1-schnell). Testing more models, such as SDXL or SD3, would support the generality claim.
Questions for authors

Please refer to the critical weaknesses

Reviewer_r97uOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SAEMark, an SAE-based watermarking method for text-to-image models. It selects a model-specific SAE feature, strengthens that feature during generation, and detects the watermark by measuring its activation in a forward pass on the query image. The authors evaluate the method on Stable Diffusion v2.1 and FLUX.1-schnell against eight watermarking baselines and nine image transformations or regeneration attacks. The results suggest strong detection performance under the tested conditions.

Critical strengths

The paper explores a timely direction by using interpretable internal representations as a carrier for generative-image watermarks. Its embedding and detection pipeline is clearly described, and the evaluation covers two different generation paradigms, several watermarking baselines, and a range of image transformations. The reported detection speed and ablations on SAE feature selection are also useful.

Critical weaknesses

(1)The central operation is to increase the activation of a selected feature during generation. However, the paper does not compare against direct activation steering, random or frequency-based feature selection, or other non-SAE representation directions. Consequently, it remains unclear whether the gains arise specifically from SAE representations or simply from intervening on a model’s internal activations. (2)The study does not test attacks designed to identify and suppress the selected SAE feature, optimization-based removal, or repeated editing and regeneration. (3)The quality analysis relies mainly on FID and CLIP Score, which do not directly measure paired perceptual changes caused by embedding. Paired metrics such as LPIPS and PSNR, human evaluation, and comparisons against the corresponding unwatermarked generations would better substantiate the claim that generation quality is preserved.

Questions for authors

Please refer to the weakness

Reviewer_iCNKOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes SAEMark, an image watermarking method that uses sparse autoencoder (SAE) representations. The method selects an SAE feature, modifies its activation during image generation, and detects the watermark by measuring the same feature's activation. The authors evaluate the approach on Stable Diffusion 2.1 and FLUX.1-schnell.

Critical strengths
  • The reported robustness results are strong, particularly against image transformations and regeneration attacks.
Critical weaknesses
  • My main concern is that the method essentially increases the activation of a selected SAE feature and detects that activation later. But where is the secret key or the actual security mechanism? The paper only considers single-bit watermark detection, without a clear mechanism for distinguishing different sources or assigning different watermark identities. The authors evaluate several image transformations and regeneration attacks, but none specifically targets the selected SAE feature. What prevents someone from suppressing this feature to remove the watermark or activating it to forge one?

  • Another major concern for me is that watermark detection requires access to the original generative model. The query image is encoded again and passed through the model to extract its internal SAE feature. I don't find this a convincing verification setup, especially for proprietary models. Providing external verifiers with the model and its internal representations could introduce information leakage risks. If verification is performed only by the model owner, then how can an independent party verify the watermark? The paper does not sufficiently discuss this access and trust assumption, which I think is important for a method motivated by provenance and authentication.

  • The authors claim to decouple watermarking from the underlying generative architecture. However, the method still requires a different SAE, model-specific feature selection, and access to the corresponding internal model blocks. I understand that the same general procedure can be applied to different architectures, and the experiments demonstrate this on two models. However, I think this is different from having an architecture-agnostic watermarking mechanism.

  • The authors evaluate different SAE features and their effects on watermark detectability and image quality, which is useful. However, I would like to see a comparison with simpler internal activation-steering approaches that do not use SAEs. Without this, it is difficult to understand how much of the improvement comes specifically from using SAE representations.

Questions for authors
  • What is the security mechanism of SAEMark? How would the method support different watermark identities and prevent targeted removal or forgery?
  • In what sense is the method architecture-agnostic when a separate SAE, feature, and internal model access are required for each architecture?
  • Can the authors demonstrate that SAE-based feature steering provides an advantage over simpler internal activation interventions?

Comments on the paper

1 comment
Area_Chair_NHVUOpenReview ↗

Participate in the discussion period

Dear reviewers,

The authors have replied to your comments and the discussion is close to the end. We encourage you to participate in the discussion to see whether your questions and concerns have been answered by the authors. Thank you!

Best, Your AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.