acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Image Watermarking in the Era of Frontier Generative Models

39 views

Abstract

Image watermarks are increasingly used to establish the provenance of AI-generated content. We demonstrate a watermark-removal attack that repeatedly regenerates a watermarked image with GPT-Image-2. The attack requires only API access, without knowledge of the watermarking method, detector access, or local model deployment. It achieves 89% average removal across 13 post-hoc and in-generation methods, with a removal-fidelity trade-off comparable to leading specialized attacks. The evaluated content-based watermarks resist regeneration. However, we show that an existing method that inserts visual concepts from a fixed private inventory is vulnerable to concept extraction. Our extraction attack recovers 79% of its concepts from 3,000 watermarked images. The recovered candidates enable 93.9% removal after ten targeted edits, at a cost to image fidelity, and 70.4% spoofing success when generating images outside the watermarking system. We propose K-SemMark-R, a registered content-based watermark that independently samples prompt-compatible carrier values for each image and stores verified carriers in a private registry. Experiments show that most initially detected images remain detectable by carrier verification after regeneration, although the method does not enroll every prompt. These findings highlight the potential of content-based watermarking and motivate research into image-specific designs that resist regeneration without relying on a shared concept inventory.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Image watermarking, AI-generated image provenance, watermark removal, multimodal generative models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_AvC6OpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper evaluates image watermark robustness under repeated regeneration with general-purpose image models, develops a concept-extraction attack against IConMark, and proposes K-SemMark-R, a registered watermarking scheme based on image-specific semantic attributes. The experiments demonstrate the practical accessibility of regeneration attacks and expose vulnerabilities associated with reusing a fixed concept inventory. However, the paper’s central methodological contribution and the benefits of the proposed defense remain insufficiently established.

Critical strengths

The paper addresses a relevant security problem and provides a broad evaluation of watermark removal using readily accessible image-generation APIs. Its most compelling contribution is identifying and exploiting the leakage of IConMark’s fixed concept inventory to enable both removal and spoofing. The exploration of image-specific semantic carriers is also a potentially useful direction.

Critical weaknesses
  1. Regeneration-based watermark removal already has clear precedents. For example, CtrlRegen removes watermarks through controllable regeneration from clean noise, while IConMark already introduces interpretable concept-based watermarking and hybrid schemes. Much of this paper is devoted to an engineering exercise in reproducing and evaluating existing ideas, without offering distinctive insights into regeneration-based attacks or defenses.
  2. For example, although the paper compares several regeneration models, it does not explain why GPT-Image-2 achieves a better removal–fidelity trade-off or identify the underlying factors driving these performance differences. This illustrates the paper’s emphasis on empirical evaluation rather than deeper insights into regeneration-based attacks.
  3. The paper’s organization and logical flow need improvement. The current presentation does not clearly establish its main focus or core technical novelty, while a substantial portion of the paper is devoted to demonstrating the practical accessibility of general-purpose image-generation APIs and evaluating their attacks across watermarking methods.
  4. IConMark already performs well against the evaluated regeneration attacks, and the proposed design does not demonstrate a substantial improvement over it. The evaluation also lacks comparisons with stronger in-generation watermarking baselines designed to resist regeneration attacks. One relevant example is Attack-Resistant Watermarking for AIGC Image Forensics via Diffusion-based Semantic Deflection (ICLR 2026). Discussing such methods and their applicability to the sequential editing setting would help contextualize the proposed approach.
  5. K-SemMark-R stores the original image embedding and semantic attributes, first retrieves a similar registered image, and then checks its attributes. The detector therefore has access to per-image registration information. Consequently, the observed robustness may primarily come from storing and matching reference images in the registry, rather than from actively embedding a watermark.
Questions for authors

Please refer to the weaknesses discussed above, particularly points 1 and 2.

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

The paper shows that an API-only regeneration attack strips most existing watermarks at competitive fidelity. Content-based watermarks survive it, but IConMark's small concept inventory can be recovered and exploited for removal and spoofing. As a remedy, the authors propose K-SemMark-R, which uses per-image carriers stored in a private registry and is more robust to regeneration, though it sometimes abstains and its verifier is not fully stable.

Critical strengths
  • Realistic threat model. The attacker only has API access, gets no detector feedback, does no per-image optimization.
  • A new insight on IConMark. A shared secret concept list leaks through objects that keep showing up where they shouldn't. The extraction is well calibrated and transfers across models (GPT-Image-2 to FLUX.2)
Critical weaknesses
  • No adaptive attack on K-SemMark-R, and the main attack favors it. The regeneration prompt explicitly says to "preserve colors, lighting, style," which are the very attributes K-SemMark-R uses as carriers. A looser prompt already drops detection too much. The obvious attack is to ask the model to keep the subject but re-roll materials, lighting, and layout, and it isn't tested. The authors admit this, but without it the defense claim is weak.
  • Small samples, no error bars. The headline result is 10 of 68 images, which is close to the 5 of 68 flips on unattacked images. Spoofing uses 27 inputs per setting, but Fig. 4 says 25 scenes. The main cells use 100 images. There are no confidence intervals or repeated runs.
Questions for authors
  • Can you try an adaptive prompt that keeps the subject but re-samples materials, lighting, and layout? How much does detection drop?
  • How does concept recall scale with the number of prompts and images? An ablation of the two terms in Eq. 3 would show which one matters.
  • How does image-level FPR grow with registry size? Is p0 stable across carrier types?
  • Will you release code, prompts, and concept lists? Did you notify the IConMark authors and the API providers?
Reviewer_FwuxOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies watermark removal with frontier image editing models and has three parts. First, it shows that regenerating a watermarked image four times with GPT-Image-2, through the API and with a fixed prompt, removes 89% of watermarks on average across 13 post-hoc and in-generation methods. The removal and fidelity trade-off is comparable to CtrlRegen, and the cost is about $0.58 per image. Second, it shows that content-based IConMark resists this attack, but its fixed concept database can be recovered. An extraction pipeline based on unexpected and persistent objects recovers 79 of 100 concepts from 3,000 images, which then enables 93.9% removal after ten edits and 70.4% spoofing. Third, it proposes K-SemMark-R. Here an LLM selects scene details left open by the prompt, a random value is sampled for each, and the values are stored with the image embedding in a private registry. Detection retrieves records by embedding and checks the stored values with a VLM. Regeneration removes 14.7% of these watermarks.

Critical strengths
  1. The attack evaluation is broad and careful. It covers 13 watermark methods and 60 baseline attack configurations. Detection thresholds are documented per method at 1% FPR (Appendix B), refusals count as failures, and full per-method and per-round tables are reported. The finding that a cheap public API matches specialized attacks is practically important now that watermarking is being deployed and regulated. For example, it removes 86% of Gaussian Shading where CtrlRegen removes 0%.
  2. The analysis of content-based watermarking is useful. The paper explains why content-based marks resist regeneration, and the concept extraction attack identifies a real weakness of a shared concept inventory. The alignment cost of IConMark is also measured (CLIP drops 11 to 14%).
Critical weaknesses
  1. The attack has limited technical novelty. Regeneration attacks are well established (Regen-Diff, Rinse, CtrlRegen), FLUX img2img is already one of the baselines, and Pleimling et al. (2026) use off-the-shelf image-to-image models against protective perturbations. The main new element is a stronger commercial model with a good prompt. The results also depend on specific proprietary model versions that may change or add safeguards, which limits reproducibility and may age the finding quickly.
  2. K-SemMark-R is not evaluated against the most obvious baseline. The registry stores each image's embedding, and retrieval alone keeps the correct record for 95.9% of attacked queries. A plain retrieval defense that matches queries against stored embeddings might therefore detect regenerated images nearly as well without any carriers. The paper does not show what the carriers add over retrieval alone.
  3. The K-SemMark-R evidence is small and noisy. The results rest on 68 detected images. Re-verifying unchanged images flips 7.4% of decisions, which is half of the reported 14.7% removal. Enrollment fails for 25 of 100 prompts. No confidence intervals or repeated runs are given.
  4. The false positive claim does not reflect deployment. One false acceptance in 4.5M images is measured against a registry of about 75 records. A per-record FPR of 1% does not bound the image-level FPR when queries match many records. Registry scale, retrieval cost, and false matches at millions of records are not evaluated.
  5. There are no adaptive attacks on K-SemMark-R. An attacker who knows the design can regenerate while asking the model to vary minor details such as lighting, materials, and small objects. A looser prompt already lowers detection from 89.5% to 79.5% on the development set. The recurrence analysis also does not run the full extraction attack.
  6. The IConMark attacks lack simple baselines. Removal should be compared with asking the editor to remove unusual or prompt-irrelevant objects, without extraction. Spoofing should be compared with inserting random or generic objects. Removal also comes with a large fidelity cost (CLIP 0.81, DreamSim 0.34), and spoofing uses only 27 images per setting. Extraction precision is 35.3%.
  7. Some smaller points:
  • The evaluation uses 100 images per method at 512×512 with no intervals.
  • No deployed watermark such as SynthID is tested.
  • The paper does not discuss whether GPT-Image-2 outputs carry their own provenance signals, such as C2PA metadata. If they do, the attack replaces one provenance signal with another.
Questions for authors
  1. What detection rate and FPR does a retrieval-only baseline achieve on the same registry and attacks, without carriers?
  2. How does the image-level FPR of K-SemMark-R scale with registry size, for example at 100K or 1M records?
  3. How does K-SemMark-R perform under an adaptive regeneration prompt that asks the model to vary unspecified details?
  4. How do naive removal and spoofing baselines compare with the extraction-based attacks on IConMark?
  5. Can you report repeated runs or intervals for the K-SemMark-R results, given the 7.4% verifier variability?
  6. Do GPT-Image-2 outputs carry C2PA metadata or an invisible watermark, and how does this affect the practical claim?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.