acceptodds
Withdrawn from ICLR 2027

ClaimMark: Decoder-Differential Watermarking for Claimed-Key Image Provenance

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multi-user image provenance requires more than watermark detection: a verifier must determine whether the evidence supports a submitted key. We introduce ClaimMark, a training-free watermarking method for reference-assisted key attribution. ClaimMark constructs reproducible, high-dimensional carriers from secret keys and shapes their frequency content and spatial amplitude while constraining their deviation from the original pseudorandom directions. Decoder-differential embedding adds only the watermark-induced change in variational autoencoder (VAE) decoder output to the original image, avoiding transfer of the VAE's reconstruction error. Given a suspect image and a trusted original, the verifier compares the observed residual with a regenerated key-specific template using spatial and spectral correlations. It accepts a claim only when the score satisfies both absolute and decoy-relative thresholds. ClaimMark requires no watermark-specific training, diffusion inversion, or iterative latent optimization. It achieves PSNR values of dB on MS-COCO and dB on DiffusionDB. Across the evaluated attribution conditions, average genuine-claim acceptance is and wrong-key acceptance is . These results support reference-assisted key verification under the evaluated image transformations.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Invisible Watermarking, Image Provenance, Key Attribution, Multi-User Watermarking, Training-Free Watermarking, Robust Watermarking

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_jisuOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ClaimMark, a training-free watermarking method for checking whether a submitted key matches an image. Each secret key generates a pseudo-random latent pattern. Only the change in VAE decoder output is added to the original image. This avoids VAE reconstruction error and gives high PSNR (about 51–53 dB). The verifier compares the suspect image with the trusted original. The method reports 99.20% genuine-claim acceptance and 0.162% wrong-key acceptance.

Critical strengths
  • The method achieves high fidelity and strong robustness to many attacks.
  • The paper is well written and clearly organized.
  • The method needs no training.
Critical weaknesses
  • The verifier must store every original image. This raises storage costs and limits deployment. It also makes ClaimMark a non-blind method, while most recent watermarking research targets blind verification. Furthermore, all compared baselines are blind and need no original image. The comparison is therefore unfair.
  • Geometric results assume the attack is known. Real deployment needs robustness to unknown attacks.
  • The evaluation uses only 65 keys. Real systems may serve millions of users.
  • Ablation uses only 20-100 images.
Questions for authors

Please refer to the critical weaknesses

Reviewer_o1AhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces ClaimMark, a training-free reference-assisted watermarking framework for claimed-key image provenance that uses decoder-differential embedding to avoid VAE reconstruction error transfer and decoy-calibrated verification. These results demonstrate strong signal discrimination under the conditions of known reference image, known or reproducible geometric transformations, non-adaptive attacks, and a controlled candidate key pool. However, they are not sufficient to support broader claims of general image provenance or robust attribution under adversarial conditions.

Critical strengths
  • The decoder-differential embedding design elegantly cancels out the VAE reconstruction error floor, achieving a 24+ dB PSNR improvement over direct latent decoding. This is a simple, well-motivated, and impactful technical contribution.
  • The paper frames the problem as claimed-key verification (accept or reject a submitted key) rather than mere detection or closed-set identification, and introduces a decoy-calibrated four-threshold decision rule to operationalize false-claim rejection.
Critical weaknesses
  1. Limited applicability requires narrower claims. The paper does not address general blind watermarking or open-world image provenance, but rather a strongly prior-dependent reference-assisted claim verification problem. The verifier must store a per-image trusted reference, know user-key associations, and re-run the same VAE—making this more akin to a residual consistency check on server-side registry records rather than a source proof usable by independent third parties. If the reference is missing, compressed, re-encoded, or version-mismatched, both the observed residual and the predicted template will fail to align.
  2. Geometric robustness results assume oracle attack knowledge, and the comparison protocol with baselines is asymmetric. Rotation, scale, and perspective experiments use known attack labels or transformation parameters: for rotation, the verifier exhaustively tries four directions; for scale and perspective, the verifier directly reproduces the same transformation from the benchmark to generate the template. Taking perspective attack as an example, ClaimMark achieves 100% Top-1 accuracy while RingID only reaches 1.0% and Shallow Diffuse only 1.4%—but this large gap is likely because ClaimMark has access to the reference image and knows which transformation occurred, whereas baseline methods operate in blind detection mode. If RingID and other baselines were also provided with reference images and allowed to align before verification, would the gap remain so large? The current experiments fail to rule out the confound that the comparative advantage stems from asymmetric reference information rather than inherently better robustness.
  3. False-key forgery and adaptive attacks are not substantively evaluated. Wrong keys are randomly constructed, without an attacker selecting keys based on observed scores, templates, or repeated queries. However, the key-conditioned residual is deterministic, and the verifier exposes accept/reject outcomes as well as intermediate scores for each claim—an attacker could potentially search candidate keys, use gradient approximation, repeat queries, or aggregate residuals across multiple user images to find keys that pass more easily. In particular, intermediate scores returned by the verifier (e.g., residual energy, correlation coefficients) may provide gradient information that lowers the bar for forgery.
  4. Computational cost and verification efficiency are not adequately discussed. Verification requires running the decoder and computing multi-dimensional scores for 64 decoy keys plus the claimed key, with complexity scaling linearly with the number of decoys N. The paper only provides asymptotic complexity analysis without reporting any empirical runtime data—per-image verification latency, GPU memory footprint, or batch throughput are all absent. Increasing N from 64 to 128 would further reduce false acceptance but nearly double decoding cost, making the error-rate vs. service-cost tradeoff a central practical concern. No potential acceleration strategies are discussed.
Questions for authors
  • Is the intended scenario third-party-verifiable provenance or a server-side registry check? How does verification behave when the stored reference is missing, compressed, re-encoded, or from a different VAE version?
  • In the geometric-robustness experiments, ClaimMark knows the transformation and holds the reference image while baselines detect blind. If baselines were given the same reference and allowed to align before verification, how much of the reported gap (e.g., 100% vs. 1.0% Top-1 under perspective attack) would remain?
  • Does the verifier expose intermediate scores (residual energy, correlation coefficients), and is the key-conditioned residual deterministic? If so, is there any evidence against score-guided or repeated-query key search by an attacker?
  • Can you report empirical verification cost—per-image latency, GPU memory, throughput—at N=64 and N=128 decoy keys? The false-acceptance vs. decoding-cost tradeoff seems central to practical deployment.
Reviewer_D2xDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents ClaimMark, a method for verifying whether an image was generated using a claimed key, with a trusted reference. Given an original image and a suspect image with the perceived owner's key, should the claim be accepted is what it studies.

The method constructs a key derived perturbation in the VAE latent space and adds the resulting change in the decoder output to the original image. During verification, the observed difference between the original and the suspect image is compared against the expected difference for the claimed key. The claim is accepted if four verification thresholds are satisfied, including two based on comparisons against 64 incorrect keys.

The method does not require additional training or optimization, but requires access to the original image during verification. The authors evaluate it on 500 images each from MS-COCO and DiffusionDB, using 65 candidate keys and several image attacks. They report PSNR values of around 51–53 dB, an average genuine-claim acceptance rate of 99.20%, and a wrong-key acceptance rate of 0.162%.

Critical strengths
  1. The paper addresses an important gap in how watermark verification is evaluated. The distinction between watermark detection, closed-set key identification, and claimed-key verification is useful, as strong performance on one does not necessarily imply strong performance on the others (ref: Table 5b where presence-only verification accepts 93.175% of wrong-key claims). I consider this problem framing and the corresponding evaluation metrics to be one of the main strengths of the paper.
  2. The proposed verifier is simple and has some theoretical support. Using wrong keys as decoys to set the verification thresholds is a useful approach that does not require additional training or optimization. The reconstruction-error cancellation argument and rank bound also support the method under the stated assumptions.
Critical weaknesses
  1. The reported fidelity and verification performance may depend on how the images are saved, rather than the proposed embedding method. The PSNR stays near 51.1 dB across embedding strengths, even though the pre-serialization PSNR is 82–87 dB. Figure 2 also shows that the difference between the watermarked and original images is almost entirely 0 or -1, which suggests truncation during 8-bit conversion rather than the described clip(x + r) operation. In my independent reimplementation (SD 2.1-base VAE, 50 COCO images, one key, float32, no null-space gain), truncation reproduced 51.20 dB PSNR, while rounding left 44 of 50 images unchanged and substantially reduced the true-key verification score. My setup differs from the authors', so this is not conclusive, but the pattern in Figure 2 needs an explanation. This matters because it is unclear how much of the reported fidelity and verification performance comes from the proposed embedding method itself.

  2. The evaluation does not fully support the claimed verification performance or the comparisons with prior work. ClaimMark has access to the original image and, for scale and perspective attacks, the transformation parameters, while the baselines do not. This makes comparisons such as 100% versus roughly 1% Top-1 accuracy difficult to interpret. The reported 0.162% wrong-key acceptance rate is also measured per claim, rather than per image. Table 8 suggests that the probability of accepting at least one wrong key among 64 candidates could be closer to 10%, although the evaluation protocols are not identical. The results are also sensitive to the chosen verification threshold, and the headline averages exclude regeneration, learned compression, and combined attacks, where performance drops substantially. These issues make it difficult to assess how much the method improves on existing approaches and how reliable the reported verification performance is.

Questions for authors

Clarifications:

  1. How were the verification thresholds chosen? What held-out data was used, how large was it, and was it separate from the evaluation data? Section E.4 seems to suggest the thresholds were selected using the evaluated trials.
  2. Table 12 reports decoder RMS around 1.3e-4, compared with serialization RMS around 5.5e-3. In Figure 2, pixels appear to change by -1 exactly when the residual is negative, and remain unchanged otherwise. What quantization or rounding procedure is used when saving the images? Is this behavior intentional? How do the fidelity and verification results change if standard rounding is used instead of truncation?
  3. Which values of M and rho, model, and numerical precision were used for each main result? The paper mentions M=1 and rho=0.25, but the supplied scripts default to M=4 and rho=0.75. I would also like to know whether the null-space gain is used in the reported results.
  4. Why does the same nominal configuration give Claim-WDR of 92.27% in Table 11 and 99.55% in Table 5, and Top-1 accuracy of 94.09% versus 99.55%? Are the images, attacks, keys, or evaluation protocols different?
  5. Table 5's caption mentions 100 images, but the reported Claim-WDR values appear to be multiples of 1/220, which would correspond to 20 images across 11 attacks. How many images were actually used?
  6. Which wrong-key construction was used for the main WKAR results? Table 8 shows that WKAR changes considerably depending on how wrong keys are constructed. How does this affect the reported 0.162% rate?
  7. Can the authors clarify ZoDiac's results in Table 2? A WDR of 0.998 at p=0.90 with 21.6% wrong-key acceptance seems inconsistent with the reported mean true-key rank of 37.32. How were these metrics calculated?

Additional results that could change my assessment:
8. Could the authors report Claim-WDR, WKAR, and Any-Wrong under regeneration, learned compression, and combined attacks? I would also like to see the geometric attack results without providing the transformation parameters to the verifier. Image-level confidence intervals would help assess how stable these results are.
9. How does the proposed method compare with a simple keyed pattern added directly in image space, using the same embedding energy and verification procedure? Currently, it is unclear whether the VAE-based embedding provides an advantage over this simpler alternative.

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

Related papers

Discussion (0)

Sign in to comment.