acceptodds
Under review as a conference paper at ICLR 2027

Not All Recovered Symbols Are Equal: Reliability-Aware Watermark Verification for Autoregressively Generated Images

Abstract

Recovering verification symbols through retokenization is a key step in verifying watermarks in autoregressively generated images. However, the recovered symbols may differ from those used during generation, especially after image transformations. Importantly, these recovery errors are not uniform: some symbols remain stable, whereas others are more likely to change. Existing watermarking methods, however, typically assign equal decision weight to recovered verification symbols despite their different recovery reliability. This limitation may lead to watermark verification failure. To address this issue, we propose Reliability-Aware Verification (RAV), which predicts symbol survival reliability from recovery-side features, including quantization geometry, posterior uncertainty, consistency, and spatial context, and uses the predicted reliability to weight watermark evidence. By emphasizing reliable symbols and suppressing unreliable ones, RAV reduces the risk of verification failure caused by unreliable recovered evidence. Experiments on WMAR and ClusterMark across three autoregressive image generation models, RAR-XL, TAMING, and LlamaGen GPT-B, show that RAV consistently improves watermark verification under diverse transformations at strict false-positive rates.

Primary area:
generative models
Keywords:
Image Watermarking, Autoregressive Image Generation, Watermark Verification, Reliability-Aware Verification

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_8DpNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how to detect watermarks in images generated by autoregressive models. Current watermarks for these models, like WMAR and ClusterMark, bias the generator toward a secret set of tokens. To check for the watermark, the detector tokenizes the image and statistically tests the token distribution. The paper shows that after edits like JPEG compression, some token positions are reliably recovered and others are not. It then proposes RAV, which is a small network that predicts how likely each token is to be recovered correctly, and gives more weight to the more reliable tokens during detection. On three ImageNet generators, RAV improves detection by a few percentage points in most settings.

Critical strengths
  1. The method is simple and cheap. It changes only the detector, needs no secret key to compute the weights, works with different watermarks, and adds less than 20 ms per image.
  2. The detection improvement transfers to neural compression, which was not used during training. On Taming, RAV outperforms a simple weighting for all six compression models (Table 11).
Critical weaknesses
  1. The learned network improves only modestly on simpler baselines. Weighting tokens by reliability is already used in text watermarking (SWEET [1], EWD [2]), and simpler weightings that need no extra training get much of RAV's gain (Tables 2 and 4).
  2. RAV does not fix the cases where these watermarks cannot be detected reliably. After a 15-degree rotation, it detects only 3.0% of watermarked images at a 1% false-positive rate (Table 8), which the authors acknowledge as a limitation. The paper also cites forgery attacks but does not test against them.

References

[1] Taehyun Lee, Seokhee Hong, Jaewoo Ahn, Ilgee Hong, Hwaran Lee, Sangdoo Yun, Jamin Shin, and Gunhee Kim. Who Wrote this Code? Watermarking for Code Generation. ACL 2024.

[2] Yijian Lu, Aiwei Liu, Dianzhi Yu, Jingjing Li, and Irwin King. An Entropy-based Text Watermarking Detection Method. ACL 2024.

Questions for authors
  1. On ClusterMark, is RAV's gain over simple confidence weighting statistically significant?
  2. What is the false-positive rate on unwatermarked images from other sources, such as real photos or other generators? The weights depend on image content, so this rate could differ.
  3. All experiments use models that generate 256×256 ImageNet images from a class label. Does RAV also improve detection on a text-to-image autoregressive model?
  4. The paper shows that RAV transfers to neural compression, but it does not say whether the removal attacks it tests were used in training. How much of RAV's gain remains on unseen removal attacks?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 8DpN

Question 1:

Thank you for this valuable question. We added image-paired significance analyses comparing RAV with simple confidence weighting on ClusterMark. RAV improves TPR on both RAR-XL and GPT-B; the gain is statistically significant on RAR-XL but not on GPT-B. The smaller GPT-B gain may reflect the strong Confidence baseline and limited headroom.

Specifically, we performed 5,000 image-paired bootstrap resamples across six attack groups. On ClusterMark/RAR-XL, RAV improves TPR over Confidence by 1.22 percentage points (pp), with a 95% CI of [0.35, 2.00], excluding zero. On ClusterMark/GPT-B, the gain is 0.46 pp (95% CI [-0.23, 0.70]), including zero. A possible explanation is that Confidence already reaches 96.55% six-group macro TPR on GPT-B, leaving limited headroom for RAV's reliability-prediction CNN, versus 91.68% on RAR-XL.

Changes: Sec. 5.3 and Appendix F.3.

Question 2:

Thank you for highlighting possible false positives under source shift. We evaluated genuinely unwatermarked images from another generator and real photographs. RAV maintains low FPRs across both external sources.

Without retraining or threshold recalibration, we applied the original detector keys and previously frozen nominal-1%-FPR thresholds to 1,000 unwatermarked LlamaGen GPT-B images and 1,000 real COCO photographs. Observed FPRs on GPT-B/COCO were 1.00%/1.50% for WMAR-RAR-XL and 1.30%/0.50% for ClusterMark-RAR-XL. FPR remains low despite modest source-dependent variation.

Changes: Sec. 5.5 and Appendix C.

Question 3:

Thank you for suggesting evaluation on text-to-image AR generation. We added watermark verification experiments on Janus-Pro-1B. RAV improves detection over both Equal and Margin, supporting its applicability to an additional AR image-generation setting.

We adapted a WMAR-style keyed-token watermark and RAV to Janus-Pro-1B, using disjoint prompt splits. For each condition, 350 genuinely unwatermarked images calibrated the threshold; another 350 watermarked and 350 genuinely unwatermarked images formed the test set. Across PNG round-trip, JPEG Q50, and blur R1, macro TPRs were 94.00%, 93.43%, and 89.90% for RAV, Margin, and Equal, respectively. RAV gains were 4.10 pp over Equal and 0.57 pp over Margin.

Changes: Sec. 5.6 and Appendix D.

Question 4:

Thank you for this suggestion. We evaluated two held-out removal-attack families, neural compression and regeneration, and checked the available training and model-selection records. RAV retains gains over Margin or Confidence on these unseen attacks.

For WMAR, six CompressAI q=3 neural-compression models were absent from predictor training. At frozen nominal-1%-FPR thresholds, RAV improved TPR over Margin by 0.58 pp on RAR-XL and 4.23 pp on TAMING. For ClusterMark, we examined three regeneration attacks absent from recorded training conditions. On RAR-XL, RAV attained 95.67% TPR versus 93.92% for Confidence, a 1.75 pp gain. On strict GPT-B regeneration, RAV exceeded Confidence by 0.50 pp at their respective frozen thresholds.

Changes: Sec. 5.6 and Appendix E.

Weakness 1:

We thank the reviewer for asking about the practical benefits of learned weighting. We added comparisons with training-free weighting methods and computational-cost analyses. The results show that RAV further improves autoregressive image-watermark detection over simple weighting methods while maintaining manageable online verification overhead.

We also analyzed statistical significance against simple weighting. For WMAR, RAV improves TPR over training-free Margin at nominal 1%/0.1% FPR by 0.94/1.11 pp on RAR-XL (95% CIs [0.65, 1.24]/[0.82, 1.42]) and 1.53/2.32 pp on TAMING ([1.19, 1.86]/[1.93, 2.71]). See Question 1 for ClusterMark significance results.

The additional WMAR online cost is approximately 18.95 ms/image. ClusterMark verification takes 8.534 ms/image with RAV versus 3.945 ms/image with Hard. Overall, the cost is manageable, with statistically significant gains in three of four generator-watermark settings.

Changes: Sec. 2, Secs. 5.2 and 5.3, and Appendix F.3.

Weakness 2:

Thank you for identifying severe rotation and targeted-forgery limitations. Under a 15-degree rotation at nominal 1% FPR, RAV increases TPR from 1.75% with Margin to 3.00%. Although absolute detection remains low, this suggests that reliability weighting may exploit some residual watermark evidence; weighting cannot replace failed symbol recovery or synchronization.

The cited studies by Müller et al. (CVPR 2025 on diffusion-watermark forgery and a 2026 study of AR-watermark removal and forgery) provide threat context, not evidence that RAV resists these attacks. We explored feasible targeted-forgery evaluations but have not completed them owing to time and computational constraints.

Changes: Sec. 2, Sec. 6, and the geometric-attack analysis in Appendix Additional WMAR Results.

Thank you for the response and the new experiments. However, the learned predictor is only slightly better than simple weighting without training, and in some settings, the gain is not significant. Therefore, I keep my score.

AuthorsOpenReview ↗

Response to Reviewer 8DpN

We thank the reviewer for the follow-up feedback. We would like to further clarify that RAV's main contribution goes beyond the incremental gain of the learned CNN. Our primary contribution is a reliability-aware, non-uniform evidence-weighting framework for autoregressive image watermark verification, which explicitly accounts for the unequal survival reliability of recovered symbols.

Compared with the original WMAR detector (NeurIPS 2025), RAV improves TPR by 2.17–3.22 percentage points (pp) at 1% FPR and 3.07–4.56 pp at 0.1% FPR across RAR-XL and TAMING. For ClusterMark (CVPR 2026), using the same frozen Cluster Predictor, RAV improves six-attack macro TPR over hard, uniform-weighted verification by 6.34 pp on RAR-XL and 6.48 pp on GPT-B at the empirical 1% FPR operating point. These results demonstrate the benefits of reliability-aware evidence aggregation across two established watermarking schemes.

Beyond this framework-level contribution, we further explore learned spatial reliability prediction. Although training-free weighting captures much of the benefit, RAV achieves an average additional gain of 1.04 pp over Margin/Confidence across four generator-watermark settings at nominal 1% FPR, with statistically significant improvements in three of the four settings.

We believe our contribution lies in both demonstrating the value of reliability-aware verification and exploring the potential of learned reliability estimation. We hope this framework can inspire further research on more effective reliability prediction and watermark evidence aggregation.

Reviewer_1nDFOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RAV framework for watermark detection in autoregressive image generation. The main idea is that recovered visual tokens are not equally reliable, and treating them equally during verification may reduce detection performance. The authors propose learning a reliability predictor using recovery-side features and using the predicted reliability to weight the watermark evidence. The method is evaluated on two watermarking schemes, WMAR and ClusterMark, across different autoregressive image generators.

Critical strengths
  • The problem is clearly motivated. I like the ideas that go beyond simply recovering watermark symbols and consider how reliable the recovered evidence actually is and what it means.

  • The method is practical for existing watermarking schemes, as it operates at the verification stage without modifying the watermark embedding or image generation process.

Critical weaknesses
  • While the proposed method generally improves detection performance compared with the original detectors, I find the additional benefit of the learned reliability predictor less convincing. For WMAR, the margin baseline already achieves competitive results, while for ClusterMark, max probability and posterior soft perform similarly to the learned predictor in several settings. This makes me question whether the additional training and complexity of the learned predictor are justified by the relatively modest improvements. I think a clearer comparison between the detection improvements and the additional training and verification cost would help demonstrate when the learned predictor is actually beneficial.

  • The authors train the predictor to estimate whether a recovered token matches the original token. However, I wonder whether exact token recovery is necessarily the best measure of reliability for watermark verification. A token may change during recovery while still preserving useful watermark evidence. I understand the intuition behind using token survival probability, but I would appreciate a clearer explanation of how this relates to the actual watermark evidence. Also, the weighting functions seem to be heuristic choices. I think an ablation of these functions would help justify the design.

Questions for authors
  • Since this work already considers recovery reliability, I think it would be interesting to discuss whether this information could help distinguish between the absence of watermark evidence and cases where there is insufficient reliable evidence to make a verification decision. I understand that this is beyond the current scope, but I think it could strengthen the broader motivation and impact of the work.
2 comments
AuthorsOpenReview ↗

Response to Reviewer 1nDF

Weakness 1: Incremental benefit and computational cost

Thank you for this important comment. We agree that a learned predictor should be evaluated not only against the original detector but also against simpler non-learned baselines, considering training, model complexity, and inference costs.

For WMAR, Margin is competitive under routine conditions, but RAV-WMAR gains grow under some geometric perturbations. With a 16-pixel translation, RAV-WMAR achieves 94.88%/85.25% TPR at frozen nominal 1%/0.1% FPR, versus Margin's 84.38%/64.00% (gains of +10.50/+21.25+10.50/+21.25 percentage points). This suggests better use of remaining evidence after the same recovery pipeline.

For ClusterMark, we compared the main RAV-ClusterMark (Unary CNN) with Max Probability and Posterior Soft:

Setting / metricUnary CNNMax Prob.Posterior Soft
RAR-XL AUROC0.99220.99010.9872
RAR-XL empirical [email protected]% FPR92.31%90.54%87.58%
GPT-B AUROC0.99300.99400.9954
GPT-B empirical TPR@1% FPR97.01%96.55%96.67%

The learned predictor exploits complementary features on RAR-XL but does not outperform simple weighting on every setting or metric: on GPT-B, its low-FPR TPR is higher but Posterior Soft has higher AUROC. Learning is therefore an optional enhancement.

We also measured computation. For WMAR, feature preparation plus five-fold CNN training/validation totals 356.81 s (RAR-XL) and 390.80 s (TAMING); the spatial CNN has 231,553 parameters and adds 18.95 ms/image to verification. For ClusterMark, Unary CNN training/validation takes 42.34 s (RAR-XL, RTX 4090) and 49.11 s (strict GPT-B, RTX 4080 SUPER), with 39,585 parameters. Full RAV versus Hard verification takes 8.534 vs. 3.945 ms/image (+4.589 ms). These are stage timings, not end-to-end adaptation; hardware and protocols differ. Training is offline and embedding is unchanged. The added computation is real, but measured training costs are relatively limited and the predictors are lightweight. Learning helps when recovery geometry and spatial context add information; otherwise, simpler methods remain reasonable. Changes: Secs. 5.2-5.4, 6; App. A.3, E.2, F.3-F.4.

Weakness 2: Token survival and weighting functions

Thank you for raising this important issue. We agree that exact token recovery and retention of watermark evidence are not equivalent: a changed token can preserve the same group or keyed detection contribution. Survival is a practical proxy: matching symbols and context preserve the detector input; mismatches need not destroy evidence. Original and recovered symbols provide automatic supervision without secret keys or greenlist labels.

We studied 800 blind ClusterMark/RAR-XL images under nine attacks. Mean centered keyed-green evidence is +0.6753 when both transition endpoints survive and -0.0151 otherwise: a difference of Δ=+0.6905\Delta=+0.6905 (image-paired 95% CI [0.6881, 0.6927]). Moreover, 23.55% of originally green transitions with mismatched endpoints remain green. Survival therefore tracks useful evidence without being identical to it.

We also agree that the weight functions need empirical justification. Holding the predictor, recovered evidence, and detector statistic fixed, we compared four weight rules per scheme using independently calibrated nominal-1%-FPR thresholds:

SchemeWeight ruleMean TPR@1% FPR
WMAREqual87.33%
WMARRaw89.39%
WMARFloor05 (RAV-WMAR)89.50%
WMARFloor1089.47%
ClusterMarkEqual88.88%
ClusterMarkMean93.64%
ClusterMarkMin94.17%
ClusterMarkProduct (RAV-ClusterMark)94.40%

The original rules rank highest among those tested, exceeding Equal by 2.17 and 5.52 percentage points, respectively. These results support our weighting choices without assuming universal optimality. We clarified the proxy's limits and added the ablations. Changes: Secs. 3.1-3.3, 5.4; App. F.2, G.2-G.3.

Question: Watermark absence vs. insufficient evidence

Thank you for this insightful suggestion. We agree that recovery reliability could help distinguish absence of detectable watermark evidence from insufficient reliable evidence to decide. RAV remains binary, so low scores cannot separate them. As a future extension, token reliabilities could yield evidence coverage or effective evidence size; a separately calibrated sufficiency threshold could support three outcomes: watermark detected; not detected with sufficient evidence; or inconclusive due to insufficient evidence. Reliability alone cannot establish watermark absence. Designing and calibrating abstention is beyond the present scope; we added this direction to Sec. 6 (Limitations).

Thank you for the detailed rebuttal. I appreciate the additional analysis. I still think the improvements over simpler weighting methods are relatively modest, but I find the idea of considering the reliability of recovered evidence during verification useful, especially since it can be applied to existing watermarking methods without changing the embedding process. Overall, the rebuttal addresses my main concerns.

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

This work aims to handle the task of verifying watermarks in autoregressive (AR) generated images. The key insight is that the recovered symbols may differ from the ones used in generation. Specifically, the recovery errors are not uniform: some symbols remain stable, whereas others are more likely to change. Existing AR image watermarking methods ignore this issue and treat the recovered verification evidence uniformly, leading to watermark verification failure. To address this issue, the authors propose Reliability-Aware Verification (RAV), which predicts symbol survival reliability from recovery-side features and uses these predictions to weight watermark evidence. By emphasizing reliable symbols and suppressing unreliable ones, RAV reduces the risk of verification failure caused by unreliable recovered evidence. The experimental results on WMAR and CLUSTERMARK show consistent improvement across three different AR image generation models, including RAR-XL, TAMING, and LlamaGen GPT-B.

Critical strengths
  1. This work is the first to investigate verification-symbol survival reliability as a detector-side signal for autoregressive image watermark verification.
  2. To solve the aforementioned issue, the authors introduce a unified Reliability-Aware Verification (RAV) framework that predicts symbol survival reliability from recovery-side information and converts the predictions into watermark-structure-aware evidence weights.
  3. The experimental results across 2 different AR image watermarking schemes (WMAR, CLUSTERMARK) and 3 different AR image generators (RAR-XL, TAMING, LlamaGen GPT-B) demonstrate the broad applicability across different watermark schemes and different model architectures.
Critical weaknesses
  1. [Robustness to geometric transformation]. Could RAV handle arbitrary geometric transformations, like rotation, translation, or crop? Since RAV reweights the watermark evidence based on the predicted symbol survival reliability, the spatial correspondence between the original and recovered token grids is not explicitly restored. Hence, I am wondering about the robustness of RAV towards stronger geometric transformations. Could the authors provide some extra experimental results on this issue? If not, please provide a detailed discussion of this issue.
  2. [Reliability Predictor]. Does the reliability predictor have to be retrained for every generator and tokenizer? If so, how much paired training data and computation are required for adaptation to different models?
Questions for authors
  1. [Robustness to geometric transformation].
  2. [Reliability Predictor].

Please refer to the details in the critical weaknesses part.

2 comments

Weakness 1. Robustness to geometric transformations.

We thank the reviewer for pointing out the impact of spatial correspondence under geometric transformations on watermark detection. We agree that rotation, translation, and cropping may disrupt correspondence between the original and recovered token grids. Our experiments show that RAV behaves differently across geometric transformations: it can better use surviving watermark evidence, but cannot itself restore severely lost alignment.

(1) Results under geometric transformations. Rotation and cropping results were already reported in Sec. 5.2 and Appendix E.2. To address translation specifically, we additionally applied rightward shifts of 0, 4, 8, and 16 pixels to the original 800 frozen WMAR/RAR-XL blind-test images, using black padding without interpolation. The synchronization and retokenization pipeline, predictor, and detection thresholds were kept unchanged. TPR (%) at frozen nominal FPR operating points is shown below:

Shift (px)Margin @1%RAV @1%Margin @0.1%RAV @0.1%
099.6399.7599.5099.63
491.8893.5082.3886.63
899.3899.5098.2599.38
1684.3894.8864.0085.25

At 16 pixels, RAV improves over Margin by 10.50/21.25 percentage points (pp) at nominal 1%/0.1% FPR; image-paired bootstrap 95% CIs are [8.38, 12.75]/[18.38, 24.13] pp. The 0-pixel control reproduced the original results. Performance is non-monotonic: 4 pixels was more challenging than 8 pixels, with 46/800 synchronization fallbacks at 4 pixels. At 16 pixels, observed wrong-key FPRs were 1.148%/0.127%, differing slightly from the nominal operating points.

(2) Limits under strong geometric transformations. RAV reweights evidence from the existing synchronization/recovery pipeline; it does not reconstruct spatial correspondence. Under 15-degree rotation, TPR at nominal 1% FPR was only 3.00% (Margin: 1.75%). Thus, RAV complements synchronization and cannot resolve severe transformations that defeat the underlying recovery pipeline.

Changes: Sec. 5.2; Appendix E.2 (geometric-severity table).

Weakness 2. Generator/tokenizer adaptation and cost.

We thank the reviewer for the valuable suggestions regarding adaptation of reliability predictors across generators and tokenizers. We agree that training requirements and transferability matter for practical deployment. In our current evaluation, a separate predictor is trained for each generator/tokenizer configuration, but we also evaluated whether predictor retraining can be avoided.

(1) Training data and measured costs. For WMAR/RAR-XL and WMAR/TAMING, each configuration uses 1,000 development image IDs across nine conditions (9,000 recovery records) with five-fold image-grouped training. Feature preparation took 189.36/225.04 s and CNN training/validation 167.45/165.76 s, respectively, giving measured subtotals of 356.81/390.80 s (RTX 4090). ClusterMark/RAR-XL used 800/200/200 training/selection/calibration image IDs; 12 epochs took 42.34 s (RTX 4090). Formal ClusterMark/GPT-B used 800/100/100 IDs; training and validation took 49.11 s across 16 epochs (best epoch 8), and the measured full process took 57.00 s (RTX 4080 SUPER). These timings exclude image generation, recovery-backbone training, and full detector evaluation. WMAR uses five folds whereas ClusterMark trains one model, so the timings are not matched end-to-end comparisons.

(2) Bidirectional zero-shot transfer. We transferred frozen ClusterMark reliability CNNs between RAR-XL and GPT-B, retaining the source feature construction, normalization, Platt calibration, and weights without target-domain predictor retraining. For RAR-XL to GPT-B, macro TPR@1% FPR over eight atomic attacks was 97.00%, versus 97.36% for the native predictor: -0.36 pp (5,000 image-paired bootstrap resamples; 95% CI [-0.65, +0.11] pp), with no statistically significant difference detected. For GPT-B to RAR-XL, macro TPR@1% over nine attacks was 93.94%, versus 95.38% natively: -1.43 pp (95% CI [-2.00, -0.88] pp), a significant loss. Thus, direct reuse is feasible in the tested configurations but is direction-dependent. The attack sets differ across directions, and both generator and tokenizer change. The reported macro TPRs use attack-wise empirical ROC points; separate frozen-threshold tests use independent target-domain negatives for calibration. No predictors were retrained, and target test cohorts had been previously explored.

Changes: Secs. 3.1 and 5.4; Appendix A.3 (adaptation data, timing, and transfer results).

Thanks for the detailed response from the authors. My concerns about geometric transformation and reliability predictor's training cost have been addressed. Hence, I am keeping my positive rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Reproducibility Materials

Dear Area Chair and Reviewers,

Thank you for your careful reviews and constructive suggestions. We have revised the manuscript and expanded the supplementary material in response to the reviews and subsequent discussions. The principal revisions and newly added experimental analyses are highlighted in blue in the revised PDF. Our detailed point-by-point responses have also been posted. The main changes are summarized below.

1. Stronger baselines and statistical evidence (Secs. 5.2–5.3; App. F.3). We added image-paired bootstrap comparisons against training-free weighting. At frozen nominal 1%/0.1% FPR, RAV improves WMAR over Margin by 0.94/1.11 percentage points (pp) on RAR-XL and 1.53/2.32 pp on TAMING, with positive 95% confidence intervals. For ClusterMark, the gain over Confidence is 1.22 pp (95% CI [0.35, 2.00]) on RAR-XL, whereas 0.46 pp (95% CI [-0.23, 0.70]) on GPT-B is not statistically significant. We discuss conditions where simple confidence weighting is competitive.

2. Mechanism, ablations, and protocol corrections (Secs. 3–5; Apps. F–G). A new ClusterMark analysis links transition survival to useful keyed-green evidence without equating them: the mean centered-evidence difference is +0.6905 (paired 95% CI [0.6881, 0.6927]); 23.55% of originally green transitions with mismatched endpoints remain green. We added fixed-predictor weight-function ablations for WMAR and ClusterMark and pointwise/pairwise comparisons. We corrected the GPT-B permutation control: in the paired audit, Unary RAV versus corrected Permuted is 0.9701 versus 0.8861, a drop of 8.40 pp (95% CI [6.48, 10.52]). We distinguish empirical nearest-1%-FPR ROC, non-exceeding-FPR readouts, and independently frozen thresholds.

3. Geometric robustness and unseen attacks (Sec. 5.2; Apps. B, E). We added 0/4/8/16-pixel horizontal translations to the frozen WMAR/RAR-XL evaluation. At 16 pixels, RAV reaches 94.88%/85.25% TPR at nominal 1%/0.1% FPR, compared with Margin's 84.38%/64.00%. We also clarify that weighting cannot restore spatial alignment when severe rotation defeats recovery/synchronization. Additional held-out neural-compression and regeneration evaluations include training-provenance checks; targeted forgery is identified as an unevaluated threat rather than a demonstrated defense.

4. Cross-source FPR and text-to-image generalization (Secs. 5.5–5.6; Apps. C–D). With original keys and frozen thresholds, we tested 1,000 unwatermarked GPT-B images and 1,000 COCO photographs. RAV's observed FPR on GPT-B/COCO is 1.00%/1.50% for WMAR-RAR-XL and 1.30%/0.50% for ClusterMark-RAR-XL, showing source-dependent calibration shift. A new Janus-Pro-1B text-to-image experiment uses a WMAR-style watermark with disjoint splits: macro TPR is 94.00% (RAV), 93.43% (Margin), 89.90% (Equal); the gain over Margin is not statistically significant.

5. Adaptation, efficiency, and scope (Sec. 6; App. A.3). We now detail generator/tokenizer-specific splits, automatically labeled key-independent survival, offline training costs, and detection overhead. WMAR development-stage partial subtotals are approximately 357/391 s (RAR-XL/TAMING); ClusterMark unary-CNN fitting takes 42/49 s (RAR-XL/strict GPT-B), under differing protocols and hardware. WMAR adds 18.95 ms/image; ClusterMark RAV versus Hard takes 8.534/3.945 ms/image. Direct frozen-predictor transfer is feasible in the tested settings but direction-dependent, and target-domain threshold calibration remains necessary. We also clarify negative evidence versus insufficient reliable evidence as a future direction.

6. Reproducibility materials. We provide an updated complete supplementary archive combining original and new experiments, including source code, protocols, machine-readable results, calibration records, audit manifests, and frozen RAV reliability-predictor checkpoints for the principal WMAR, ClusterMark, strict GPT-B, and Janus evaluations. Checkpoint hashes and reproduction guidance are included. Upstream generator/tokenizer/recovery checkpoints and raw image corpora are obtained from the cited projects rather than redistributed.

We hope these revisions clarify RAV's evidence, scope, limitations, and reproducibility. Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.