acceptodds
Withdrawn from ICLR 2027

Watermark Any Image: A Unified Framework

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

25 views

Abstract

Embedding data in images while maintaining visual quality and ensuring reliable detection and decoding remains a challenge in image watermarking. We introduce UniMark, a unified watermark framework for zero-bit detection, multi-bit detection, and multi-bit decoding that operates directly on existing images without accessing or modifying the generator. UniMark calibrates projection scores in reconstructable image representations using unwatermarked distributions and modifies these scores to embed watermark information. We develop statistical watermark constructions and corresponding likelihood-ratio tests for detection and decoding. These tests are optimal under independence, while finite-sample analysis demonstrates near-optimal performance. Experiments on three real-image datasets and images from four generative models demonstrate high image fidelity at near-perfect detection rates across diverse image sources. UniMark supports payloads of up to 32,768 bits, twice the largest payload of the existing largest payload watermark method, and achieves a 92% full-message recovery rate at 8,192 bits under clean conditions, which existing methods have not demonstrated. Our anonymous code repository is available at [this anonymous repository](https://anonymous.4open.science/r/WatermarkyAnyImage-UniMark-58C5).

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Image Watermarking, Statistical Watermarking, Likelihood Ratio Test, Multi-bit Watermarking, High-Capacity Watermarking

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_d9tmOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes UniMark, a unified framework for watermark detection and message decoding in existing images without generator access. It combines projection statistics, quantile calibration, and Binary or Gaussian watermark constructions. Although the reported image fidelity, processing speed, and high-capacity results are promising, the contribution of the central components is not sufficiently established through controlled experiments. The near-optimality claims are broader than the supporting analysis, and reliable message recovery remains limited under common image transformations. These concerns leave the main contributions insufficiently supported.

Critical strengths
  • Table 1 reports substantially higher PSNR at approximately 64-bit payloads while maintaining nearly perfect detection and decoding. Table 3 reports 92% complete recovery for 8,192-bit messages. These constitute meaningful empirical strengths.
  • Calibration, threshold selection, and auditing use separate data, and decoding certification avoids treating all bits within an image as independent samples. Tables 9 and 10 also provide empirical evidence beyond the main DCT implementation.
Critical weaknesses
  • The abstract states that “finite-sample analysis demonstrates near-optimal performance” lacks clear qualifications. B.3 only certifies detection rate and bit-accuracy gaps under evaluated conditions, not optimality in capacity, fidelity, or robustness. Its generic error-based bounds do not quantify the loss from ignoring dependence, and the multi-bit certification only covers payloads up to 1,024 bits.
  • Quantile calibration and key-dependent orthogonal projection are presented as important elements of the framework, but the paper does not isolate their individual contributions. The experiments primarily compare the complete system with other methods and demonstrate applicability across representations. These results support the effectiveness of the overall system, but do not establish whether the gains arise from the proposed calibration and projection designs rather than frequency selection, embedding strength, or the number of projections assigned to each message bit.
  • Section 2.1 permits approximately invertible representations, while the subsequent description uniformly uses “inverse transform” and xwm=T−1(zwm), without clearly distinguishing exact inversion from approximate decoding. Satisfying the target projections in representation space through Eq. 7 does not automatically guarantee that those projections are preserved after image reconstruction and re-encoding.
  • Figure 2(a) matches clean accuracy rather than robustness, so higher PSNR alone does not establish a better overall trade-off. The claim of highest PSNR and SSIM is not directly supported by Table 3 because those metrics are unavailable for MaxMark. Table 14 also includes identical Clean and Resize 1.00 results in its average; an average excluding reference conditions would be more informative.
Questions for authors
  1. What is the main technical advance over the closest methods? Do experiments show how much calibration and orthogonal projection each contribute?
  2. Dose the false-positive rate remain within the target range when unwatermarked images undergo the same compression and resizing?
Reviewer_C8y8OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

UniMark modifies calibrated projection scores in an image representation and reconstructs the watermarked image. It provides binary and Gaussian constructions for detection and decoding, with experiments on image quality, payload size, runtime, and image transformations.

Critical strengths

Watermarking existing images without access to the generator is useful. Reporting complete-message recovery alongside bit accuracy helps interpret the high-capacity results, although these strengths do not resolve the concerns below.

Critical weaknesses
  1. The scope of the claimed unification is not clear enough. Using the same projection-and-calibration pipeline for several tasks does not establish reliable performance across arbitrary image distributions. Appendix C.2 estimates both projection distributions and detection thresholds from a mixture of four generated-image sources. The paper needs to explain what remains valid when the deployment distribution differs from that mixture. Generator-independent embedding should not be conflated with distribution-independent calibration. Testing several sources is useful, but does not resolve this distinction.

  2. The theoretical and empirical evidence leave the main advantage uncertain. The near-optimality argument largely certifies that a detector with a small measured error rate has little room to improve; it does not explain when the proposed construction controls dependence or achieves a favorable distortion-robustness trade-off. This limits the conceptual insight of the theoretical contribution. The distinction matters in practice: Table 11 reports only 1.56% complete-message recovery after JPEG Q80 for the 64-bit method, compared with 99.73% for TrustMark. Strong clean-image performance and high watermark-presence detection do not compensate for losing the identifier under routine compression.

Overall, I recommend rejection. The claimed breadth and theoretical contribution are not sufficiently established, and the practical message-recovery limitation is consequential.

Questions for authors

Which assumptions are required for calibration to transfer to a new image source? What are the source-specific false-positive rates with fixed calibration? At matched JPEG message-recovery performance, does the image-quality advantage remain? Please distinguish existing evidence from proposed future experiments.

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

This paper proposes UniMark, a training-free post-hoc watermarking method that modifies calibrated projection scores in reconstructable image representations. A shared statistical construction supports watermark detection and message decoding. Experiments report improvements in clean-image fidelity, payload capacity, and processing speed across several image sources.

Critical strengths

The method reports useful improvements in image fidelity, clean-image payload capacity, and processing speed without watermark-specific training.

Critical weaknesses
  1. The research gap and the source of the claimed performance advantage are unclear. A unified statistical formulation does not by itself establish an improvement over existing transform-domain watermarking. The evaluation does not sufficiently isolate the benefit of calibrated projection remapping from frequency selection and embedding strength. Comparisons at matched payload, image quality, and robustness are needed to establish the practical contribution.
  2. Robustness to common distortions is a substantial limitation. In Table 11, JPEG Q80 reduces 64-bit complete-message recovery to 1.56%, compared with 99.73% for TrustMark's 61-bit payload. In Table 14, resizing to 95% and restoring the original resolution reduces complete-message recovery to 0% even at 8,192 bits. Continued watermark detection does not resolve the loss of the embedded message. These results limit the practical value of the reported capacity gains.
Questions for authors

Please address the above two concerns.

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

Related papers

Discussion (0)

Sign in to comment.