acceptodds
Under review as a conference paper at ICLR 2027

From Fixed JND to Learned Attenuation in Invisible Image Watermarking

Abstract

In invisible image watermarking, the attenuation map sets the per-pixel strength of the watermark residual. We ask whether replacing fixed just-noticeable-difference (JND) attenuation with a learned attenuation map improves the quality-robustness trade-off, and whether any robustness gain survives when the embedded distortion, rather than the embedding coefficient, is held fixed. Freezing Pixel Seal's embedder and extractor, we train a U-Net that predicts the attenuation map. At the same coefficient, the base learned map raises bit accuracy under a combined JPEG-crop-brightness attack by 3.81 percentage points (pp) but lowers PSNR by 5.37 dB; adding an LPIPS loss and Shape-Aligned Attenuation (SAA), which aligns the mean-normalized learned and JND maps, recovers 2.05 dB at a 0.31 pp accuracy cost. At matched PSNR, however, this advantage disappears: fixed JND takes a small but significant lead (0.36-0.38 pp over the best learned condition at all but the highest target), and among the learned conditions the robustness differences come from SAA (+0.55-2.59 pp, with or without the LPIPS loss). The learned conditions' lower LPIPS at matched PSNR also largely vanishes at matched robustness, where fixed JND needs the least distortion. In exploratory tests on MaskWM, WAM, and VideoSeal, matching PSNR likewise narrows the JPEG-attack gap to under 1.2 pp (learned attenuation stays slightly ahead on VideoSeal). With Pixel Seal's backbone frozen, learned attenuation therefore improves on fixed JND neither in robustness at equal distortion nor clearly in quality at equal robustness, and single-coefficient comparisons overstate its benefit.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image watermarking, Invisible watermarking, Learnable attenuation, JND

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

This paper replaces fixed JND attenuation with a learned network and studies the effects of perceptual and shape-alignment regularization. Its main finding is that robustness gains at a fixed embedding coefficient largely disappear when PSNR is matched. The controlled evaluation is useful, but the methodological novelty and explanation of the negative result remain limited.

Critical strengths
  1. The frozen-backbone setup and factorial ablation provide a clear comparison of attenuation strategies and regularization choices.

  2. The matched-PSNR evaluation, three training seeds, and paired bootstrap analysis strengthen the empirical findings. The paper also reports unfavorable results and experimental limitations transparently.

Critical weaknesses
  1. Limited additional insight. The attenuation network and normalized shape-matching loss offer modest methodological novelty. Since a fixed embedding coefficient does not ensure equal distortion, the paper’s contribution depends on explaining the observed ranking reversal. The current analysis provides limited guidance on why JND remains competitive or when learned attenuation would be useful.

  2. The key scalar control is not compared at matched distortion. Appendix B.1 includes an image-dependent scalar multiplying the native JND map, but evaluates it only at a fixed coefficient. Consequently, the experiments do not adequately establish whether dense spatial prediction adds value beyond adjusting the overall strength for each image. Matching average PSNR with one global coefficient does not remove this ambiguity.

  3. The source of the negative result remains unclear. The frozen backbone was trained with JND, whereas the new attenuation network is trained from scratch at one embedding strength and rescaled afterward. This leaves unresolved whether the gap arises from the attenuation representation, optimization, or compatibility with the pretrained backbone.

  4. The decisive comparisons have limited coverage. The main matched-PSNR analysis emphasizes Combined attacks, while the equal-robustness perceptual comparison relies on one interpolated operating point. These results provide limited support for broader conclusions about the practical quality–robustness trade-off.

Questions for authors
  1. How does the scalar-JND baseline in Appendix B.1 compare with dense attenuation under the main matched-PSNR protocol and comparable training seeds?

  2. Would a JND-initialized correction, or training directly for a representative distortion budget, change the matched-PSNR ranking?

  3. Does the ranking reversal persist across the other evaluated attack groups? Can the equal-robustness comparison be verified at several directly measured operating points?

Reviewer_zww4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper substitutes Pixel Seal's fixed JND attenuation with a learned per-pixel module (a compact U-Net, together with an optional LPIPS loss and Shape-Aligned Attenuation). Under a fixed coefficient, the learned attenuation yields a +3.81pp accuracy improvement but incurs a 5.37dB PSNR cost; once PSNR is matched, this gain vanishes and fixed JND comes out slightly ahead. The contribution is methodological: comparisons at a single coefficient overstate the benefit of learned attenuation.

Critical strengths

The paper isolates precisely one component (the attenuation map) on a frozen backbone, assesses five conditions in a 2×2 factorial design with three seeds, applies image-level paired bootstrap significance tests, and demonstrates that the headline gain (+3.81pp) at fixed α is reversed once PSNR is matched.

Critical weaknesses

1.The scope is narrow and the technical novelty is limited. The method itself is a small U-Net that predicts a per-pixel scalar map plus two fairly standard regularizers (LPIPS and an MSE-to-JND shape loss). The architecture and losses offer little novelty. Relatedly, the idea of learned attenuation (adaptive per-pixel/adaptive strength embedding) has considerable prior art (e.g., Bagheri et al. 2020, Wang et al. 2023 Adaptor), which the paper acknowledges but does not compare against experimentally. 2.The key comparisons rest on a narrow empirical base. (a) The headline matched-PSNR reversal on Pixel Seal depends on three seeds and a 5-point α grid, and the margins are small (0.36–0.38pp); it is unclear how sensitive the conclusion is to the grid or to the choice of Combined attack. (b) The cross-backbone evidence is single-seed, uses a reduced 36-attack common subset, and reports matched-PSNR results only for JPEG attacks, so the central "matched-distortion" claim is fully validated on only one backbone. (c) There is no human perceptual study; quality is judged solely by PSNR/SSIM/LPIPS, and the paper itself notes metric disagreement (SAA improves PSNR/LPIPS but lowers SSIM).

Questions for authors

1.On Pixel Seal, how sensitive is the matched-PSNR ranking to the number/density of α grid points and to interpolation? Would a denser calibration or a different summary (e.g., area under the PSNR–accuracy curve) alter the conclusion that "NATIVE leads at all but the highest target"? 2.Why is the Combined attack used for Pixel Seal but JPEG-only for the cross-backbone matched-PSNR analysis (because Combined is "uninformative for MaskWM")? Could you report a common robust metric for which the matched-PSNR conclusion holds consistently across backbones? 3.You report a ~2.1× embedding-time cost for the learned module. Given that at matched PSNR/robustness it offers no clear benefit, is there any operating regime (e.g., extreme downscaling, resize-only, targeted/localized watermarking à la WAM) where learned attenuation demonstrably wins? 4.The four main-text hyperparameters were chosen from single-seed sweeps on COCO. Could you confirm that the matched-PSNR conclusion (not just Table 1's fixed-α trends) is stable under at least one alternative λ choice? 5.Since gradients flow through the frozen extractor, is the learned attenuation merely re-tuning the effective embedding strength for the extractor it was trained with? Would results change if the extractor were fine-tuned jointly, or is that out of scope by design?

Reviewer_9wRKOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes to evaluate the robust-quality trade-off of modern post-hoc image watermarking when the fixed psycho-visual mask is replaced with a learned module. The authors's training method can be summarized simply: take PixelSeal recipe, freeze the embedder and extractor, add a U-Net to learn an attenuation map and train jointly PixelSeal's discrimnator from scratch. The main contribution is the specific loss of the attenuation map: Shape-Aligned Attenuation (SAA). The idea is to make the "structure" of the learned module as close as possible to the standard JND, but allow freedom in how the energy is distributed within this structure. The authors then validate their approach through an ablation study. Notably they study the impact of imposing a structure on the learned module through their SAA versus using only an LPIPS loss as a guide.

Critical strengths

I approve of the main objective of the paper: studying the impact and limitations of using a learned attenuation mask is a problem relevant to the practice, the JND mask being rarely questioned.

Critical weaknesses

Limited claims and limited validation scope Because the authors focused on training a module matching the JND structure, the resulting method is quite weak in terms of results and novelty. The authors are left with studying the quality-robustness trade-off of distributing the energy differently over the JND mask. A more interesting work would have either 1) tried to find a better method of designing a training recipe for an attenuation map when finding out that an LPIPS loss is not sufficient by itself and/or 2) analyzed what properties made the JND work to inform the design of a better attenuation map. In any case, the experimental results are performed for a very specific set of hyper-parameters (something which the authors state directly) and are very underwhelming, with performance barely varying in terms of bit-accuracy between attenuation maps and a pretty clear PSNR/LPIPS trade-off that does not show any method to outperform the other. Worse, the experiment presented in Figure 2 basically demonstrates that there is no reason to choose the author's approach instead of the JND (especially given the computational cost).

AI writing makes the paper unclear and unnecessarily detailed on unimportant matters The paper, as all AI-written paper nowadays, looks good at first sight; a focused reading shows that a lot of attention is given to useless details (line 362 dares to explain to me that the line between two points in a visual graph is a linear interpolation...), so much so that I believe this paper could lose one or two pages if all the fat was removed, for example Section 3.2 re-explains basic training objectives that are not contributions of the paper (LPIPS loss and cross-entropy loss on the estimated bits). On the other hand, some important parts are never described, for example the "patch discriminator" (which I guess is simply taken from PixelSeal) is not described in the main paper nor in Appendix A (where the authors state that it actually is!).

I would like to point out to the authors the impact of outsourcing most of the paper to AI agents (which they directly claim on line 488-489): the paper might look good from a distance but it clearly misses the nuances of the subjects which should nevertheless be known by researcher working in watermarking. "Funnily", the AI rambling often highlights the lack of rigor of the paper itself: line 366 "this is a single interpolated operating point, not a full equal-robustness curve.", the Limitation and future work paragraph, line 421-423 "This is exploratory evidence [...], not evidence of a backbone-independent improvement or a ranking of methods,". I think it is pretty telling that such sentences are still present in a paper: a rigorous researcher would have judged some of these remarks to be irrelevant and removed them or addressed the exposed problems when serious (and most of them are!). A sneaky researcher would have hidden these faults. But when a paper is clearly not written by hand and not proof-read correctly, this is what we get.

Questions for authors

Due to the weaknesses I highlighted, it is clear to me that this paper is not at the level of quality expected for an A* conference. I remind the authors that ICLR favors 1) substantial conceptual novelty 2) technical and experimental rigor and replicability and 3) irreproachable clarity and structure.

Obviously the paper does not propose anything conceptually novel: it is a purely empirical evaluation of a practical question in modern post-hoc image watermarking. The experimental rigor is lacking: the "positive" results seem cherry-picked (the Combined column in Table 1), others are not correctly validated (see questions). In any case, the results themselves are very underwhelming: variations between methods are at best around a few percent, but most often a lot smaller. The overall conclusion of Figure 4 is that there is a PSNR/bit-accuracy trade-off with the attenuation map, something which is extremely obvious. Finally, regarding clarity, I point to the long discussion in the weakness section.

Given all my gripes, I advise a strong rejection of the paper. If the authors really want to publish this work, I would advise lower rank conference where incremental, fully empirical work such as this one would be a better fit.

Some work can be done that would slightly improve my score -- up to a weak reject:

  • The authors show that purely maximizing LPIPS is not enough to learn a valuable attenuation map. They choose to force the JND structure to fix this problem. This looks to me quite a weak fix when the goal was to see the gain available when foregoing the JND itself. It is thus somewhat evident that they won't truly beat a JND based scheme with this approach. It would have been more interesting to understand what properties make the JND works in watermarking and isolate these properties as loss functions. This would have allowed to see if the JND structure can be improved upon, how and why. Can the authors propose such JND-free loss function? Or at least provide properties of interest.
  • As is, the qualitative analysis of the attenuation map is pure nonsense. Figure 3 is pretty good to understand the difference between a pure LPIPS learned mask vs a JND mask. But how do the author link this qualitative analysis with the quantitative impact on robustness and quality? Quality might be easily explainable through the well-known masking effect around edges (but the authors never speak of it), but what about robustness? On the "AI rambling" side, I would enjoin the authors to explain to me, in human language, what lines 370-375 mean. In particular, why "multipli[ying] the attenuation map’s high-frequency DCT magnitude fraction by 3.8–4.2× and its mean Sobel gradient magnitude by 2.6–2.7×" allows to "pulls that response back toward edges and textures". I understand what the authors want to say in essence (the mask concentrate on high frequencies and it matches with an edge filter), but they never introduce either of these metrics, nor discuss their relevance to the attenuation map properties.
  • WAM and VideoSeal are basically the same backbone (this is even reflected in the code if the authors are inclined to look: when the video flag is set to false, VideoSeal falls back to WAM). Can the authors then explain what differs between the two in their setting? In particular, is WAM used in its "localizable" mode or without any mask? The authors should also state what version of VideoSeal they use in the main paper (they do in the Appendix but it is important given that the baseline version is 96-bit whereas the newest 256bit).
  • Substantial (human) proofreading is required for any improvement of the score. For example "At matched PSNR, the same 2 × 2 ablation separates cleanly." does not mean anything, same for "at 43–46.04 dB every seed-level paired interval against either SAA condition excludes zero". Overall, they should pay attention to nonsensical jargon and make it so the experiment discussion is not simply a long list of number followed by a dubious conclusion -- the Comparison at matched robustness. paragraph at line 361 is a good example of what not to do.
  • For the matched PSNR comparison, the PSNR is matched on average over the watermarked images. This means that the signal is stronger on some images and weaker on others, which actually does not allow for a fair evaluation. In any case, the experiments shows that when matching the PSNR on average, JND always dominates the author's approach: what's the point of the method then?
  • Watermarking schemes based on neural networks have an unintuitive property: they are trained at a fixed average PSNR, hence forcing a different yet fixed PSNR on all images usually degrades the robustness substantially. The consequence is that Figure 4. is misleading in my opinion: if the authors were to fix the PSNR for all images, they would see that even decreasing it does not improve on the bit-accuracy. Consequently, It might be the case that the lower bit-accuracy they observe at a higher PSNR is not due to the impact of their attenuation map but only due to the extractor not being trained at this PSNR level. Could the authors test the low PSNR case? And could the authors demonstrate that the attenuation map is what actually changes the robustness (such as is Figure 2)?
1 comment

I note the absence of response to my review, given the remaining timeframe, I will maintain my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.