acceptodds
Under review as a conference paper at ICLR 2027

Optimizing and Securing the Modern Watermarking Channel for Images

Abstract

To comply with recent regulations requiring traceable generated content, modern watermarking has adopted multi-bit post-hoc watermarking schemes. These modern designs rest on an encoder-decoder pair implemented as deep neural networks. These models are usually treated as pure black-boxes trained end-to-end, with the noise of the watermarking channel modeled through a fixed set of geometric and valuemetric transforms applied to watermarked images. We argue that this purely empirical approach leads to unquestioned design flaws and a lack of theoretical performance guarantees. This work proposes a general theoretical model of modern post-hoc watermarking schemes grounded in a statistical analysis of the outputs of the encoder/decoder pair. We show that these deep neural networks implicitly define a watermarking channel modeled as parallel AWGN channels, with messages transmitted using BPSK modulation. This imposes a binary alphabet, greatly limiting the capacity of these watermarking systems. Another fatal flaw is their lack of a secret key, making them intrinsically insecure. We make this notion of watermarking security precise for post-hoc schemes by linking it to the possibility of estimating the secret key under a given statistical model of the decoder's output. By putting together the results from this theoretical analysis, we introduce SNW: a novel post-hoc watermarking system that significantly outperforms existing state-of-the-art baselines in terms of capacity while also providing strong security guarantees. Notably, it does not depend on a fixed codebook or binary alphabet, allowing it to reach a rate close to Shannon capacity through the use of capacity-achieving error-correcting codes.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Watermarking, Image, Deep Neural Network

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_uF9dOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents an information-theoretic analysis of modern neural watermarking systems and proposes Secure Neural Watermarking (SNW). By learning isotropic watermark representations and incorporating a secret orthogonal projection, SNW aims to improve watermark capacity while providing security against PCA-based key estimation attacks. Experiments demonstrate substantially higher estimated capacity than several existing watermarking methods.

Critical strengths

The paper investigates the communication channel and security properties of neural watermarking systems, providing useful insights beyond conventional architecture-driven improvements.

Critical weaknesses

The claimed perfect security is restricted to PCA-based key estimation. Isotropic covariance does not imply statistical independence or general key secrecy, and higher-order attacks such as ICA may still recover information about the secret projection. Moreover, Proposition 1 uses the capacity formula for a hard-decision binary symmetric channel rather than the soft-output BPSK-AWGN channel. These issues limit the strength of the theoretical claims.

SNW uses a 768-dimensional codeword, compared with 256 bits for PixelSeal and VideoSeal. Its absolute capacity advantage therefore does not necessarily imply better per-symbol efficiency. Indeed, Table 7 shows lower normalized capacity than PixelSeal under JPEG Q50, rotation, and cropping. The evaluation also omits the closely related high-capacity ChunkySeal method. Furthermore, the claimed capacity is estimated theoretically without demonstrating reliable message recovery using actual error-correcting codes.

The watermark capacity of the proposed method is not particularly impressive, considering that the input image resolution is 1024 × 1024. Existing methods, such as MBRS (ACM MM 2021) and FIN (AAAI 2023), have demonstrated the ability to embed dozens of bits in much smaller images. For example, if a method can embed 128 bits in a 256 × 256 image, scaling the payload proportionally to the number of pixels would correspond to 2,048 bits at 1024 × 1024 resolution. Although this scaling is not necessarily achievable in practice, it suggests that the reported capacity of 768 bits may not be particularly high. Therefore, the authors should compare their method with existing approaches under the same image resolution and payload length, while reporting both watermarked image quality (e.g., PSNR) and robustness.

Questions for authors

See weaknesses

4 comments

We thank the reviewer for pointing out that the channel analysis provides useful improvements in the construction of a watermarking scheme. We address the concerns below.

The claimed perfect security is restricted to PCA-based key estimation. Isotropic covariance does not imply statistical independence or general key secrecy, and higher-order attacks such as ICA may still recover information about the secret projection.

The paper is very clear about that fact, and we don't see why it should be considered a critical weakness given that we are the first to propose:

  1. A formal notion of security for modern watermarking systems.
  2. A general method for securing such systems against a large class of attacks, namely PCA attacks.

Secondly, the burden of proof that "ICA may still recover information about the secret projection." does not lie with us since we never made any claim about it. We would appreciate the reviewer to exhibit how such an attack would be written down and performed, otherwise, one could always find a "possible" attack for any watermarking system by claiming such an attack "could" exist.

In any case, we are highly doubtful that mounting such an attack is trivial because:

  • Our model of the noise in Eq. (7) is isotropic
  • The isotropy that we impose, during training, on how the steering vectors vv are distributed, means that if Eve has access to independent watermarked samples, the "sources" in the ICA setting would be isotropic (or close to it)

If both of these things are true, then ICA is impossible (because either independence is broken or the sources are Gaussian due to the isotropy). Nevertheless, we don't rule out the possibility of mounting such an attack, for example by imposing some structure on the steering vectors (breaking the isotropy of the samples) or providing more information to Eve. But we hope that the reviewer sees that describing the conditions under which ICA is possible or not, and how to perform it, is completely different work that might require another paper.

Moreover, Proposition 1 uses the capacity formula for a hard-decision binary symmetric channel rather than the soft-output BPSK-AWGN channel. These issues limit the strength of the theoretical claims.

Again, it is unclear to us why this is a "critical weakness". We chose the hard-decision BSC because it is the actual channel used by every current post-hoc modern watermarking system, and as such accurately model their limits. Of course, we could have used a soft-decision BSC for SNW, or better yet, the continuous alphabet model which SNW readily accomodates. However, this would have been unfair to VideoSeal and PixelSeal, and one could have critiqued our evaluation on the ground that we compare the capacity of systems with different channel models. Consequently, we respectfully disagree that this "limits the strength of the theoretical claims", it is a conscious decision made to allow both an accurate model of the current art as well as a fair comparison with SNW (which could have boasted larger capacity simply by using a continuous alphabet on the AWGN channel).

SNW uses a 768-dimensional codeword, compared with 256 bits for PixelSeal and VideoSeal. Its absolute capacity advantage therefore does not necessarily imply better per-symbol efficiency. Indeed, Table 7 shows lower normalized capacity than PixelSeal under JPEG Q50, rotation, and cropping. The evaluation also omits the closely related high-capacity ChunkySeal method. [...]

We respectfully disagree that this is a valid criticism. Two major points in the paper are:

  1. It is highly non-trivial to "add more bits" for a fixed backbone. One cannot change the payload size of modern watermarking with the same ease as what was possible with classical schemes: the DNNs need to be fully retrained. We would also like to point out that, contrary to ChunkySeal, we don't "add more bits" simply by making the DNN larger, a point which was raised by the reviewer themselves. One important lesson of the paper is that we can "add more bits" by simply using the watermarking space more efficiently.
  2. Bit-accuracy is meaningless by itself. As we already state in the paper (line 448), even for PixelSeal which boasts a bit-accuracy of 0.997 the probability of decoding error is still 0.53! Maybe one of the most important message we wanted to deliver in the paper is that Shannon capacity IS the better metric since it allows to measure the expected number of transmitted bits while still being agnostic to implementation details (choice of error-correcting code, block-length, etc...) and valid for any strictly positive probability of decoding error. Consequently, we argue that higher capacity still means that bits are also "stronger" overall.

We chose not to include ChunkySeal in the experiments because it is based on exactly the same design as PixelSeal but with a larger backbone (2048 vs. 768). The whole point of the paper was to show that current approaches don't use the latent space efficiently and that one can reach far better performance without ever touching the architecture.

As a sign of good faith, however, here are the results requested by the reviewer:

AttackCodeword bit-accuracyCapacity (bits)
Identity0.87445
Contrast_2.00.85394
Brightness_0.20.87441
CenterCrop_0.90.86421
CenterCrop_0.80.85403
CenterCrop_0.70.84374
CenterCrop_0.60.82334
CenterCrop_0.50.75230
JPEG_50_augly0.85395
JPEG_80_augly0.86430
GaussianBlur_3_3_1.0_1.00.87445
Rotate_900.86419
Rotate_450.51
Resize_0.50.87445
HorizontalFlip0.87441
MedianFilter_3_30.87445
Saturation_1.50.87442

As the reviewer can see, ChunkySeal consistently underperforms compared to SNW! To be perfectly transparent, we were extremely surprised by these results. This lack of performance can be linked to the whitening operation that we perform to make individual bits independent -- see Appendix F. Before this operation, the bit accuracy of ChunkySeal is indeed close to 100%, after it, it drops close to 85%! This means that baseline ChunkySeal is heavily biased towards certain messages. Even worse, if one were to use its baseline bit-accuracy to compute its actual capacity, one would overestimate it significantly due to the dependence and bias between the message bits. To be clear, this does not mean that ChunkySeal could not be made to perform better with a more sound training setup. However, the official implementation and weights significantly underperform compared to SNW which uses a smaller model.

The watermark capacity of the proposed method is not particularly impressive, considering that the input image resolution is 1024 × 1024. Existing methods, such as MBRS (ACM MM 2021) and FIN (AAAI 2023), have demonstrated the ability to embed dozens of bits in much smaller images. For example, if a method can embed 128 bits in a 256 × 256 image, scaling the payload proportionally to the number of pixels would correspond to 2,048 bits at 1024 × 1024 resolution. Although this scaling is not necessarily achievable in practice, it suggests that the reported capacity of 768 bits may not be particularly high. Therefore, the authors should compare their method with existing approaches under the same image resolution and payload length, while reporting both watermarked image quality (e.g., PSNR) and robustness.

We thank the reviewer for the opportunity to clarify a point that was not precise enough in the paper. All images are rescaled to 256×256256 \times 256 before embedding and decoding, following the method used by VideoSeal and PixelSeal (TrustMark resizes to 245×245245 \times 245). So every model in this work is compared for the same input size. Note that this not by choice: a limitation of watermarking based on neural networks is that they require a fixed input size (contrary to classical watermarking systems such as Broken-Arrows which could take an arbitrary input size) .

The references proposed by the reviewer are outdated, as they embed far fewer bits in smaller images compared to the current state-of-the-art that we study in this paper (i.e. the Seal family + Trustmark).

As an aside, we would like to point out to the reviewer that, contrary to their claim, doubling the input size does not double the capacity: a well-known result of the Information Security community is the square root law of capacity which states that under very general conditions the capacity of a data hiding scheme is only sublinear in the dimension of the host -- a good reference, albeit derived for steganography (the arguments would not differ much), can be found in Ker's landmark paper [1].

We argue that a fair watermarking comparison compares two schemes for a given input size and distortion budget. In this context, it is possible to compare different payload as the goal is to measure the expected number of transmitted bits without error. One could say that Chunkyseal is more efficient (though it is not, as we have just shown), but at the cost of a larger model (backbone of 2048 vs. 768). The main point of this paper is that the SNW construction method is more efficient for the same model size.

[1] Ker, A. D. (2017, June). The square root law of steganography: Bringing theory closer to practice. In Proceedings of the 5th ACM Workshop on Information Hiding and Multimedia Security (pp. 33-44).

Thank you for the detailed rebuttal and additional experiments.

The clarification that the security guarantee is specifically established against PCA-based attacks is reasonable. I also appreciate the explanation of using the hard-decision BSC model to ensure a fair comparison with existing watermarking systems.

The additional ChunkySeal results are valuable and address my concern regarding the missing high-capacity baseline. Furthermore, the clarification regarding the common model input resolution and the fixed distortion budget resolves my concern about the capacity comparison with earlier methods. I agree that linearly extrapolating watermark capacity based on image pixel count is not justified.

The rebuttal has strengthened my confidence in the proposed method and its experimental evaluation. I increase my score.

Reviewer_5kqoOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the capacity and security of neural post-hoc image watermarking. Through statistical analysis of existing encoder–decoder systems, the authors model their effective watermarking channels as parallel AWGN channels with BPSK modulation and identify underutilized latent dimensions. They propose Secure Neural Watermarking (SNW), which trains neural projections to embed and recover continuous unit vectors while using a secret orthogonal matrix to map codewords into the latent space. The paper analyzes decoding performance and resistance to PCA-based subspace estimation under a statistical perturbation model. Experiments using the PixelSeal backbone report substantially higher estimated total capacity than existing baselines under a shared PSNR budget.

Critical strengths
  1. The paper examines how the learned representation and codeword mapping jointly determine watermarking performance. The empirical analysis in Section 3 provides an informative view of latent-space utilization and motivates a concrete alternative to the standard bit-classification training objective.

  2. The keyed mapping and energy-allocation mechanism provide a clear relationship between codeword dimensionality, decoding reliability, and covariance-based leakage.

  3. SNW achieves higher estimated total capacity under many tested transformations without relying on a larger backbone.

Critical weaknesses
  1. The connection between the security definition and the PCA analysis needs clearer explanation. Definition 1 describes security in terms of forgery success following key estimation, whereas Section 5.2 analyzes whether the watermark subspace can be identified from covariance statistics. It is not sufficiently clear how these two perspectives are connected. The authors should explain how the spectral criterion in Proposition 3 translates into the security property in Definition 1, and clarify the distinction between recovering the watermark subspace and recovering the secret basis required to associate directions with particular codewords. It would also help to explain how the assumptions underlying this analysis are supported by the learned representations.

  2. The experimental comparisons do not isolate the contribution of the proposed design to the capacity improvement. SNW uses a 768-dimensional binary codeword, while PixelSeal and VideoSeal use 256-dimensional codewords. The results demonstrate that SNW can transmit a larger codeword using the existing backbone, but do not establish how much of this improvement requires the proposed training objectives and representation design. A particularly important missing baseline is the same backbone trained with the conventional end-to-end objective to transmit 768 bits under comparable conditions. Without this comparison and targeted ablations, it remains unclear whether the proposed design resolves a difficulty that a straightforward increase in codeword length would encounter.

  3. The experimental evaluation has limited coverage of relevant baselines and geometric distortions. The main performance comparison includes only PixelSeal, VideoSeal, and TrustMark, omitting relevant alternatives such as ChunkySeal, which is already discussed in the introduction and directly addresses high-capacity watermarking. Including this baseline, while accounting for differences in model size and distortion budget, would better establish the benefits of the proposed channel design. WAM could provide a complementary comparison for robustness to cropping. Furthermore, although Table 7 includes cropping, a 90° rotation, horizontal flipping, and resizing, the evaluation does not systematically cover non-right-angle rotations, a wider range of scaling factors, or combined geometric transformations. Given the emphasis on geometric robustness in the motivation, broader testing would help establish whether the reported advantages extend to these more challenging conditions.

Questions for authors
  1. Could the authors explain how the PCA-based subspace analysis in Section 5.2, particularly the spectral criterion in Proposition 3, relates to the forgery-based security definition in Definition 1? Please clarify the relationship between recovering the watermark subspace, recovering the secret basis, and forging a specified codeword. What evidence supports the applicability of the statistical assumptions to the representations produced by the trained network?

  2. How does SNW compare with a 768-bit version of the same backbone trained using the conventional end-to-end objective, under comparable training conditions and distortion budgets? Could the authors provide targeted ablations to distinguish the benefit of increasing the codeword dimensionality from the contributions of continuous-vector training and the isotropy objective?

  3. Could the authors include a comparison with ChunkySeal, accounting for differences in model size and distortion budget? A comparison with WAM could additionally assess robustness to cropping. Beyond the transformations already reported in Table 7, how does SNW perform under non-right-angle rotations, a wider range of scaling factors, and combined geometric transformations? Please examine whether its capacity advantages over the baselines remain consistent across these conditions.

6 comments

We sincerely thank the reviewer for taking the time to engage with both the theory and the finer details of the paper. We notably appreciate their acknowledgement that our analysis in Section 3 is useful to motivate an alternative construction of neural watermarking, that the proposed design makes the robustness/security design explicit, and that the proposed SNW model achieves high performance for the same backbone size. We address the remaining concerns below.

Could the authors explain how the PCA-based subspace analysis in Section 5.2, particularly the spectral criterion in Proposition 3, relates to the forgery-based security definition in Definition 1? Please clarify the relationship between recovering the watermark subspace, recovering the secret basis, and forging a specified codeword. What evidence supports the applicability of the statistical assumptions to the representations produced by the trained network?

We thank the reviewer for this opportunity to clarify the security aspect. As we state in the paper, Definition 1 means that a system is insecure as soon as it is possible for an adversary to improve their estimation of a secret key compared to a brute-force attack. The goal is to find how many watermarked observations they need to get an improvement. For SNW, the secret key is the secret matrix \mathbf{U}, which defines a secret subspace.
Under the PCA attack, we follow the same worst-case setup: as soon as the support of the Marchenko-Pastur distribution of the watermarked covariance is not perfectly included in the support of the non-watermarked one, the system is deemed insecure. Indeed an attacker could improve their brute-force attack by observing that some region of the latent space has a variance that is either too small or too large from what is expected in the non-watermarked case. They can thus focus their sampling effort within this region to increase their chance of finding the secret matrix. Proposition 3 quantifies the number of samples that allows to observe this "information leakage".
For SNW, as soon as the secret matrix is known, the attacker can recover the codeword associated to the user identity by applying the decoding pipeline. To summarize: recovering the watermark subspace is equivalent to recovering the secret basis (these are two different point of view on \mathbf{U}), and recovering \mathbf{U} allows to recover the codeword \mathbf{c}.

Regarding the statistical assumptions, the main one is that the dimension of the latent space L is large enough. More precisely, the Marchenko-Pastur theorem on which the proposition rests, assumes that, as N \rightarrow \infty, we have \frac{N}{L} \triangleq \eta for some \eta \in [0, +\infty). The qualitative result that disjoint support of the eigenvalues is equivalent to insecurity is "universal" in this sense. The quantitative result on the other hand does rest on a further assumption that a perturbation in pixel space can be modeled as a Gaussian perturbation in latent space (Eq. 7 works with the normalized perturbation). This last assumption is supported by our analysis in Section 3 as well as the corresponding figure for SNW in Appendix H.1 which shows watermarked eigenvalues match non-watermarked ones (hence our theoretical quantitative result for "perfect security" matches what is observed empircally).

How does SNW compare with a 768-bit version of the same backbone trained using the conventional end-to-end objective, under comparable training conditions and distortion budgets? Could the authors provide targeted ablations to distinguish the benefit of increasing the codeword dimensionality from the contributions of continuous-vector training and the isotropy objective?

We thank the reviewer for the advice. We fully agree that comparing with the baseline architecture and training recipe directly on 768 dimensions is an important element. Due to the limited time, we performed the following ablation study on 30k steps of training, from scratch. We contrast three different settings:

  • The same architecture with the objective to embed a 768-bits message using PixelSeal BCE loss.
  • SNW without the isotropy loss
  • SNW with the isotropy loss
    To present these results, we report the bit accuracy and the isotropy (lower is better in this table) as a function of the number of training steps.
    Here is the Table of results for bit accuracy / isotropy:
Method1k steps10k steps20k steps30k steps
Baseline BCE0.506 / 0.7250.510 / 0.7690.517 / 0.5490.516 / 0.552
SNW w/o iso0.506 / 0.8520.509 / 0.9920.508 / 0.9920.508 / 0.995
SNW w iso0.502 / 0.6200.509 / 0.5630.518 / 0.3940.525 / 0.269

Again, codeword bit-accuracy does not tell the whole story in our case since embedding a continuous vector allows us to use a continuous alphabet, which in theory increases the maximum capacity of our system. We thus complete the previous table with a measure of the cosine distance with the embedded vector for SNW methods (closer to 1 is better):

Method1k steps10k steps20k steps30k steps
SNW w/o iso0.0060.0270.0260.026
SNW w iso0.0060.0250.0590.091

We observe that combining the alignment loss with the isotropy loss leads to linear learning in the first steps, whereas using only the BCE causes the model to stagnate. We also notice that without the isotropy loss, the model does not create the expected structure, and the learning of the continuous vector stagnates after a few steps.

To complete the analysis, we wanted to determine the importance of the isotropy loss when the model has already learned to distribute the watermarking signal uniformly over the watermark space (i.e. is the isotropy loss still necessary later during training). To do so, we fine-tuned the model presented in the paper with and without the isotropy loss for 30k additional steps (again we report the cosine distance):

Method+10k steps+20k steps+30k steps
SNW w/o iso0.908 / 0.0610.907 / 0.0460.887 / 0.191
SNW w iso0.909 / 0.0330.910 / 0.0320.910 / 0.032

We observe that, the model loses performance without controlling the isotropy, which is not the case when it is controlled.
We will add this ablation analysis to the paper to justify the choice of the losses.

Could the authors include a comparison with ChunkySeal, accounting for differences in model size and distortion budget? A comparison with WAM could additionally assess robustness to cropping.

We chose not to include ChunkySeal in the experiments because it is based on exactly the same design as PixelSeal but with a larger backbone (2048 vs. 768). The whole point of the paper was to show that current approaches don't use the latent space efficiently and that one can reach far better performance without ever touching the architecture.

As a sign of good faith, however, here are the results requested by the reviewer:

AttackCodeword bit-accuracyCapacity (bits)
Identity0.87445
Contrast_2.00.85394
Brightness_0.20.87441
CenterCrop_0.90.86421
CenterCrop_0.80.85403
CenterCrop_0.70.84374
CenterCrop_0.60.82334
CenterCrop_0.50.75230
JPEG_50_augly0.85395
JPEG_80_augly0.86430
GaussianBlur_3_3_1.0_1.00.87445
Rotate_900.86419
Rotate_450.51
Resize_0.50.87445
HorizontalFlip0.87441
MedianFilter_3_30.87445
Saturation_1.50.87442

As the reviewer can see, ChunkySeal consistently underperforms compared to SNW! To be perfectly transparent, we were extremely surprised by these results. This lack of performance is linked to the whitening operation that we perform to make individual bits independent -- see Appendix F. Before this operation, the bit accuracy of ChunkySeal is indeed close to 100%, after it, it drops close to 85%! This means that baseline ChunkySeal is heavily biased towards certain messages. Even worse, if one were to use its baseline bit-accuracy to compute its actual capacity, one would overestimate it significantly due to the dependence and bias between the message bits. To be clear, this does not mean that ChunkySeal could not be made to perform better with a more sound training setup. However, the official implementation and weights significantly underperform compared to SNW which uses a smaller model.

Regarding WAM, we have trouble understanding what the reviewer expects. WAM, in its base form, only hides 32bit, and has been vastly outperformed by VideoSeal and PixelSeal which are directly based on its design. Maybe the reviewer is speaking about the "localization" mode of WAM based on masking objects in an image. In this case the experimental setting for the comparison is unclear: what model should we use for segmentation? For a fair comparison, at least 8 objects should be present so that 256bits of payload are actually hidden, which might be somewhat rare in practice. Should we use random masks instead? In any case, it is not clear to us what more can be gained from these results.

As an aside, WAM has been demonstrated to be extremely insecure: it is possible to retrieve and/or remove the watermarking signal by simply applying a Fourier transform on a set of watermarked images [1]. As such, even if WAM had a somewhat better robustness to cropping (which we doubt given its extremely small maximum capacity), one should never use it given this vulnerability.

Again, we are nevertheless ready perform another experiment provided the reviewer provides a little more detail on what they expect.

Reviewer_2okyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper models post-hoc image watermarking as parallel AWGN channels with BPSK modulation. It argues that current systems have three flaws: no secret key, wasted latent dimensions, and a 1-bit-per-symbol limit. To overcome these limitations, it proposes SNW. SNW adds a secret semi-orthogonal key U and trains the network for isotropic latents. The paper proves SNW is secure against PCA key estimation.

Critical strengths
  • “Modelling modern DNN watermarking as AWGN channels is new and well explained
  • Separating watermarking security from adversarial robustness is a valuable point.
  • The capacity gain is large and comes without changing the backbone.
Critical weaknesses
  • The embedder is public. Therefore, Eve can decode Camille’s vector from one image and re-embed it without the secret key U.
  • The paper says adversarial attacks are out of scope (Sec. 2.3). Yet it runs WM Forger, an optimization-based attack, for removal (Table 6). However, SNW is not tested against WM Forger in its forgery mode.
  • SNW is not re-attacked after the anti-spoofing fine-tuning. Lack of adaptive attacks weakens the model’s security claim since the model is public.
  • The gain comes from more bits, not better bits. SNW is weaker per bit. Under many transformations in Table 7, SNW has higher total capacity but a lower rate than PixelSeal. At Crop 50%, PixelSeal even has more total capacity (143.7 vs. 102.0 bits). The main text shows only total capacity (Figures 3–4), which may be misleading.
  • The method uses multiple losses and three stages, but only the anti-spoofing loss is ablated.
  • Although LPIPS values are in the same order of magnitude, it is about 5× worse than TrustMark, and the paper does not report SSIM, which is common in watermarking literature.
  • The introduction says SNW needs significantly less training compute. However, it only reports 590 hours on one NVIDIA L40S GPU (Reproducibility section). No training cost is given for other baselines.
Questions for authors

Please refer to the critical weaknesses

8 comments

We thank the reviewer for the review, for noting the originality of the approach, for valuing the distinction between robustness and watermark security, and for recognizing the improvement in SNW capacity using the same backbone. We address the remaining concerns below.

The embedder is public. Therefore, Eve can decode Camille’s vector from one image and re-embed it without the secret key U.

The reviewer is correct, but this does not fall under our threat model. If that was the case, then all current post-hoc systems would be useless. We need to clarify the scenario. As mentioned in the paper, the key is fixed. The attacker, Eve, has white-box access to the decoder but access to the embedder only through an API (where the key is fixed). The embedder takes only a codeword as input, since the key is fixed. In the threat model, Eve wants to spoof Camille's identity. With knowledge of the key, Eve can decode Camille's codeword and then use the codeword with the embedder API, which would otherwise be impossible without the key. We will clarify this scenario in Section 2 of the paper.

The paper says adversarial attacks are out of scope (Sec. 2.3). Yet it runs WM Forger, an optimization-based attack, for removal (Table 6). However, SNW is not tested against WM Forger in its forgery mode.

We apologize to the reviewer for what seems to be a misunderstanding. In this paper, we make a precise distinction between operations that are deemed "adversarial" and those that are deemed "blind". We consider an operation to be adversarial if they are made out of a optimization loop which directly try to minimize or maximize the value of a (possibly imperfect) oracle of a detector or decoder score. Otherwise we say that the attack is blind (in the paper we call it an "augmentation"). In this setup, adversarial operations would be white-box attacks such as DDN Attack [1] or black-box attacks such as CGBA or Surfree [2,3]. Obviously, regeneration attacks or VAE-attacks are blind since they don't optimize anything. In the case of WMForger, we still say the attack is blind because its optimization loop is not based on an oracle of the true detector, it is only based on a gradient ascent of an ad-hoc preference model. We make further distinction and discussion in Appendix E. This is the reason why we say that "adversarial attacks are out of scope" yet still study what some would call "adversarial attacks". We will clarify this point Section 2.3.

Now, with regard to the forgery mode of WM Forger. Before performing any experiments, it is expected that this mode cannot work on SNW which was fine-tuned with an anti-spoofing loss. Indeed, the best possible forgery attack of WMForger would happen if the exact watermarking residual was extracted by WMForger and copied over a clean image. Yet, this exactly the attack against which SNW is trained to be robust against. To demonstrate that this the case empirically, we first estimate the watermarking residual using WMForger on 1000 image watermarked with SNW and transfer each residual to another, different, clean image. In the paper, Table 5, we already present SNW results against transferring the exact watermarking residual, which is the optimal estimate. We show that after spoofing-loss fine-tuning, residual transfer does not work. Likewise, we end up with same conclusion with WM Forger forgery mode:

AttackSpoofed bit-accuracy
WMForger / 100 steps0.502
WMForger / 50 steps0.495

SNW is not re-attacked after the anti-spoofing fine-tuning.

We thank the reviewer for pointing out a missing information in the paper. Indeed, the results presented in the paper were computed using the fine-tuned model, which as we have just shown, is robust against spoofing attacks.

Lack of adaptive attacks weakens the model’s security claim since the model is public.

As mentioned in the paper, this work is out of scope for white-box and black-box adversarial attacks. However, Appendix E discusses adversarial robustness. To complete the analysis, we will add the following discussion.

The first step in performing these attacks on the detector is to estimate the key, which, as shown in the paper, is not trivial. If we assume the key is known, we return to the neural watermarking baselines following the HiDDeN paradigm. As said in Appendix E, to give adversarial robustness guarantees in terms of distortion, we need to control the Lipschitz constant of the decoder's projection function. For neural networks, this constant is very high, as exhibited in Table 5. However, classic watermarking approaches, such as Broken-Arrows, rely on 1-Lipschitz functions. On the other hand, this classic watermarking construction lacks robustness to geometric transformations.

To complete this analysis, we examine geometric robustness as a function of the projection function's Lipschitz constant. Starting from our last checkpoint, we fine-tuned the model and progressively reduced the Lipschitz constant. Here are the results on the training cosine similarity:

Lipschitz constantIdentityJPEGCrop
690.960.960.84
600.960.960.82
500.960.960.83
400.960.960.81
300.960.960.81
200.960.950.75
100.950.940.63
40.940.940.50

During training, the crop is a random crop with a factor uniformly sampled between 0.5 and 1.0. The cosine reported in the Table is a mean. As the Lipschitz constant declines, robustness and detectability decline globally. However, this decline is more important for geometric transformations than for valuemetric transformations.

In summary: There is a trade-off between robustness to geometric transformation and watermarking security.

The gain comes from more bits, not better bits. SNW is weaker per bit. Under many transformations in Table 7, SNW has higher total capacity but a lower rate than PixelSeal. At Crop 50%, PixelSeal even has more total capacity (143.7 vs. 102.0 bits). The main text shows only total capacity (Figures 3–4), which may be misleading.

We respectfully disagree that this is a valid criticism. Two major points in the paper are:

  1. It is highly non-trivial to "add more bits" for a fixed backbone. One cannot change the payload size of modern watermarking with the same ease as what was possible with classical schemes: the DNNs need to be fully retrained. We would also like to point out that, contrary to ChunkySeal, we don't "add more bits" simply by making the DNN larger, a point which was raised by the reviewer themselves. One important lesson of the paper is that we can "add more bits" by simply using the watermarking space more efficiently.
  2. Bit-accuracy is meaningless by itself. As we already state in the paper (line 448), even for PixelSeal which boasts a bit-accuracy of 0.997 the probability of decoding error is still 0.53! Maybe one of the most important message we wanted to deliver in the paper is that Shannon capacity IS the better metric since it allows to measure the expected number of transmitted bits while still being agnostic to implementation details (choice of error-correcting code, block-length, etc...) and valid for any strictly positive probability of decoding error. Consequently, we argue that higher capacity still means that bits are also "stronger" overall.

The method uses multiple losses and three stages, but only the anti-spoofing loss is ablated.

We thank the reviewer for the remark. Due to the limited time, we performed the following ablation study on 30k steps of training, from scratch. We contrast three different settings:

  • The same architecture with the objective to embed a 768-bits message using PixelSeal BCE loss.
  • SNW without the isotropy loss
  • SNW with the isotropy loss
    To present these results, we report the bit accuracy and the isotropy (lower is better in this table) as a function of the number of training steps.
    Here is the Table of results for bit accuracy / isotropy:
Method1k steps10k steps20k steps30k steps
Baseline BCE0.506 / 0.7250.510 / 0.7690.517 / 0.5490.516 / 0.552
SNW w/o iso0.506 / 0.8520.509 / 0.9920.508 / 0.9920.508 / 0.995
SNW w iso0.502 / 0.6200.509 / 0.5630.518 / 0.3940.525 / 0.269

Again, codeword bit-accuracy does not tell the whole story in our case since embedding a continuous vector allows us to use a continuous alphabet, which in theory increases the maximum capacity of our system. We thus complete the previous table with a measure of the cosine distance with the embedded vector for SNW methods (closer to 1 is better):

Method1k steps10k steps20k steps30k steps
SNW w/o iso0.0060.0270.0260.026
SNW w iso0.0060.0250.0590.091

We observe that combining the alignment loss with the isotropy loss leads to linear learning in the first steps, whereas using only the BCE causes the model to stagnate. We also notice that without the isotropy loss, the model does not create the expected structure, and the learning of the continuous vector stagnates after a few steps.

To complete the analysis, we wanted to determine the importance of the isotropy loss when the model has already learned to distribute the watermarking signal uniformly over the watermark space (i.e. is the isotropy loss still necessary later during training). To do so, we fine-tuned the model presented in the paper with and without the isotropy loss for 30k additional steps (again we report the cosine distance):

Method+10k steps+20k steps+30k steps
SNW w/o iso0.908 / 0.0610.907 / 0.0460.887 / 0.191
SNW w iso0.909 / 0.0330.910 / 0.0320.910 / 0.032

We observe that without isotropy, the model loses performance without controlling the isotropy, which is not the case when it is controlled.
We will add this ablation analysis to the paper to justify the choice of the losses.

Although LPIPS values are in the same order of magnitude, it is about 5× worse than TrustMark, and the paper does not report SSIM, which is common in watermarking literature.

We compare model performance at a fixed PSNR budget of 48 dB, which is standard for PixelSeal and VideoSeal. At this distortion level, TrustMark's signal is very weak, with a capacity of 63 bits when no attack is performed. Compare to > 200 capacity of PixelSeal and VideoSeal and > 650 capacity for SNW. We will correct the claim to state explicitly that it refers to VideoSeal and PixelSeal, which propose higher detectability at this distortion. As requested by the reviewer, we computed the SSIM. Results are reported in the following Table, and will be added to the paper's Table 2:

PixelSealVideoSealTrustMarkSNW
0.9950.9910.9950.990

As expected, the values are in the same order of magnitude.

Comments on the paper

1 comment

We sincerely thank all reviewers for a discussion that we genuinely believe to have been fruitful. The results obtained during this round have both clarified and strengthened our arguments.

We have clarified our threat model and the link between our definition of watermarking security and one of its possible measure through the number of samples necessary to successfully perform a PCA attack (reviewers 2oky, 5kqo and uF9d).

Most importantly, a more complete ablation study was added at the request of the three reviewers.

Finally we added several experiments which substantially reinforced our claims about the importance of using the watermarking space efficiently:

  • we added an analysis of the impact of the Lipschitz constant on the robustness, demonstrating the trade-off between (classic) robustness and adversarial robustness. In passing, we also showed that SNW boasts the smallest Lipshchitz constant of the tested methods despite having the highest capacity. (reviewers 2oky, 5kqo)
  • we performed more attacks to complete our benchmark, including combined operations, non right-angle rotations, resizing factors and a state-of-the-art "adaptive" attack, WMForger. In the latter case, we demonstrated that WMForger and other spoofing attack based on copying the watermarking signal on a cover image do not work on SNW. (reviewer 2oky,5kqo)
  • we compared SNW to a larger backbone, ChunlySeal which boast a codeword of size 1024 bits. Most surprising to us, we demonstrated that our smaller SNW has both a higher total capacity -- 465 bits vs 290 bits on average over common image operations -- and a higher rate per pixel than ChunkySeal -- 0.61 vs 0.28. (reviewers 2oky, 5kqo and uF9d)

At this point of the rebuttal, all reviewers have noted and validated the improvements, with two of them already having increased their scores.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.