acceptodds
Under review as a conference paper at ICLR 2027

Compositional Adversarial Training for Robust Visual Watermarking

Abstract

Robust watermarking is typically trained with random post-processing augmentation, but random sampling under-covers the combinatorial space of realistic attack pipelines and rarely encounters the rare compositions that actually break detection. This leads to unstable training and poor sample efficiency. We instead formulate watermark robustness as a min-max problem over a structured space of compositional transformations. We propose Compositional Adversarial Training (CAT), a plug-in framework that learns a sequential differentiable adversary that observes the current intermediate image and adaptively selects an attack family at each step to maximally disrupt message recovery. CAT combines a straight-through Gumbel-Softmax attack selection with entropy regularization, allowing the backward pass to be end-to-end differentiable and aggregate gradient information across attack families, yielding faster, smoother convergence without collapsing to a single attack mode. We evaluate CAT on post-generation watermarks VideoSeal 0.0, VideoSeal 1.0, and PixelSeal and in-generation WMAR under both single-step and two-step attack suites, on in-distribution and multiple out-of-distribution image and video benchmarks. CAT consistently outperforms random-augmentation baselines trained for the same number of gradient updates and the same augmentation budget, with the largest gains on hard composed attacks and OOD evaluations; in the SA-1B and CLIC image benchmarks, it improves overall watermark capacity by up to 74.3% in the single-step setting and 24.2% in the compositional setting. In the autoregressive setting, CAT raises difficult geometric TPR@FPR by up to 18 percentage points. These results show that robust visual watermarking benefits from training against adaptive compositional adversaries rather than independent random corruptions.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
visual watermarking, adversarial training, gumbel softmax, entropy regularization

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_U74NOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Compositional Adversarial Training (CAT). It uses a controller to adaptively select and compose image transformations, so as to improve the robustness of watermarking models against complex post-processing attacks. Experiments show that CAT outperforms the random augmentation baseline across multiple watermarking models and datasets, and achieves favorable performance under compositional attack tests.

Critical strengths
  1. Compared with single perturbations, the research is more aligned with practical scenarios. The paper considers the impacts on watermarks brought by sequentially composed multiple image processing operations, and conducts evaluations on long attack chains that simulate social media reposting and editing.
  2. CAT selects attack types and their order step by step conditioned on intermediate images. It introduces an adversarial controller during training without modifying the model inference pipeline.
  3. With the same number of training updates and augmentation budget, CAT generally outperforms the random augmentation baseline. It achieves strong performance under challenging compositional attacks and out-of-distribution evaluations.
Critical weaknesses
  1. The appendix experiments show that CAT is nearly ineffective against diffusion-based regeneration, with accuracy close to random guesses. It exhibits limited performance against semantic-level watermark removal. If the method is only effective for conventional processing operations, its practical value is constrained. The paper could further discuss how this applicability boundary impacts real-world watermark protection scenarios.
  2. CAT’s adversarial search space is constrained by a predefined library of attack primitives. The controller can only select and compose attack sequences from the given image transformations, essentially searching for the worst-case compositional perturbations within a known transformation space. Although the paper evaluates several held-out attacks, the results do not establish how well CAT generalizes to a broader range of transformations beyond those considered.
  3. The paper compares CAT’s performance with random augmentation, Faster AutoAugment, and several alternative adaptive strategies. However, for RoMark, GAT and Adversarial AutoAugment, it mainly reports training overhead without robustness comparisons on the same watermarking model and attack test sets. Therefore, existing experiments support that CAT has lower training cost under the evaluated settings, but cannot directly confirm whether its robustness surpasses these methods.
Questions for authors
  1. Could the author further analyze and explain the main reasons why CAT performs poorly against semantic watermark-destruction attacks such as diffusion-based regeneration? How does this limitation affect the method’s practical scope?
  2. How well does CAT generalize to transformations that are not included in the predefined attack primitive library?
  3. Under the same evaluation settings, how does CAT perform in terms of robustness compared with RoMark, GAT and Adversarial AutoAugment?
2 comments
AuthorsOpenReview ↗

Response to Reviewer U74N (1/2)

(1/2) We thank the reviewer for recognizing the practical relevance of composed processing operations, effectiveness against long and realistic attack chains, and CAT's strong performance under compositional and out-of-distribution attacks. In what follows, we address all of the reviewer's concerns.

W1 & Q1: Diffusion-based regeneration and practical scope

Diffusion-based regeneration removes every imperceptible post-hoc watermark, which prior work proves theoretically, so the chance-level accuracy reflects a limit of any posthoc watermark rather than of CAT. Zhao et al. (NeurIPS 2024) prove that regeneration attacks, which add Gaussian noise to the watermarked image and reconstruct it with a generative model, remove any watermark whose perturbation is small in ℓ2 norm. After the noise is added, the image is nearly indistinguishable in distribution from a noised unwatermarked image, so no decoder can recover the message, however it was trained. Saberi et al. (ICLR 2024) prove a fundamental trade-off between evasion and spoofing errors for watermarks that introduce subtle perturbations, and show that diffusion purification removes such watermarks in practice. Our DiffPure attack follows their setup. Consistent with these results, we find that all six models in Appendix F, trained with random augmentation or with CAT, decode at 50.0% to 51.6% bit accuracy under DiffPure. No training procedure can restore information that the attack destroys, and a watermark strong enough to survive the injected noise would no longer be imperceptible.

CAT does improve robustness to generative attacks that preserve part of the signal. Averaged over the three models, we improve bit accuracy over random augmentation by +4.4 pp under VAE regeneration, +2.9 pp under learned neural compression, and +3.0 pp under semantic region inpainting.

The applicability boundary is therefore the boundary of post-hoc imperceptible watermarking, and CAT targets the threat that falls inside it. Watermarks applied after generation, as in VideoSeal and PixelSeal, must survive the processing that content undergoes when it is distributed: recompression, resizing, cropping, color edits, transcoding, and repeated reposting. These operations are the most common cause of watermark loss, require no attacker, and are what we train CAT against. Deliberate regeneration with a diffusion model requires a different kind of watermark, such as semantic watermarks embedded during generation (e.g., Tree-Ring, Wen et al., NeurIPS 2023), which need control of the generator and cannot mark existing content. We will add this discussion and both citations to the paper.

W2 & Q2: Generalization beyond the primitive library

CAT's gains carry over to transformations outside its library. Every method we compare with draws from a fixed library, and CAT changes how compositions are chosen, not which primitives exist. To answer Q2 directly, we evaluate PixelSeal (depth 1) on SA-1B images against 10 families outside the training library, each at several fixed strengths. We report bit accuracy averaged over strengths, with the LPIPS to the watermarked image:

TransformationLPIPSRandomCAT
No attack–92.098.6
WebP (quality 50, 30, 10)0.1589.497.0
Salt-and-pepper (1%, 3%, 5%)0.7285.093.8
Pixelation (4, 8, 16 px)0.4588.894.6
Gaussian noise (σ 0.02, 0.05, 0.1)0.3990.697.3
Gamma (0.5 to 2)0.1091.197.2
Sharpening (2, 4, 8)0.2391.598.2
Median filter (3, 5, 7)0.1891.798.1
Resize (0.75 to 0.2)0.3191.998.2
Grayscale0.2091.998.5
Shear (10°, 20°, 30°)0.4982.383.3
Average–89.495.6

CAT is ahead on all ten families, by +6.2 pp on average. We further compose these families with each other and with every training family one step beyond its training range (e.g., JPEG quality 30, crop 0.45, rotation 15°), and apply all 342 ordered pairs of two different transformations:

CompositionPairsLPIPSRandomCAT
Unseen + unseen900.4489.295.5
Unseen + out-of-range1800.5478.684.4
Out-of-range + out-of-range720.6270.375.5
All3420.5379.685.4

CAT is ahead on 310 of the 342 compositions, by +5.8 pp on average. Against a model-aware attacker that selects, per image, the most damaging transformation or composition of two from a pool including unseen families and training families far beyond their ranges, CAT stays ahead at every LPIPS budget, by +5.9 pp at LPIPS ≤ 0.1 and +3.0 pp at LPIPS ≤ 0.3. The realistic attacks of Appendix F are also outside the library, and there CAT gains +2.9 pp and +20.4 bits under neural compression and +11.1 pp and +41.1 bits under repeated video transcoding. We will add these results to the paper.

AuthorsOpenReview ↗

Response to Reviewer U74N (2/2)

W3 & Q3: Robustness comparison with RoMark, GAT, and AAA

Prior adversarial methods attack with the wrong kind of distortion or with too much strength, and on the same watermarking model and attack test sets, CAT outperforms RoMark, GAT, AAA, and DADW under an equal compute budget. We fine-tune the same PixelSeal checkpoint, trained for 100 epochs without attacks, with each method for 24 hours on 4 A4000 GPUs. We keep the published hyperparameters of each method: component-wise PGD with T=7 and an ℓ∞ budget of 4/255 for GAT, a 6-point severity grid for RoMark, M=8 policies for AAA, and, for Distortion Agnostic Deep Watermarking (DADW), a two-layer convolutional attack network with 5 gradient steps per watermarking step. We measure a training step at 1.33 s for CAT, 3.75 s for DADW, 4.53 s for AAA, 9.57 s for RoMark, and 16.15 s for GAT, so slower methods complete fewer epochs. We evaluate all models with the attack suite of the main tables and report bit accuracy (%) averaged over SA-1B, CLIC-train, CLIC-test, DIV2K, and MetFaces:

MethodDepthEpochsIdentityValueCompressionGeometricCombined
GAT–2249.949.949.949.949.9
RoMark13655.554.753.451.154.7
DADW–9794.793.780.160.794.1
AAA15794.893.992.486.093.5
CAT1210100.099.599.094.598.6
AAA25891.190.390.388.089.5
CAT2183100.099.599.597.298.7

We find that each prior method fails for a reason that follows from its design:

  • RoMark and GAT optimize the worst-case strength of fixed transformations for each image. In the first epoch, this pushes the decoding loss above chance level, the decoder converges to an uninformative output, and accuracy stays at chance. When we remove GAT's ℓ∞ component, it still collapses (50.0%), and RoMark's own Algorithm 1 avoids collapse but does not improve on the starting checkpoint (82.6% vs. 83.3% without attacks).
  • AAA samples policies that ignore the image and trains on 8 augmented copies of every batch, so it completes only 57 epochs. With CAT, we improve on it by +6.5 pp at depth 1 and +9.2 pp at depth 2, with the largest gains on geometric attacks (+8.5 pp and +9.2 pp), at matched image quality.
  • DADW learns local pixel-level perturbations, which cannot express geometric transformations, and reaches only 60.7% under geometric attacks, against 94.5% for CAT.

CAT avoids all three failure modes. We compose real-world transformations, including geometric ones, chosen with their order by a controller conditioned on the image and the intermediate attacked state. We sample strengths from the training distribution, so every training example remains learnable while the adaptive compositions supply the difficulty. CAT is also the cheapest method per step. We will add these results and the per-step costs to the paper.

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

This paper proposes Compositional Adversarial Training (CAT), a framework that improves visual watermark robustness through adaptive, sequential attack selection. It uses a recurrent controller, straight-through Gumbel-Softmax, and entropy regularization to construct challenging attack compositions. Experiments across image, video, and autoregressive watermarking demonstrate improvements over random augmentation, particularly under geometric and compositional distortions.

Critical strengths

Effective adaptive training strategy. CAT achieves consistent improvements over random augmentation across multiple watermarking models and datasets. The sequential attack selection mechanism is practically motivated and introduces moderate training overhead.

Extensive experiments and ablations. The paper evaluates image and video watermarking, out-of-distribution generalization, and longer attack sequences. The ablation studies provide useful evidence for sequential factorization, intermediate-state conditioning, and entropy regularization.

Critical weaknesses

Learning adversarial transformations has been explored in previous watermarking methods, such as Distortion Agnostic Deep Watermarking (CVPR 2020) and RoMark, while GAT and Adversarial AutoAugment have investigated compositional or adaptive adversarial training. CAT mainly combines existing optimization techniques with sequential, image-conditioned attack selection. Although several alternative controllers are evaluated, direct robustness comparisons with closely related adversarial training approaches are missing.

The evaluation uses different payload lengths (96, 100, 128, and 256 bits), making absolute capacity comparisons difficult without controlling image resolution and visual fidelity. Notably, Table 1 reports higher overall capacity for MBRS (256 bits) than for CAT-trained VideoSeal 1.0 (256 bits) under single-step attacks on both SA-1B and CLIC. Furthermore, the reported capacity is estimated using a binary symmetric channel formula rather than validated through actual reliable message transmission. Stronger matched-condition comparisons are needed to establish the practical capacity advantage.

Questions for authors
  1. Can the authors clarify the methodological differences from DADW, RoMark, and GAT, and provide a performance comparison with at least one closely related adversarial training method under a matched computational budget?

  2. Can the authors provide comparisons under matched payload, resolution, and PSNR settings, particularly against MBRS and other strong watermarking baselines?

  3. Can the authors validate the estimated channel capacity through actual message recovery with error-correcting codes, or provide evidence supporting the assumed bit-error channel model?

5 comments
AuthorsOpenReview ↗

Response to Reviewer DGsr (1/3)

We thank the reviewer for recognizing the effectiveness of our adaptive training strategy, its consistent improvements over random augmentation across watermarking models and datasets, and the extensive experiments and ablations. In what follows, we address all of the reviewer's concerns.

W1 & Q1: Relation to DADW, RoMark, GAT, and AAA

Prior adversarial methods attack with the wrong kind of distortion or with too much strength, and CAT outperforms all of them under an equal compute budget. We fine-tune the same PixelSeal checkpoint, trained for 100 epochs without attacks, with each method for 24 hours on 4 A4000 GPUs. We keep the published hyperparameters of each method: a two-layer convolutional attack network with 5 gradient steps per watermarking step for DADW, a 6-point severity grid for RoMark, component-wise PGD with T=7 and an ℓ∞\ell_{\infty} budget of 4/255 for GAT, and M=8 policies for AAA. We measure a training step at 1.33 s for CAT, 3.75 s for DADW, 4.53 s for AAA, 9.57 s for RoMark, and 16.15 s for GAT, so slower methods complete fewer epochs. We report bit accuracy (%) averaged over SA-1B, CLIC-train, CLIC-test, DIV2K, and MetFaces:

MethodDepthEpochsIdentityValueCompressionGeometricCombined
GAT–2249.949.949.949.949.9
RoMark13655.554.753.451.154.7
DADW–9794.793.780.160.794.1
AAA15794.893.992.486.093.5
CAT1210100.099.599.094.598.6
AAA25891.190.390.388.089.5
CAT2183100.099.599.597.298.7

We find that each prior method fails for a reason that follows from its design:

  • DADW learns an attack network that adds local pixel-level perturbations. These cannot express geometric transformations, so its decoder never learns them and reaches only 60.7% under geometric attacks, against 94.5% for CAT.
  • RoMark and GAT optimize the worst-case strength of fixed transformations for each image. In the first epoch, this pushes the decoding loss above chance level, the decoder converges to an uninformative output, the embedder's gradient vanishes, and accuracy stays at chance. To rule out an artifact of our adaptation, we remove GAT's ℓ∞ component so that it optimizes only its semantic transformations, and it still collapses (50.0% after 26 epochs). We also run RoMark's own Algorithm 1, which optimizes the strength of one family per sub-batch: it avoids collapse but does not improve on the starting checkpoint (82.6% vs. 83.3% without attacks, 57.4% geometric).
  • AAA samples augmentation policies that ignore the image and trains on 8 augmented copies of every batch. It completes only 57 epochs and does not reach full accuracy even without attacks. With CAT, we improve on it by +6.5 pp at depth 1 and +9.2 pp at depth 2 averaged over the four attack categories, with the largest gains on geometric attacks (+8.5 pp and +9.2 pp), at matched image quality (PSNR 32.9 to 33.3 dB).

CAT avoids all three failure modes. We compose real-world transformations, including geometric ones. We train a recurrent controller that chooses which transformation to apply next, and in which order, conditioned on the image and the intermediate attacked state, end to end with straight-through Gumbel-Softmax and entropy regularization. We sample strengths from the training distribution rather than optimizing them to the worst case, so every training example remains learnable while the adaptive compositions supply the difficulty. CAT is also the cheapest method per step. We will add these results and the per-step costs to the paper.

AuthorsOpenReview ↗

Response to Reviewer DGsr (2/3)

W2 & Q2: Matched comparison with MBRS

The gap between MBRS and VideoSeal 1.0 in Table 1 reflects the VideoSeal 1.0 architecture, and both training methods are limited by it. Even without attack, VideoSeal 1.0 decodes only 89.7% of bits with random augmentation and 94.0% with CAT, while MBRS, InvisMark, TrustMark, and our other models trained with CAT reach 98.7–99.8%. Its overall accuracy is the lowest of the three architectures we train under both methods (84.6% and 85.9%). A fair comparison with MBRS therefore needs a capable model at the same payload. We train PixelSeal with 256 bits using the main recipe, and raise the strength factor of the released MBRS model, which is trained with real and simulated JPEG compression, from 1.0 to 1.33 to match PSNR (43.1 vs. 42.8 dB). Both embed at 256×256 and upsample the watermark residual to the original resolution, and we evaluate both on the same attacks. Bit accuracy (%) / capacity (bits) on SA-1B:

CAT is better on geometric attacks, the hardest category and the cheapest for an attacker, by +17.5 pp and +48 bits. Cropping, rotating, or rescaling an image costs an attacker nothing and leaves the content intact. On compositions, which are the more realistic threat, CAT is better whenever a compression or geometric step is involved and by +23.1 pp and +79 bits overall, while MBRS falls to chance under two geometric steps.

AttacksMBRS (256)PixelSeal (256) + CAT
Identity99.9 / 25596.8 / 207
Value94.9 / 21491.6 / 169
Compression89.7 / 20392.1 / 174
Geometric65.7 / 8383.2 / 131
Combined99.0 / 24694.5 / 184
Val+Val95.5 / 21592.6 / 171
Val+Comp88.8 / 16894.5 / 186
Val+Geom61.8 / 5487.8 / 139
Comp+Comp83.3 / 12796.4 / 202
Comp+Geom57.2 / 2592.5 / 166
Geom+Geom49.6 / 187.5 / 135
All compositions67.3 / 7790.4 / 156

Sweeping the strength of both models from 39.5 to 47.5 dB does not change this. Bit accuracy (%), MBRS / PixelSeal (256) + CAT:

PSNR (dB)GeometricAll compositionsGeom+Geom
47.5 / 47.464.4 / 75.862.9 / 79.149.5 / 78.2
45.6 / 46.065.3 / 79.165.0 / 84.649.5 / 82.9
44.0 / 43.965.6 / 82.866.5 / 89.449.6 / 86.9
43.1 / 42.965.7 / 83.667.3 / 90.649.6 / 87.9
42.1 / 42.165.7 / 84.068.1 / 91.249.6 / 88.5
40.7 / 41.065.8 / 84.768.9 / 92.449.7 / 89.4
39.6 / 39.465.8 / 85.169.5 / 92.949.7 / 89.8

A stronger MBRS watermark improves its accuracy without attacks and under value attacks, but its geometric accuracy stays at 64–66% and it remains at chance under two geometric steps at every strength. Its weakness to geometric transformations is therefore structural and cannot be traded against image quality, and CAT at its weakest strength exceeds MBRS at its strongest.

AuthorsOpenReview ↗

Response to Reviewer DGsr (3/3)

W3 & Q3: Validation with error-correcting codes

With an actual error-correcting code, CAT recovers the exact message far more often than random augmentation, at both depths. As a deployed system would, we protect each message with a BCH code, which adds redundancy so that the message is recovered exactly as long as at most t bits are flipped. We encode a random message, embed the codeword in PixelSeal (128 bits) trained at depth 1 or 2, attack the image, decode the bits, and correct them with the BCH decoder. With a code that carries 92 message bits and corrects up to 5 flipped bits, the share of SA-1B images whose message is recovered exactly (%) is:

TrainingDepthIdentityValueCompressionGeometricCombinedDepth-2 attacks
Random12.81.81.60.61.81.1
CAT191.071.670.948.679.858.1
Random295.982.482.470.470.474.9
CAT299.289.891.980.793.384.6

At depth 1, random augmentation recovers at most 2.8% of messages, while CAT recovers 48.6% to 91.0%. At depth 2, CAT improves exact recovery by +10.3 pp under geometric attacks, +22.9 pp under combined attacks, and +9.7 pp under depth-2 compositions. A longer message leaves fewer bits for error correction, and choosing the best code for each model, CAT transmits 54 bits per image against 22 for random augmentation at depth 1, and 87 against 70 at depth 2.

Reviewer_N6WdOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Compositional Adversarial Training (CAT), a sequential adversarial training framework for robust watermarking. CAT conditions attack selection on intermediate attacked images and learns both attack types and their order using Gumbel-Softmax and entropy regularization. Across VideoSeal, PixelSeal, and WMAR, CAT improves robustness and watermark capacity over random augmentation, especially under difficult compositional and OOD attacks, while preserving perceptual quality.

Critical strengths
  1. This paper addresses the limited coverage of rare compositional attacks under random augmentation.

  2. The evaluation spans image, video, and autoregressive watermarking, including compositional, OOD, and editing attacks.

Critical weaknesses
  1. The adversarial optimization is limited to selecting attack families and their order, while attack strengths are sampled from predefined distributions. The robustness to unseen attacks remains unclear.

  2. The main comparisons are against random augmentation and FAA; RoMark, GAT, and Adversarial AutoAugment are mainly evaluated through computational-cost analyses rather than full performance comparisons.

  3. Many OOD experiments reuse the same attack families and parameter ranges, so generalization to unseen or adaptive attacks is still unclear.

Questions for authors
  1. Can the authors evaluate CAT when attack strengths are adversarially optimized rather than sampled from fixed distributions?

  2. Can the authors provide full robustness comparisons with RoMark, GAT, and Adversarial AutoAugment under comparable computational budgets?

  3. Can the authors evaluate CAT against model-aware attacks that adapt their transformations or strengths to the target watermarking model, beyond the fixed attack families and parameter ranges used in training?

4 comments
AuthorsOpenReview ↗

Response to Reviewer N6Wd (1/2)

We thank the reviewer for recognizing the breadth of our evaluation across image, video, and autoregressive watermarking, including compositional, out-of-distribution, and editing attacks. We especially appreciate the reviewer's recognition that CAT addresses the limited coverage of rare compositional attacks under random augmentation. In what follows, we address the reviewer's concerns.

W1: Sampled strengths and robustness to unseen attacks

We sample strengths from the training distribution by design, because adversarially optimizing them makes training collapse. We show in our response to W2 that the two methods that optimize worst-case strengths during training, RoMark and GAT, collapse to chance accuracy, and in our response to W3 that CAT remains more robust than random augmentation when an attacker optimizes strengths at test time.

CAT's gains carry over to transformation families never seen in training. We evaluate PixelSeal (depth 1) on 200 SA-1B images against ten families outside the training library, each at several fixed strengths. We report bit accuracy (%) averaged over strengths, with the LPIPS of the attacked image to the watermarked image:

TransformationLPIPSRandomCAT
No attack–92.098.6
WebP (quality 50, 30, 10)0.1589.497.0
Salt-and-pepper (1%, 3%, 5%)0.7285.093.8
Pixelation (4, 8, 16 px)0.4588.894.6
Gaussian noise (σ 0.02, 0.05, 0.1)0.3990.697.3
Gamma (0.5 to 2)0.1091.197.2
Sharpening (2, 4, 8)0.2391.598.2
Median filter (3, 5, 7)0.1891.798.1
Resize (0.75 to 0.2)0.3191.998.2
Grayscale0.2091.998.5
Shear (10°, 20°, 30°)0.4982.383.3
Average–89.495.6

CAT is ahead on all ten families, by +6.2 pp on average, with the largest gains under salt-and-pepper noise (+8.8 pp) and WebP (+7.6 pp). We further compose these families with each other and with every training family one step beyond its training range (JPEG quality 30, crop 0.45, rotation 15°, perspective 0.6, blur kernel 21, brightness 0.4, contrast 2.5, saturation 2.5, hue 0.15), and apply all 342 ordered pairs of two different transformations:

CompositionPairsLPIPSRandomCAT
Unseen + unseen900.4489.295.5
Unseen + out-of-range1800.5478.684.4
Out-of-range + out-of-range720.6270.375.5
All3420.5379.685.4

CAT is ahead by +5.8 pp on average. The realistic attacks of Appendix F are also outside the training library, and averaged over VideoSeal 0.0, VideoSeal 1.0, and PixelSeal, CAT improves bit accuracy and capacity by +2.9 pp and +20.4 bits under neural compression, +4.4 pp and +9.1 bits under VAE regeneration, +3.0 pp and +3.4 bits under region inpainting, and +11.1 pp and +41.1 bits under repeated video transcoding.

AuthorsOpenReview ↗

Response to Reviewer N6Wd (2/2)

W2, Q1 & Q2: Adversarially optimized strengths and comparison with RoMark, GAT, and AAA

Prior adversarial methods attack with the wrong kind of distortion or with too much strength, and CAT outperforms all of them under an equal compute budget. We fine-tune the same PixelSeal checkpoint with each method for 24 hours on 4 A4000 GPUs, keeping the published hyperparameters (PGD with T=7 and ℓ∞ budget 4/255 for GAT, a 6-point severity grid for RoMark, M=8 policies for AAA, and a two-layer attack network for Distortion Agnostic Deep Watermarking, DADW). A training step takes 1.33 s for CAT and 3.75 s to 16.15 s for the others, so slower methods complete fewer epochs. Bit accuracy (%) averaged over five datasets:

MethodDepthEpochsIdentityValueCompressionGeometricCombined
GAT–2249.949.949.949.949.9
RoMark13655.554.753.451.154.7
DADW–9794.793.780.160.794.1
AAA15794.893.992.486.093.5
CAT1210100.099.599.094.598.6
AAA25891.190.390.388.089.5
CAT2183100.099.599.597.298.7

We find that each prior method fails for a reason that follows from its design:

  • RoMark and GAT optimize the worst-case strength of fixed transformations for each image, which is what Q1 asks about. In the first epoch, this pushes the decoding loss above chance, and the decoder collapses to an uninformative output. This is not an artifact of our adaptation: without its ℓ∞ component, GAT still collapses (50.0%), and RoMark's own Algorithm 1, which optimizes one family per sub-batch, does not improve on the starting checkpoint (82.6% vs. 83.3% without attacks).
  • AAA samples augmentation policies that ignore the image and trains on 8 augmented copies of every batch. It completes only 57 epochs, and with CAT we improve on it by +6.5 pp at depth 1 and +9.2 pp at depth 2 averaged over the four attack categories, with the largest gains on geometric attacks (+8.5 pp and +9.2 pp), at matched image quality (PSNR 32.9 to 33.3 dB).
  • DADW learns an attack network that adds local pixel-level perturbations. These cannot express geometric transformations, so it reaches only 60.7% under geometric attacks, against 94.5% for CAT.

CAT avoids all three failure modes. We compose real-world transformations, including geometric ones, chosen with their order by a controller conditioned on the image and the intermediate attacked state, and we sample strengths from the training distribution, so every example remains learnable while the adaptive compositions supply the difficulty. We will add these results and the per-step costs to the paper.

W3 & Q3: Model-aware attacks

CAT remains more robust than random augmentation when the attacker adapts transformations and their strengths to the target watermarking model. We evaluate PixelSeal on SA-1B against two attackers that do not know the message. In a query-based search, the attacker runs the decoder and, per image, selects the transformation whose decoded bits change most within an LPIPS budget, from a pool that extends every training family far beyond its range (e.g., JPEG quality 10, rotation ±45°), adds four unseen families, and includes compositions of two. We also run GAT's attack (PGD with T=10), which optimizes the strengths of hue, saturation, brightness, contrast, and rotation through the decoder. Bit accuracy (%):

MethodDepthNo attackSearch, LPIPS ≤ 0.05≤ 0.1≤ 0.2≤ 0.3PGD
Random191.790.186.779.271.886.7
CAT198.596.992.683.274.896.0
Random299.197.293.685.877.296.6
CAT299.898.896.287.980.097.2

Under the search, CAT stays ahead at every budget, by +5.9 pp at depth 1 and +2.6 pp at depth 2 at LPIPS ≤ 0.1, and by +3.0 pp and +2.8 pp at the loosest budget. Under PGD within the training ranges, CAT loses 2.5 pp at depth 1 against 5.0 pp for random augmentation. The widened ranges produce heavily distorted images (LPIPS 0.68, PSNR 9 dB), where CAT is ahead at depth 1 and behind at depth 2. We will add these results to the paper.

Thanks for the rebuttal and additional experiments. Most of my concerns are addressed in the rebuttal. However, the claim that adversarially optimizing attack strengths inevitably causes collapse has not yet been fully verified, as the authors did not directly evaluate constrained attack strength optimization in CAT, and the performance of the baseline methods may be influenced by specific experimental settings. I therefore maintain my current rating.

AuthorsOpenReview ↗

Response to Reviewer N6Wd

We thank the reviewer for the follow-up. We now evaluate adversarially optimized strengths directly inside CAT, both unconstrained and constrained, and rerun GAT and RoMark in settings that favor them. The results confirm our design: unconstrained optimization collapses robustness, constraints prevent this only partially, and sampling beats every constrained variant on every dataset.

Strength optimization in CAT. For every run below, we use the protocol of our previous table: we fine-tune the same PixelSeal checkpoint for 24 hours on 4 A4000 GPUs and report bit accuracy (%) averaged over SA-1B, CLIC-train, CLIC-test, DIV2K, and MetFaces. We keep CAT's controller and add, for each attack family, an MLP that outputs the strength from the controller's hidden state. We train it adversarially with the controller, through the differentiable attacks. Unconstrained, the MLP chooses the strength anywhere within the training range. To constrain it, we let the MLP choose only within a window of ±ε of the range width (ε = 0.5 or 0.25), placed around a random point of the range for each example. For JPEG quality, trained on 40 to 80, ε = 0.25 gives a window such as 50 to 70. The window moves so that training still covers the whole range, as a fixed window would simply shrink it. Depth-2 is the average over composed attacks of two steps:

Strengths in CATIdentityValueCompressionGeometricCombinedDepth-2
MLP, anywhere in range99.898.591.279.796.881.6
MLP within ±½ range window100.099.598.091.999.193.3
MLP within ±¼ range window99.999.198.693.898.295.2
Sampled (CAT)100.099.599.094.598.695.9

Unconstrained optimization collapses robustness to the hardest attacks. The MLP drives nearly every family to the hardest end of its range (e.g., JPEG quality 40, crop 0.5, perspective 0.5, contrast 2.0), so the decoder trains only on the most extreme strengths and generalizes poorly to the rest of the range. Compared with CAT, accuracy falls by 14.8 pp under geometric attacks, 7.8 pp under compression, and 14.3 pp under depth-2 compositions.

Constraints prevent the collapse only partially, and CAT remains the best. Each window recovers part of the loss, and the narrower the window, the higher the accuracy. Under geometric attacks and depth-2 compositions alike, CAT remains ahead by 2.6 pp of the ±½ window and 0.7 pp of the ±¼ window. CAT is ahead of every variant on all five datasets. Every step away from sampling costs accuracy. Sampling covers the whole range, while the controller supplies the adversarial pressure through the choice and order of attacks.

GAT and RoMark in settings that favor them. If the collapse came from our settings being too harsh, making the methods easier to train with should remove it. We do this in two ways. First, we weaken the adversary. GAT and RoMark apply their worst case to only half of each batch, and the other half is attacked at sampled strengths as in CAT, so the decoder always sees learnable examples. We also cut GAT's ℓ∞ budget from 4/255 to 1/255, a 4× smaller pixel perturbation. Second, we start GAT, with its published settings, from a decoder that is already robust: the model fine-tuned for 24 hours with random augmentation (95.2% under geometric attacks), instead of the 100-epoch checkpoint.

MethodEpochsIdentityValueCompressionGeometricCombinedDepth-2
GAT, half batch, ℓ∞ 1/2552249.949.949.949.949.949.9
GAT, from robust model2250.050.050.050.050.050.0
RoMark, half batch3679.177.068.756.176.758.5
CAT210100.099.599.094.598.695.9

GAT collapses even with these advantages. From the robust model, validation accuracy falls from 95.6% to 49.8% within the first epoch and never recovers, so the collapse comes from GAT's adversary itself, not from our settings or starting point. Softened RoMark no longer collapses, but in 36 epochs it reaches only 56.1% under geometric attacks and 58.5% under compositions, about 38 pp behind CAT. Part of the gap is cost: RoMark's grid search makes each step about 7× slower than CAT's, while CAT's controller adds almost no overhead.

We will add both tables to the paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Common Response to Reviewers and Area Chair

We sincerely thank all reviewers (DGsr, n6wd, U74N) for their time, thoughtful feedback, and constructive suggestions. We appreciate the reviewers' recognition of the practical importance of compositional attacks, the effectiveness of our adaptive training strategy, and the breadth of our evaluation.

We would like to provide a concise overview of why CAT addresses an important problem, what distinguishes our approach, and how our existing and newly added experiments support its contributions.

1. We address a practical and underserved threat: compositional attacks.

  • Real-world processing is compositional (n6wd, U74N). Shared content is recompressed, resized, cropped, color-edited, and reposted, so a watermark faces sequences of transformations rather than single ones. These operations need no attacker and are the most common cause of watermark loss.
  • Random augmentation misses the compositions that matter (n6wd). The compositions that break a watermark are rare, and sampling attacks uniformly seldom produces them during training.

2. We introduce an adaptive, compositional adversary with no inference cost.

  • Sequential, state-conditioned attack selection (DGsr, U74N). A recurrent controller chooses which transformation to apply next and in which order, conditioned on the image and the intermediate attacked state. We train it end to end with straight-through Gumbel-Softmax and entropy regularization.
  • Sampled strengths keep training stable (DGsr, n6wd). The controller supplies difficulty through the choice and order of attacks, while strengths are sampled from the training distribution, so every example remains learnable.
  • Efficient and training-only (DGsr, U74N). The controller is discarded after training, and a CAT training step (1.33 s) is cheaper than that of every prior adversarial method we compare (3.75 s to 16.15 s).

3. Extensive experiments demonstrate consistent gains.

  • Across models and settings (DGsr, n6wd, U74N). CAT improves over random augmentation on VideoSeal 0.0, VideoSeal 1.0, and PixelSeal, across image, video, and autoregressive watermarking, under compositional, out-of-distribution, and editing attacks, and on long attack chains that simulate reposting.
  • Realistic attacks outside the training library (Appendix F). Averaged over the three models, CAT improves bit accuracy by +2.9 pp under neural compression, +4.4 pp under VAE regeneration, and +11.1 pp under repeated video transcoding.

4. Additional rebuttal experiments further strengthen our findings.

  • Equal-compute comparison with prior adversarial methods (DGsr, n6wd, U74N). Fine-tuning the same PixelSeal checkpoint for 24 hours, CAT improves on Adversarial AutoAugment by +6.5 pp at depth 1 and +9.2 pp at depth 2. DADW reaches only 60.7% under geometric attacks against 94.5% for CAT, and GAT and RoMark collapse to chance accuracy.
  • Adversarially optimized strengths (DGsr, n6wd). Letting an MLP choose strengths inside CAT reduces accuracy by 14.8 pp under geometric attacks. Constraining it recovers only part of the loss, and sampling remains ahead on all five datasets. GAT still collapses with a 4× smaller perturbation budget and when started from an already robust model.
  • Unseen transformations (n6wd, U74N). CAT improves by +6.2 pp on ten transformation families outside the training library, and by +5.8 pp across 342 compositions of unseen and out-of-range transformations.
  • Model-aware attacks (n6wd). Against an attacker that searches for the most damaging transformation per image, CAT stays ahead at every LPIPS budget (+5.9 pp at LPIPS ≤ 0.1). Under PGD on strengths within the training ranges, CAT loses 2.5 pp against 5.0 pp for random augmentation.
  • Matched comparison with MBRS (DGsr). At 256 bits and matched PSNR, PixelSeal with CAT exceeds MBRS by +17.5 pp under geometric attacks and +23.1 pp across compositions, while MBRS falls to chance under two geometric steps at every watermark strength.
  • Error-correcting codes (DGsr). With a BCH code, CAT recovers the exact message for 48.6% to 91.0% of images at depth 1, against at most 2.8% for random augmentation, and transmits 54 bits per image against 22.

We believe CAT makes a timely contribution by turning the compositional nature of real-world processing into an adaptive training signal at no inference cost, supported by broad empirical evaluation and the additional evidence of robustness, efficiency, and practical deployability above.

We thank the reviewers and Area Chair again for their consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.