acceptodds
Under review as a conference paper at ICLR 2027

TrajBind: In-Trajectory Content Binding for Forgery-Resistant Diffusion Watermarking

Abstract

Existing diffusion watermarking methods mainly consider distortion and removal attacks, while forgery attacks pose a different threat: an adversary may transfer valid watermark evidence to content that was not generated by the watermarking system. We introduce TrajBind, an in-trajectory, content-bound watermarking method that ties the watermark pattern to the generated content. At a late denoising step, TrajBind extracts a compact content code from the model's predicted clean image and combines it with a secret key to determine the watermark pattern embedded during the remaining denoising steps. This design requires no auxiliary generation or external semantic models, while verification only requires the received image. We further provide a general formulation of diffusion-watermark forgery that covers both detection-level and exact-message attacks, together with a theoretical analysis that bounds detection-level forgery success in terms of content-code collision. Across five existing forgery attacks, TrajBind has lower worst-case forgery success than each of the 12 baseline watermarking methods while maintaining robust watermark detection and message recovery under benign transformations and removal attacks.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
diffusion models, image watermarking, watermark forgery, content-bound watermarking, generative model provenance, robust watermarking

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_y87aOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a watermark for diffusion models that is tied to the image it marks, so that a watermark copied from one image should not verify on another. A few denoising steps before generation ends, the model's prediction of the final image already shows its coarse layout. The method turns this layout into a short binary code. It combines this code with a secret key to decide the watermark pattern added in the remaining denoising steps, and for detection it recomputes the code from the received image. The authors prove an upper bound on forgery success that depends mainly on the chance that a forged image shares a code with a watermarked image.

Critical strengths
  1. Against the attacks the paper tests, the method's worst-case robustness to forgery is higher than that of all baselines.
  2. Unlike SemBind, the method does not generate an extra image or use a separate model to embed the watermark into the image.
Critical weaknesses
  1. The method can fail if the forger copies the watermark from a watermarked image with a similar coarse layout, and the paper does not test such an attack. In one of the tested attacks, every forgery whose layout happened to match passed verification (Table 10).
  2. A forger could also take a watermarked image and change its content, for example by inpainting a different object into one region. Because the watermark is tied only to coarse layout, it may still verify, and the paper does not test such edits.
  3. Close baselines are missing, including CSGuard [1], SEAL [2], and WIND [3].

References

[1] Jiewei Lai, Lan Zhang, Chen Tang, Pengcheng Sun, Zhaopeng Zhang, Yunhao Wang, and Hui Jin. CSGuard: Toward Forgery-Resistant Watermarking in Diffusion Models via Compressed Sensing Constraint. arXiv:2605.01479, 2026.

[2] Kasra Arabi, R. Teal Witter, Chinmay Hegde, and Niv Cohen. SEAL: Semantic Aware Image Watermarking. ICCV 2025.

[3] Kasra Arabi, Benjamin Feuer, R. Teal Witter, Chinmay Hegde, and Niv Cohen. Hidden in the Noise: Two-Stage Robust Watermarking for Images. ICLR 2025.

Questions for authors
  1. Suppose an attacker wants to forge the watermark into an image. They collect many watermarked images, say 1,000 to 100,000, pick the one whose coarse layout is closest to their image, and possibly nudge their image's layout toward it. How often does copying that image's watermark onto their image then pass verification?
  2. If a watermarked image is edited by inpainting one region, or by regenerating it with a changed prompt, does its watermark still verify?
  3. How many forged images are used for each forgery result?
6 comments
AuthorsOpenReview ↗

Response to Reviewer y87a (1/3): Threat Model and Coarse-Layout Forgery (W1/Q1)

We thank the reviewer for taking the time to review our submission. We address each weakness and question point by point below. All line, table, section, and appendix numbers refer to the revised PDF unless stated otherwise. We first clarify the motivation and the threat model defined in our work, then report the requested experimental results.

TrajBind addresses black-box watermark forgery under Sec. 3.2 and App. D. An adversary transfers watermark evidence from genuine images to a chosen target while preserving the target's appearance and meaning, as required by Sec. 3.1. The motivating threat is false attribution of malicious content to a legitimate provider or user. The adversary may use the public diffusion backbone and optimize surrogate models, while the secret key, content-code parameters and verifier remain private (Sec. 3.2, Lines 139–143). We have clarified these restrictions in the revised submission in Sec. 3.1 (Lines 111–126).

For W1/Q1, to ensure the requested attack lies within our defined threat model, our reference selection and gradient updates use public features, with no private content-code parameters or verifier feedback. The objective matches the selected reference's content features and watermark signal, with an LPIPS penalty to limit perceptual changes to the target.

Q1/W1. We added this result in Section 5.2 "Robustness to Coarse-Layout Forgery Attack" (Main paper, Lines 406–428, Tab. 2) and Section F.5 "Coarse-Layout Comparison with SemBind" in the revised submission (Appendix, Lines 2041–2065, Tab. 12).

Coarse-layout experiment. We used 100 target images and 2,000 watermarked references per dataset and variant on MS-COCO, DiffusionDB and DALL-E 3. This pool size lies within the 1,000 to 100,000 range the reviewer suggested; the limited rebuttal period prevented us from scaling this experiment to 100,000 watermarked references. Each entry is Min / Opt.

DatasetASRdet\mathrm{ASR}_{\mathrm{det}}ASRmsg\mathrm{ASR}_{\mathrm{msg}}ASRdet\mathrm{ASR}_{\mathrm{det}} at LPIPS ≤ 0.10
MS-COCO0.18 / 0.180.07 / 0.070.04 / 0.04
DiffusionDB0.21 / 0.200.14 / 0.130.07 / 0.05
DALL-E 30.12 / 0.130.10 / 0.110.04 / 0.06

The first two columns count a target as successful if any tested attack output passes detection (ASRdet\mathrm{ASR}_{\mathrm{det}}) or recovers the full payload (ASRmsg\mathrm{ASR}_{\mathrm{msg}}), without filtering by LPIPS. They therefore include successful outputs even when the attack substantially changes the target's appearance. The last column counts detection successes only when the accepted output has LPIPS ≤ 0.10, which limits its perceptual difference from the original target (Eq. 2). At LPIPS ≤ 0.10, TrajBind successfully prevents watermark forgery on 93–96% of target images across all three datasets and both variants, even when the attacker selects a watermarked reference with a similar layout.

Comparison with SemBind. We used 100 target images per dataset and method on MS-COCO and DALL-E 3. Each method uses 2,000 watermarked references per dataset. The table reports ASRdet\mathrm{ASR}_{\mathrm{det}} (Tab. 12).

MethodMS-COCO: LPIPS ≤ 0.10LPIPS ≤ 0.20No filterDALL-E 3: LPIPS ≤ 0.10LPIPS ≤ 0.20No filter
SemBind-GS0.500.961.000.820.960.99
SemBind-TR0.500.950.990.810.940.95
TrajBind-Opt0.040.140.180.060.090.13

Results. TrajBind-Opt achieves stronger robustness than the state-of-the-art SemBind against this layout-based forgery attack across all evaluated settings on both datasets. At LPIPS ≤ 0.10, our ASRdet\mathrm{ASR}_{\mathrm{det}} is 4% on MS-COCO and 6% on DALL-E 3, compared with 50% and 81–82% for SemBind, respectively. TrajBind strengthens forgery resistance by combining the secret key with a content code extracted within the image's own generation trajectory to determine the watermark pattern. Verification checks this pattern against candidate codes derived from the received image, so matching a reference's coarse layout or public features alone does not ensure successful watermark transfer.

Table 10 and the requested coarse-layout experiment serve different purposes. Table 10 of the original submission is Tab. 11 in the revised PDF (App. F.4, Lines 1998–2039). There, we use private system access to identify content-code collisions and examine their relationship with forgery success, providing empirical support for our theoretical analysis. This private access exceeds the capabilities granted to the adversary in our threat model and is used only by us as evaluators. The coarse-layout experiment requested by the reviewer tests an attack that selects a similar watermarked reference and modifies the target using public features.

AuthorsOpenReview ↗

Response to Reviewer y87a (2/3): Inpainting and Changed-Prompt Regeneration (W2/Q2)

For W2/Q2, the attacker edits a genuine watermarked image and attempts to retain valid evidence for the altered content. The edits use public models or region replacement, with the region, prompt and random seed fixed independently of verification results. The attacker has no access to the secret key, private content-code parameters or verifier, following the threat model in Sec. 3.2. We apply the deployed receiver only after saving the edits, to measure detection and exact-message success.

Q2/W2. Inpainting and regeneration with a changed prompt. We added this result in Section F.7 "Robustness to Inpainting and Regeneration with A Changed Prompt" in the revised submission (Appendix, Lines 2092–2113, Tab. 14).

We evaluated 300 genuine watermarked images per dataset and variant on MS-COCO, DiffusionDB and DALL-E 3. We tested the following editing techniques:

  • Inpainting: We use the Diffusers inpainting pipeline [1] with Stable Diffusion 2 to repaint a rectangular region covering 50% of the image using an object prompt. We preserve all pixels outside the rectangle.
  • Region replacement: We replace the same rectangle with the corresponding region from another image's unwatermarked companion, again preserving all pixels outside the rectangle.
  • Img2img with a changed prompt: We use SD 2.1-base with a mismatched caption and denoising strength 0.7. The Diffusers pipeline [2] implements SDEdit's noise-and-denoise procedure [3], conditioned here on the new prompt.

Each entry is Min / Opt. Lower is better.

TechniqueStrengthMS-COCO ASRdet\mathrm{ASR}_{\mathrm{det}}MS-COCO ASRmsg\mathrm{ASR}_{\mathrm{msg}}DiffusionDB ASRdet\mathrm{ASR}_{\mathrm{det}}DiffusionDB ASRmsg\mathrm{ASR}_{\mathrm{msg}}DALL-E 3 ASRdet\mathrm{ASR}_{\mathrm{det}}DALL-E 3 ASRmsg\mathrm{ASR}_{\mathrm{msg}}
Inpainting50% area0.170 / 0.1560.016 / 0.0130.180 / 0.1830.037 / 0.0230.160 / 0.1460.007 / 0.013
Region replacement50% area0.047 / 0.0500.004 / 0.0090.083 / 0.0800.010 / 0.0200.077 / 0.0700.000 / 0.003
Img2img, changed prompt0.70.117 / 0.1230.000 / 0.0000.156 / 0.1300.000 / 0.0000.117 / 0.1070.000 / 0.000

Results. TrajBind achieves strong forgery resistance across all three datasets and both variants, including edits covering half of the image. It rejects most detection-level forgery attempts and prevents exact-message forgery in at least 96.3% of cases, including all tested changed-prompt regenerations. For 50% region edits, detection-level success is 14.6–18.3% for inpainting and 4.7–8.3% for region replacement; changed-prompt regeneration at strength 0.7 reaches 10.7–15.6% detection-level success and zero exact-message success.

[1] Hugging Face. Inpainting. Diffusers documentation. Retrieved November 9, 2026.

[2] Hugging Face. Image-to-image. Diffusers documentation. Retrieved November 9, 2026.

[3] C. Meng, Y. He, Y. Song, J. Song, J. Wu, J.-Y. Zhu, and S. Ermon. SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations. International Conference on Learning Representations (ICLR), 2022.

AuthorsOpenReview ↗

Response to Reviewer y87a (3/3): Comparison with CSGuard, SEAL, and WIND (W3) and Sample Counts (Q3)

W3. Comparison with CSGuard, SEAL, and WIND. We added this result in Section F.6 "Additional Comparisons with SEAL and WIND" in the revised submission (Appendix, Lines 2068–2090, Tab. 13). The main paper points to it in Sec. 5 (Lines 384–385).

WIND [5] and SEAL [6] are now evaluated directly. We use 1,000 watermarked images on the MS-COCO dataset and the five forgery attacks of Sec. 5.1. The reported metric is ASRdet\mathrm{ASR}_{\mathrm{det}} (lower is better):

MethodAveragingLatent-MatchingImprintingRepromptingWMCopier
WIND1.0001.0001.0001.0001.000
SEAL0.8170.0400.1000.7500.100
TrajBind-Opt0.0000.0000.0000.0030.000

Results. TrajBind-Opt achieves stronger robustness than both SEAL and WIND under every evaluated forgery attack. WIND accepts all tested forgeries, and SEAL remains particularly vulnerable to Averaging and Reprompting. TrajBind-Opt successfully rejects all tested forgeries under four attacks and 99.7% under Reprompting.

We could not locate a publicly available official implementation of CSGuard [4], so we restrict our comparison with TrajBind to their conceptual differences (which we already compared in our submission). To clarify, we revised Section "Content-Bound Watermark Methods" (App. B.1, Lines 1128–1143): CSGuard's Appendix A.4.4 reports 93.75% Reprompting attack success when both the attacker and target use SD 2.1, compared with 28.12% when an SD 2.1 attacker targets SDXL. These results show that its reported forgery resistance depends on the attacker–target model pairing. The design also differs: CSGuard constrains the denoising and inversion trajectories through a secret compressed-sensing matrix while retaining content-independent watermark bits, whereas in TrajBind the content code selects the keyed signed permutation and enters the polar encoding, so both detection and exact-message decoding depend on the content.

Q3. Number of forged images per result. 1,000 prompts per dataset per forgery attack (Sec. 5.1, Line 394), which are clearly stated in the paper. The experiments added in this rebuttal use 100 target images per dataset for the coarse-layout attack (Sec. 5.2, App. F.5), 1,000 watermarked images for SEAL and WIND (App. F.6), and 300 genuine watermarked images per dataset and variant for the editing attacks (App. F.7).

[4] J. Lai, L. Zhang, C. Tang, P. Sun, Z. Zhang, Y. Wang, and H. Jin. CSGuard: Toward Forgery-Resistant Watermarking in Diffusion Models via Compressed Sensing Constraint. arXiv:2605.01479, 2026.

[5] K. Arabi, B. Feuer, R. T. Witter, C. Hegde, and N. Cohen. Hidden in the Noise: Two-Stage Robust Watermarking for Images. International Conference on Learning Representations (ICLR), 2025.

[6] K. Arabi, R. T. Witter, C. Hegde, and N. Cohen. SEAL: Semantic Aware Image Watermarking. IEEE/CVF International Conference on Computer Vision (ICCV), pp. 16196–16205, 2025.

Reviewer_9hoTOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper extracts a content code during the later stages of diffusion generation and uses it to determine a content-bound watermark pattern. Experiments show reduced success rates for several existing forgery attacks while maintaining strong robustness against watermark removal.

Critical strengths
  • Content information is extracted within the same generation trajectory, avoiding a separate full generation pass for content binding.
  • The evaluation covers multiple datasets and baselines. Component ablations and collision statistics provide supporting evidence for the content-binding mechanism.
Critical weaknesses
  1. The visual cost is non-negligible. MIN introduces noticeable distortion. Although OPT improves fidelity, the presented examples still exhibit visible texture and background changes. The evidence does not sufficiently support the strong descriptions of high fidelity and visually subtle changes.

  2. Computational costs and implementation details are insufficiently reported. OPT requires per-image optimization, while extraction involves multi-view encoding and candidate search. Runtime measurements and complete optimization settings are missing, making deployment costs difficult to assess.

  3. The security–robustness trade-off of content coding and candidate search is insufficiently evaluated. Collision statistics explain outcomes under existing attacks, but the paper does not systematically examine how candidate-list size affects legitimate recovery and forgery success. Adaptive attacks targeting content-code collisions within the stated threat model also remain insufficiently explored.

Questions for authors
  1. How much of the forgery-resistance advantage remains at matched visual distortion?
  2. What are the runtimes for MIN, OPT, message verification, and full decoding? What optimization budget is used for OPT?
  3. How do candidate-list size changes affect correct-code coverage, legitimate recovery, and forgery success?
7 comments
AuthorsOpenReview ↗

Response to Reviewer 9hoT (1/4): Image Quality (W1) and Matched Visual Distortion (Q1)

We thank the reviewer for the detailed feedback. We address each weakness and question point by point below. All line, table, figure, section, and appendix numbers in our responses refer to the revised PDF.

W1. Image quality

We would like to clarify that image quality is not presented as the main contribution, and the paper makes no claim of best image quality. Our main contribution is resistance to watermark forgery through in-trajectory content binding, supported by conditional theoretical analysis and experiments under the stated threat model. The fidelity discussion should be read in the context of OPT's improvement over MIN (Table 4 and Figure 4) and the reported trade-off between forgery resistance and visual distortion. Tables 1 and 4 show that several baselines achieve higher visual fidelity than TrajBind-OPT but are substantially more vulnerable to the evaluated forgery attacks. Meanwhile, TrajBind-OPT improves visual fidelity over MIN and achieves the best robustness to forgery attacks.

We have added paired unwatermarked and watermarked comparisons for 10 prompts from each of MS-COCO, DiffusionDB, and DALL-E3 in Appendix G (Lines 2430-2644, Figures 5-7), which complements the late-embedding discussion in the Image Quality section (Sec. 5.4, Lines 486-491). These comparisons show that TrajBind preserves the main semantic content and layout of the original unwatermarked images, while SemBind, WIND, and SEAL change content or composition in several displayed examples. This observation is consistent with TrajBind's embedding procedure: it inserts the content-dependent watermark during the later steps of the same generation trajectory, after coarse image structure has emerged. OPT further reduces the difference from the unwatermarked image while preserving watermark recovery.

Table 4 supports this comparison. Against unwatermarked images, OPT achieves SSIM 0.8067 and LPIPS 0.0911, compared with 0.6084/0.3060 for SemBind-TR and 0.1790/0.7327 for SemBind-GS. Together, these results support TrajBind's combination of semantic content preservation and strong resistance to the evaluated forgery attacks.

Q1. Forgery resistance at matched visual distortion

Could you clarify what you mean by "at matched visual distortion"? Specifically, which distortion metric and reference image should be used, and which method parameters should be varied to obtain the match?

The metrics in Table 4 measure a different kind of change for each method, because the methods modify different parts of the generation process. Initial-noise watermarking can change the sampled scene and composition. TrajBind embeds after the content code is obtained late in the trajectory, and OPT reduces the difference from its same-seed unwatermarked companion. Decoder-based and post-processing methods introduce changes through their own embedding procedures. This distinction is reflected in the evaluation protocols of prior work, such as Tree-Ring, Gaussian Shading, Stable Signature, and PRC. These methods do not share a single embedding-strength parameter that produces equivalent changes.

Therefore, equal LPIPS to each method's own unwatermarked image reflects different kinds of change: a different generated composition or altered texture within the same composition. FID and CLIP-FID compare image distributions and do not define a per-image distortion budget.

AuthorsOpenReview ↗

Response to Reviewer 9hoT (2/4): Computational Cost and OPT Optimization Budget (Q2/W2)

Q2/W2. Computational cost

We report generation, verification, and watermark decoding on an NVIDIA L40S at batch size 1 (20 samples). Receiver measurements use watermarked images. Verification checks a claimed watermark, while watermark decoding extracts its message. Runtime includes CPU work. Peak GPU memory includes resident model weights. We added this result in Appendix F.8, Computational Cost (Lines 2116-2159, Table 15).

OperationMethodMedian runtime (s/image)Peak GPU memory (GiB)
GenerationWIND1.384.98
SEAL2.7510.99
SemBind-GS4.238.10
Gaussian Shading3.223.03
PRC3.045.77
Spherical0.993.03
TrajBind-MIN1.193.17
TrajBind-OPT2.095.72
VerificationWIND0.704.85
SEAL0.9310.77
SemBind-GS0.747.86
Gaussian Shading0.702.81
PRC21.1511.37
Spherical0.762.81
TrajBind-MIN1.310.70
TrajBind-OPT1.330.70
Watermark decodingSemBind-GS0.747.86
Gaussian Shading0.702.81
PRC21.9211.37
Spherical0.762.81
TrajBind-MIN1.000.70
TrajBind-OPT1.980.70

Results. TrajBind's computational cost remains practical relative to the evaluated methods. MIN and OPT generate an image in 1.19 s and 2.09 s, respectively, compared with 2.75 s for SEAL, 4.23 s for SemBind-GS, 3.22 s for Gaussian Shading, and 3.04 s for PRC. TrajBind-OPT uses only two optimization iterations, adding approximately 0.9 s per image over TrajBind-MIN to improve visual fidelity while preserving watermark recovery. Both variants require only 0.70 GiB for verification and decoding, the lowest receiver GPU-memory usage among the evaluated methods. These costs accompany strong forgery resistance: OPT rejects all tested forgeries under four attacks and 99.7% under Reprompting (Table 1). Together, these measurements support an acceptable computational cost for combining content binding, improved fidelity, and strong resistance to the evaluated forgery attacks.

Optimization budget for OPT. The complete OPT settings are given in Appendix E.2 (Lines 1788-1795): Adam with learning rate 0.003 for 2 iterations, an LPIPS-AlexNet plus pixel-MSE distortion objective with λpix=1\lambda_{\text{pix}}=1, and projection onto ∥u∥2≤1.5 ∥u0∥2\|u\|_2 \le 1.5\,\|u_0\|_2 after every update. Gradients pass through the final five UNet steps and the VAE decoder.

AuthorsOpenReview ↗

Response to Reviewer 9hoT (3/4): Impact of Candidate-List Size (Q3)

Q3. Impact of candidate-list size

We evaluated TrajBind-MIN and TrajBind-OPT on MS-COCO at six verification list limits, Lver∈{1,16,64,256,1024,4096}L_{\text{ver}} \in \{1, 16, 64, 256, 1024, 4096\}. For each clean version, Rinse-100, and forgery attack, we use 3,000 watermarked images from MS-COCO. We added the experiment in the revised submission, Appendix F.9, Impact of Candidate-List Size (Lines 2160-2203, Table 16).

Each pair gives MIN / OPT. Clean and Rinse-100 columns report percentages (TPR at 0.1% FPR, and bit accuracy BACC; higher is better). Forgery columns report ASRdet\text{ASR}_{\text{det}} as fractions (lower is better).

LverL_{\text{ver}}Clean TPRClean BACCRinse-100 TPRAveragingLatent-MatchingImprintingRepromptingWMCopier
120.3 / 24.160.0 / 61.89.6 / 11.20.003 / 0.0020 / 0.00030 / 00 / 0.00030 / 0
1670.3 / 77.685.0 / 88.644.7 / 49.10.002 / 0.0020.0003 / 0.00030 / 00.0003 / 0.00030 / 0
6486.2 / 90.792.8 / 94.967.1 / 71.60.002 / 0.0020.0003 / 00 / 00.001 / 0.00030.001 / 0.004
25695.7 / 97.297.3 / 98.082.7 / 85.50.002 / 0.0020.001 / 0.00030 / 00.001 / 0.0010.001 / 0.002
1,02498.6 / 98.898.6 / 98.895.1 / 96.60 / 00 / 00 / 00.003 / 0.0030 / 0
4,09699.4 / 99.499.0 / 99.097.5 / 97.70 / 00.0003 / 0.0010 / 00.007 / 0.00670 / 0

Results. At Lver=1,024L_{\text{ver}}=1{,}024, TrajBind-MIN/OPT achieve clean TPRs of 98.6%/98.8% and Rinse-100 TPRs of 95.1%/96.6%, with Reprompting attack success of only 0.3% for both variants. Increasing LverL_{\text{ver}} to 4,0964{,}096 further improves genuine-image recovery but reduces robustness to Reprompting, raising attack success to 0.70%/0.67%. We therefore select Lver=1,024L_{\text{ver}}=1{,}024 as the default, balancing reliable watermark recovery with stronger resistance to Reprompting.

Reviewer_up3uOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses forgery attacks, which transfer evidence from a legitimately watermarked image to different content and thereby induce false provenance attribution. TrajBind computes a 24-bit content code from low-frequency DCT features, obtained with the existing VAE, of a late intermediate prediction within the same generation trajectory. The code and a secret key then determine a signed permutation that selects the watermark pattern. In 50-step DDIM, the code is fixed at the zero-indexed step jr=44j_r=44 (the 45th step), after which a length-2,048 signal obtained by polar-coding a 512-bit message is embedded. MIN uses a fixed perturbation and OPT an image-specific optimized perturbation; the receiver checks a candidate list around the code estimated from the received image. Under the stated assumptions, the theory decomposes detection-level forgery success into content-code reuse, chance correlation with unused codes, and the approximation error of the pseudorandom mapping.

The experiments evaluate 12 comparison methods, five existing forgery attacks, and distortion and regeneration attacks on three prompt datasets. On MS-COCO, the detection-level forgery success rate of OPT is 0.031 under Reprompting and 0 under the other four attacks, and the detection TPR after Rinse-100 is 0.992 (Tables 1–2). The generation backbone is SD 2.1 for all three datasets. These results support the method's performance under the reported attacks and settings, but they should be distinguished from a certification of worst-case risk against all admissible attacks.

Critical strengths

S1. Determining a content-specific pattern within the same trajectory, without a separate semantic model, is a useful design. The information tied to the final image is obtained during generation, the existing VAE is reused, and the receiver avoids UNet inversion. The paper distinguishes the roles of the content code, the ECC, and the key-based signed permutation, and appropriately does not claim that exact frequency separation is preserved through the subsequent denoising and VAE processing. This is a clear system- and algorithm-level contribution that integrates defense against provenance forgery into the generation process.

S2. Improvements against the reported existing attacks and the breadth of the evaluation support the design. In a comparison that includes both SemBind variants, OPT achieves a success rate of 0 under four of the five forgery attacks on MS-COCO and maintains a TPR of 0.992 under Rinse-100 (Tables 1–2). Results on the other two prompt datasets and the ablation removing content binding also support the utility of this component (Tables 7–9). The key-leakage experiment in F.1, the collision analysis in F.4, and the reported failure under rotation help clarify the scope of the results.

Critical weaknesses

W1. The link between the core security claim and the actual collision and threshold verification should be strengthened.

The collision term in Theorem 1 depends on the attack strategy. However, in the main setting, where the secret key, the content-code parameters ΘΦ\Theta_\Phi, and the verifier remain private, I could not find an evaluation in which the attacker searches for donors that directly target code overlap or trains a proxy extractor. The five existing attacks and the direct key-leakage attack in F.1 are present, but their settings differ from this question.

Moreover, the caption of Table 10 states that for multi-donor Averaging and WMCopier, only one donor's code is checked (p. 35). The 17,909 cases "without collision" therefore cannot be interpreted as the event C(xf)∩Robs=∅C(x_f)\cap R_{\mathrm{obs}}=\varnothing over the entire observed set, and the 21 detection successes among them do not empirically measure the theorem's non-collision condition.

For the threshold, the theorem requires a τ\tau that is independent of the keyed mapping family, yet the paper does not state which key and additional mappings were used in the experimental calibration. A.5 already states that using the deployed key requires adding to the used-code set and the mapping budget, so the 0.7%–1.2% exponential term in F.4 cannot be read as a proven bound for the full current protocol.

Re-aggregating over the full RobsR_{\mathrm{obs}}, evaluating collision-inducing attacks within the main threat model, reporting results by the number of observations and candidates, and providing the actual calibration key, samples, and mapping accounting together with an independent test FPR and confidence intervals would resolve this concern and make the security claim considerably stronger.

W2. The message-recovery claim and the practical advantages of OPT need additional verification and reproducibility details.

The abstract claims robust message recovery, but Tables 2 and 11–15, which cover benign, distorted, and removal settings, report only detection TPR. The raw sign agreement in Figure 2, computed without ECC or correction, and ASRmsg\mathrm{ASR}_{\mathrm{msg}} under forgery attacks do not substitute for the rate of exactly recovering the full 512-bit payload in this population. BER, exact recovery, CRC rejection, and wrong-payload returns should be reported separately.

For OPT, the definition of x0x_0, the additional per-image optimization, the selection of the candidate with minimal D(u)D(u) among those satisfying the recovery condition, and the fallback to u0u_0 on failure are already described. E.2 also provides the backbone, DCT region, CRC, HMAC, number of candidates, and SCL settings, and Appendix G gives the unit-norm Gaussian setting. The remaining gaps are the actual settings for the optimizer, number of iterations, learning rate, loss weights, transformation set, perturbation constraints, candidate generation, and recovery-condition check; the embedding and correction strengths and schedule; the scales and weights used for candidate ranking; and the data used to estimate μ\mu and the procedure for choosing τc\tau_c. These are hard to find in the Implementation Details referenced by A.4.

To judge reproducibility and practical advantage, the total number of UNet and VAE calls, wall-clock time, and memory for MIN/OPT and SemBind should be reported on the same hardware, including the actual computation path of x0x_0 and the repeated evaluation of the remaining denoising steps during optimization. The manuscript states that code will be released after acceptance (p. 10); no executable code was available for this review, so the results were not reproduced.

Questions for authors

Q1–Q3 are clarification requests whose answers bear directly on my rating; Q4 mainly concerns reproducibility.

Q1. Clarification of existing results and additional security verification. If Table 10 is re-aggregated using the intersection with all observed donor codes, what are the detection and message success rates in the true overall non-collision event? Can you evaluate collision-inducing attacks that use donor search or a proxy code extractor while keeping ΘΦ\Theta_\Phi and the verifier private, as in the main setting? If possible, results as a function of the number of observations and candidates would help my assessment.

Q2. Clarification of threshold calibration. Was the threshold calibrated with the deployed key? Please specify the number of benign samples, the calibration/test split, the code and mapping budget used, and the independent test FPR. If the deployed key was used, what is the bound after including the additional accounting described in A.5, and how do the results change with the key-independent threshold of Corollary 1?

Q3. Verification of the message-recovery claim. What are the exact recovery rate and BER of the full 512-bit payload for benign images and after each distortion and removal attack? Reporting CRC rejections, wrong-payload returns, and the candidate coverage of the original content code at Ldec=256L_{\mathrm{dec}}=256 separately would clarify the gap between detection TPR and message recovery.

Q4. Reproducibility and cost of OPT. Please provide the actual computation path of x0x_0; the optimizer, number of iterations, learning rate, transformation set, constraints, and loss weights; and the implementation settings of the candidate selection and fallback already described. What are the total UNet and VAE calls, time, and memory for MIN/OPT and SemBind, including reference generation and repeated tail evaluation? Please clarify what "one denoising pass" covers in terms of this total cost.

6 comments
AuthorsOpenReview ↗

Response to Reviewer up3u (1/3): W1/Q1, threat model and revised collision analysis (Table 10, now Table 11)

We thank the reviewer for the detailed feedback and for recognizing the value of obtaining the content code within the generation trajectory. We address each weakness and question point by point below. This comment covers W1/Q1; the next two cover Q2 and Q3, then Q4. Unless stated otherwise, all line, table, section, and appendix numbers refer to the revised PDF.

W1/Q1. Black-box forgery and the purpose of the collision table

TrajBind addresses black-box watermark forgery under the threat model of Sec. 3.2. The adversary transfers watermark evidence from genuine images to a chosen target while preserving the target's appearance and meaning, as required by Eq. 1 in Sec. 3.1. The motivating threat is false attribution of malicious content to a legitimate provider or user. The adversary may collect genuine watermarked images, use the public diffusion backbone and optimize surrogate models. The secret key, the content-code parameters ΘΦ\Theta_\Phi and the private verifier remain unavailable.

Table 10 of the submission (Table 11 in the revision) is a collision analysis carried out by us as evaluators, and it proposes no attack. We use private system access only to identify content-code collisions and to examine their relationship with forgery success, which gives empirical support for Theorem 1. In deployment under our threat model, the content codes are private and this access is unavailable to the adversary.

Revised collision analysis over the full observed reference set

We revised Appendix F.4, "Content-Code Collisions in Forgery Attacks" (Lines 1998-2039), to state this purpose explicitly. As you requested, the revised Table 11 checks the forged image's candidate list against all content codes in the attack's observed reference set (the donor set in your review). The submitted table checked only one reference code for Averaging and WMCopier. The revised analysis uses every reference in each Averaging pool and all 1,000 training references for WMCopier, so the "without collision" column now measures the event C(xf)∩Robs=∅C(x_f)\cap R_{\mathrm{obs}}=\varnothing.

Results.

EventForgeriesPass verificationRecover full target message
Without collision14,6127 (0.048%)0
With collision3,38891 (2.69%)37

A collision does not by itself ensure successful transfer: only 11/920 Averaging cases and 3/2,391 WMCopier cases with collisions pass verification. Reprompting remains the main observed failure mode: all 71 of its collision cases pass verification, and 37 recover the target message. These results support the theorem's separation of collision and non-collision events under the evaluated attacks. In this experiment, the thresholds are selected without the deployed key, so the setting matches the fixed-threshold condition of Theorem 1 (see our response to Q2).

Requested collision-inducing attacks

Q1 requests reference search or a proxy code extractor while keeping ΘΦ\Theta_\Phi and the verifier private. Under Sec. 3.2, the adversary cannot observe the true content codes, cannot obtain private code labels for training, and cannot use private-verifier feedback to select references or attack outputs. Appendix D adds that, under the verifier-free policy, the adversary cannot select attacks or decide when to stop based on private verification results.

Search based on public image features is permitted, and the revision evaluates it. The coarse-layout forgery attack in Sec. 5.2 (Table 2, Lines 406-428) uses public models to select, among 2,000 watermarked references per dataset, a reference with a layout similar to the target, and then optimizes the target toward that reference. Under the perceptual limit LPIPS ≤0.10\le 0.10, detection-level success is 4-7% across the three datasets and both variants. Appendix F.9 (Table 16, Lines 2160-2203) reports results as a function of the candidate-list size.

Beyond this attack, the request leaves the search criterion and the proxy extractor's training signal unspecified. Could you clarify how references would be selected, or the proxy trained, using only information available under the stated black-box threat model? Direct selection by true code overlap would require private information. We need this clarification to fix the inputs and success criterion of any further experiment.

AuthorsOpenReview ↗

Response to Reviewer up3u (2/3): Q2 threshold selection, W2/Q3 full 512-bit message recovery

Unless stated otherwise, all line, table, section, and appendix numbers refer to the revised PDF.

Q2. Threshold calibration

We read "threshold calibration" as the threshold-selection procedure described in our paper. We use the term "threshold selection" throughout and introduce no separate calibration procedure.

To answer your question directly: our threshold selection does not use the deployed key or its associated mappings. The threshold τ\tau is fixed independently of the deployed keyed mapping family, as required by Theorem 1. We select thresholds from verification scores on real images and unwatermarked companion images on MS-COCO, and on unwatermarked reference images on DiffusionDB and DALL-E3, using a separate key sampled independently of the deployed key. The selected thresholds are then held fixed during evaluation under the deployed key.

The statement in Appendix A.5 describes an optional setting in which the deployment key is used for threshold selection. In that setting, the codes and mappings used during threshold selection must also be included in the theoretical accounting. Our threshold selection does not use the deployment key, so this additional accounting does not apply. The mapping budget in Theorem 1 therefore remains at most N+LN+L, where NN is the number of observed watermarked generations and LL is the verification candidate limit. To make this distinction explicit, we revised the sentence in "Watermark Extraction and Verification" (Appendix A.5, Lines 975-980) to state that the additional accounting applies only if the deployment key is used for threshold selection.

W2/Q3. Detection and full 512-bit message recovery

The main paper reports [email protected]%FPR to compare watermark detection across methods at the same false-positive operating point. This metric applies to both multi-bit and zero-bit watermarking methods. Zero-bit methods detect watermark presence without encoding a recoverable message, so exact payload recovery is undefined for them. The results below complement that detection comparison by measuring TrajBind's full message recovery.

We evaluated 1,000 watermarked images per dataset, variant and condition with a decoding candidate limit of Ldec=256L_{\mathrm{dec}}=256. The table reports the fraction of images for which all 512 original bits are recovered. Each entry is MIN / OPT, and higher is better.

ConditionMS-COCODiffusionDBDALL-E3
Clean0.972 / 0.9830.980 / 0.9930.977 / 0.983
Blur0.942 / 0.9550.957 / 0.9650.948 / 0.961
Brightness0.976 / 0.9870.989 / 0.9930.972 / 0.981
Compression0.952 / 0.9690.981 / 0.9910.939 / 0.957
Contrast0.951 / 0.9690.969 / 0.9950.933 / 0.953
Noise0.944 / 0.9560.978 / 0.9890.934 / 0.953
VAE regeneration0.943 / 0.9570.936 / 0.9420.925 / 0.941
Diff-400.931 / 0.9450.934 / 0.9470.918 / 0.932
Rinse-200.921 / 0.9300.926 / 0.9450.915 / 0.929
Rinse-600.905 / 0.9090.912 / 0.9170.911 / 0.923
Rinse-1000.868 / 0.8980.894 / 0.9050.860 / 0.875

Exact recovery of the full 512-bit message stays at or above 0.86 for both variants in every tested condition, including the strongest removal attack (Rinse-100). We added this result in Appendix F.10, "Full 512-bit Message Recovery" (Table 17, Lines 2205-2238).

AuthorsOpenReview ↗

Response to Reviewer up3u (3/3): W2/Q4, OPT implementation details and computational cost

Unless stated otherwise, all line, table, section, and appendix numbers refer to the revised PDF.

W2/Q4. OPT implementation and total cost

We added the requested implementation details for TrajBind-OPT in Appendix E.2, "TrajBind Implementation Details" (Lines 1759-1767 for the content-code parameters and candidate ranking, Lines 1780-1796 for the embedding strengths and OPT optimization settings).

For the computational cost of TrajBind-OPT, we measured generation, verification and watermark decoding on an NVIDIA L40S at batch size 1 (20 samples). Receiver measurements use watermarked images. Verification checks a claimed watermark, while watermark decoding extracts its message. Runtime includes CPU work. Peak GPU memory includes resident model weights. We added this result in Appendix F.8, "Computational Cost" (Table 15, Lines 2116-2159).

OperationMethodMedian runtime (s/image)Peak GPU memory (GiB)
GenerationWIND1.384.98
SEAL2.7510.99
SemBind-GS4.238.10
Gaussian Shading3.223.03
PRC3.045.77
Spherical0.993.03
TrajBind-MIN1.193.17
TrajBind-OPT2.095.72
VerificationWIND0.704.85
SEAL0.9310.77
SemBind-GS0.747.86
Gaussian Shading0.702.81
PRC21.1511.37
Spherical0.762.81
TrajBind-MIN1.310.70
TrajBind-OPT1.330.70
Watermark decodingSemBind-GS0.747.86
Gaussian Shading0.702.81
PRC21.9211.37
Spherical0.762.81
TrajBind-MIN1.000.70
TrajBind-OPT1.980.70

Results. TrajBind's computational cost remains practical relative to the evaluated methods. MIN and OPT generate an image in 1.19 s and 2.09 s, respectively, compared with 2.75 s for SEAL, 4.23 s for SemBind-GS, 3.22 s for Gaussian Shading, and 3.04 s for PRC. TrajBind-OPT uses only two optimization iterations, adding approximately 0.9 s per image over TrajBind-MIN to improve visual fidelity while preserving watermark recovery. Both variants require only 0.70 GiB for verification and decoding, the lowest receiver GPU-memory usage among the evaluated methods. Receiver runtime is 1.31 to 1.33 s for verification and 1.00 to 1.98 s for decoding, compared with 0.70 to 0.93 s for the baselines other than PRC, and it is more than 10 times lower than PRC's 21.15 s and 21.92 s.

These costs accompany strong forgery resistance: on MS-COCO, OPT rejects all tested forgeries under four attacks and 99.7% under Reprompting (Table 1). Together, these measurements support an acceptable computational cost for combining content binding, improved fidelity, and strong resistance to the evaluated forgery attacks.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Summary of Contribution, Rebuttal, and Revisions (to the AC and all Reviewers)

We sincerely thank the Area Chair and Reviewers y87a, 9hoT, and up3u for their time and for the detailed, constructive feedback, which has substantially strengthened our paper. This comment summarizes our contribution and how we addressed each reviewer concern during the rebuttal. All line, table, section, and appendix numbers refer to the revised PDF.

1. Contribution

TrajBind is a content-bound diffusion watermark that ties the watermark pattern to the generated image through a content code taken from the model's own denoising trajectory, with no auxiliary time-consuming generation and no external semantic model. Across five existing forgery attacks, TrajBind has lower worst-case forgery success than each of 12 baselines, supported by a theoretical bound on forgery success in terms of content-code collision, while preserving the semantic content of the original unwatermarked image and keeping detection and message recovery robust.

2. The revision and rebuttal addressed the reviewers' concerns

We responded to every weakness and question from the three reviewers.

New requested attacks. Beyond the 5 forgery attacks, 6 distortion attacks, 5 removal attacks, and 1 white-box attack evaluated in our submission, we additionally evaluated 4 attacks requested by the reviewers: the coarse-layout forgery attack (y87a W1/Q1, 9hoT W3, up3u W1/Q1; Sec. 5.2, Tab. 2; App. F.5, Tab. 12) and 3 editing attacks, namely inpainting, region replacement, and changed-prompt regeneration (y87a W2/Q2; App. F.7, Tab. 14), which go beyond the forgery setting studied in the submission. TrajBind rejected at least 93% of coarse-layout forgeries at LPIPS ≤ 0.10 and at least 81% of forgeries under the three editing attacks, on all 3 datasets and for both variants. Under the same coarse-layout attack, the closest prior work, SemBind, rejected only 50% of forgeries on MS-COCO and fewer than 20% on DALL-E 3.

New requested baselines. Beyond the 12 baselines compared in our submission, we additionally evaluated 2 baselines requested by the reviewers, SEAL and WIND, under the 5 forgery attacks (y87a W3; App. F.6, Tab. 13). TrajBind-OPT rejected at least 99.7% of forgeries under every attack, while WIND accepted every forgery under all five attacks and SEAL accepted up to 82%. We could not locate a public implementation of the third requested baseline, CSGuard, so we revised its conceptual comparison (App. B.1).

New requested analyses. Beyond the ablations, collision statistics, and image-quality results in our submission, we added 6 analyses requested by the reviewers:

  • Collision analysis over the full reference set (up3u W1/Q1; App. F.4, Tab. 11). Without a content-code collision, fewer than 0.05% of forgeries pass verification and none recover the message, which agrees with Theorem 1.
  • Threshold selection (up3u Q2; App. A.5). Thresholds are selected with a key independent of the deployed key, so the fixed-threshold condition of Theorem 1 holds.
  • Full 512-bit message recovery (up3u W2/Q3; App. F.10, Tab. 17). TrajBind recovers the entire message from at least 97% of clean images and at least 86% of images even under the strongest removal attack at its highest intensity (Rinse-100).
  • Computational cost and OPT settings (9hoT W2/Q2, up3u W2/Q4; App. F.8, Tab. 15; App. E.2). TrajBind generates an image in about 1 to 2 s, faster than SemBind, SEAL, Gaussian Shading, and PRC, and verifies it in about 1.3 s. TrajBind has the lowest receiver memory of all compared methods.
  • Candidate-list size (9hoT W3/Q3; App. F.9, Tab. 16). At the default list size, TrajBind detects at least 98% of clean watermarked images while rejecting at least 99.7% of forgeries under all five attacks.
  • Visual cost (9hoT W1/Q1; App. G, Figs. 5 to 7; Tab. 4). New paired comparisons on all 3 datasets show that TrajBind keeps the content and layout of the unwatermarked image, while SemBind, WIND, and SEAL change the content or composition in several of the displayed examples. TrajBind achieves the lowest worst-case forgery success among all 14 compared methods across the 5 forgery attacks while preserving the semantic content of the original unwatermarked image.

Dear Reviewers,

This is a gentle reminder to review the authors’ rebuttal and actively participate in the discussion on OpenReview. Your timely input is important for having an interactive discussion and reaching a well-informed assessment.

Thank you for your time and contribution!

Best,

Your AC

Dear all reviewers,

The authors have responded to your concerns for a while. Please go through the response below and the revised submission to check whether your concerns have been adequately addressed, and if there are any follow-up questions where applicable. Thank you!

Best, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.