acceptodds
Under review as a conference paper at ICLR 2027

Learning what to protect: role-aware representations for image self-recovery against tampering

Abstract

With the rapid advancement of generative AI and image manipulation techniques, distinguishing authentic content from manipulated images has become increasingly challenging, raising growing forensic and security concerns. Existing image forensics methods can detect and localize manipulation but rarely recover the original content. Image self-recovery attempted to address this limitation by embedding recovery information into the image as a watermark. However, existing methods still face a longstanding trade-off between watermark imperceptibility and recovery fidelity. They typically protect information from the entire image without considering that different information contributes differently to recovery, resulting in redundant recovery information and a higher embedding burden. Hence, we propose ProCast, a novel self-recovery framework to tackle this fundamental challenge by jointly addressing three interdependent questions: what to protect, where to embed, and how to recover. ProCast learns to organize image information into components with different recovery roles and applies different protection strategies accordingly, reducing unnecessary recovery information and the embedding burden. It further adopts content-adaptive embedding, and reliability-based recovery to achieve high-fidelity recovery with improved watermark imperceptibility. Notably, although trained only on splicing, ProCast generalizes to unseen AIGC and conventional manipulations. Extensive experiments show that ProCast achieves higher watermark imperceptibility, more accurate tampering localization, and higher-fidelity recovery, improving the quality (PSNR) of watermarked image by 3.8 dB and the quality of recovered image by around 2-8 dB over the strongest existing self-recovery baselines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Self-Recovery, Image Forensic, Tamper Detection, Deep Watermarking

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_NmnqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper proposes a self-embedding fragile watermarking method that leverages DWT frequency information as a redundancy watermark for image tampering recovery. The framework consists of three main stages: (1) RHWG: Creates a "raw" watermark by separating the image into low-frequency (Base) and high-frequency (Enhancement) components. (2) CAWE: Utilizes two U-Net networks to process the original image alongside the backup features, generating a watermark modulation signal "R" that is subsequently added to the original image. (3) RMBR: Extracts the watermark, reverses the rearrangement, and localizes the manipulated regions.

Critical strengths
  • The experimental section is a key strength of this work. The proposed method demonstrates consistent, system-level performance gains across multiple datasets and various manipulation types.
Critical weaknesses
  • While the proposed architecture is technically sound and plausible, the overall conceptual novelty remains unclear. The framework appears to combine several established techniques and ideas without sufficient differentiation from existing literature: (1) Classic works such as Fridrich & Goljan [1] and Zhang et al. [2] have already extensively discussed self-embedding, image derived references across regions for recovery, and multi-region partitioning for image recovery. (2) Even compared to the references cited within the manuscript, the framework shares substantial similarities with "RelImage," which already embeds shuffled image information and refines the recovery process using enhancement.

[1] Fridrich & Goljan, “Images with Self-Correcting Capabilities,” ICIP 1999 [2] Zhang, Wang, Qian & Feng, “Reference Sharing Mechanism for Watermark Self-Embedding,” IEEE TIP 2010

Questions for authors
  1. Are the baseline comparisons conducted at a comparable parameter scale? For instance, Table 7 indicates that ProCast uses 194M parameters, whereas RelImage uses 9.94M parameters. While this does not invalidate the observed performance gains, it remains unclear how much of the improvement comes from the core methodology rather than a significantly larger network capacity and different optimization dynamics.

  2. Tables 4 and 5 indicate that the Base+Enhancement configuration yields an improvement of approximately 0.3 dB, while the majority of the system gain appears to derive from the embedding and fusion stages. Given these results, does the learned hierarchy provide a distinct advantage, or does the system primarily benefit from a compact reconstruction representation paired with embedding and fusion? The claimed contribution regarding "learning what to protect" requires further clarity in light of these ablation results.

  3. Several design choices appear arbitrary and require further justification. For example, (1) In the RMBR module, the two decoded views share the same enhancement. Does this imply that any corruption of the enhancement component will simultaneously degrade both views? (2) The locator module mixes spatial information. Will extraction errors potentially affect a broader area than the directly manipulated carrier region? (3) While the OACB reversible construction is mathematically motivated, the rationale for its architecture and protection allocation is quite limited.

  4. The loss functions fundamentally define the operations of the "Base," "Enhancement," and "Auxiliary" components. However, the explicit training loss computations are put to the Appendix. Since these losses serve as foundational support for the method, it is recommended to move this discussion into the main text.

5 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/3)

We sincerely appreciate your constructive comments. We have carefully addressed your concerns and revised the manuscript. We are happy to discuss further concerns you may have.


W1: The framework appears to combine several established techniques and ideas without sufficient differentiation from existing literature: (1) Classic works such as Fridrich & Goljan (ICIP, 1999) [R1] and Zhang et al. (TIP, 2010) [R2] have already extensively discussed self-embedding, image derived references across regions for recovery, and multi-region partitioning for image recovery. (2) Even compared to the references cited within the manuscript, the framework shares substantial similarities with "RelImage," which already embeds shuffled image information and refines the recovery process using enhancement.

A1: We apologize for the confusion. Here, we clarify our key conceptual novelty and explain how ProCast differs from and advances beyond R1, R2, and RegImage.

Our key conceptual novelty lies in shifting from predefined recovery representations to learned functional recovery roles that guide the overall self-recovery framework.

  • (1.1): Clarification of critical differences from [R1] and [R2]
    • Representation-level difference
      • Handcrafted vs. Learned role-based representation. [R1, R2] construct recovery information using handcrafted features and predefined spatial hierarchies. In contrast, ProCast learns functionally distinct components through role-specific reconstruction constraints.
    • Framework-level difference
      • Fixed vs. Role-guided adaptive embedding. [R1, R2] rely on predefined embedding rules, which cannot adapt embedding to local image content. In contrast, ProCast allocates redundancy based on learned recovery roles and adaptively adjusts wavelet embedding according to image content.
    • Technical-level difference
      • Pixel vs. wavelet embedding. [R1, R2] embed recovery information into fragile pixel-domain (eg., LSB-replacement), making it extremely vulnerable to common post-processing operations. In contrast, ProCast embeds recovery components into role-specific Haar wavelet subbands with content-adaptive weights, improving robustness while preserving image quality.                                           
  • (1.2): Clarification of technical differences from ReImage
    • Mechanism-level difference
      • Single vs. Role-based. Representation. ReImage learns a single representation of the entire target image, while we learn representations for distinct roles to distinguish information based on its contribution to recovery.
      • Single vs. Adaptively fused multiple recovery view. ReImage provides one shuffled recovery view. But we create two differently mapped and spatially separated backups. Our reliable and adaptive fusion of the two candidates allows each location to rely more strongly on the less corrupted backup.
    • Architectural difference. ProCast uses separate, task-specific embedding and extraction networks, providing greater architectural flexibility and enabling explicit allocation of recovery roles to distinct frequency subbands. Unlike ReImage, this design avoids the invertibility constraints and auxiliary-variable estimation associated with its invertible network.
    • The effectiveness of the above designs is demonstrated in the extensive experiments.

Q1: Are the baseline comparisons conducted at a comparable parameter scale?

A1:: We acknowledge that ProCast (194M) has more parameters than ReImage (9.94M). However, ProCast is not unusually large relative to other recent self-recovery methods, such as SSEM-HIR [1] (200M) and VeriFi [3] (157.12M).

  • Scale down our ProCast. To directly examine the influence of network capacity, we replaced ConvNeXt-Base with ConvNeXt-Tiny while keeping the remaining architecture unchanged (Table R6). This reduces the total parameter count from 194M to 74M, while producing only minor changes in WM PSNR and recovery PSNR. More importantly, ProCast-Tiny still outperforms ReImage by 4.23 dB in WM PSNR and 3.38 dB in Rec. PSNR. SSEM-HIR also has slightly more parameters than the full ProCast model but achieves lower performance. These results suggest that performance is not solely determined by model size.                                           

  • Scaling up ReImage. Similarly, we increased the width and depth of ReImage’s existing subnetworks while preserving its overall architecture and core mechanisms. This increased the parameter count from 9.94M to 72.59M, making it comparable in size to ProCast-Tiny (74M). Nevertheless, ProCast-Tiny (74M) still outperforms the scaled-up ReImage-L (72.59M), suggesting that our performance gains are not simply due to increased model capacity.

Table R6 is provided in Rebuttal by Authors (2/3)

AuthorsOpenReview ↗

Rebuttal by Authors (2/3)

Table R6. Capacity-scaling and parameter-efficiency comparison under the same evaluation protocol.

ModelExtractorParamsWM PSNR↑Rec. PSNR↑
SSEM-HIR—200M28.8327.30
ReImage—9.9M37.0730.81
ReImage-Large—72.59M33.9731.30
ProCast-TinyConvNeXt-Tiny74M41.3034.19
ProCastConvNeXt-Base194M41.9535.50

A2: Tables 4 and 5 indicate that the Base+Enhancement configuration yields an improvement of approximately 0.3 dB, while the majority of the system gain appears to derive from the embedding and fusion stages. Given these results, does the learned hierarchy provide a distinct advantage, or does the system primarily benefit from a compact reconstruction representation paired with embedding and fusion? The claimed contribution regarding "learning what to protect" requires further clarity in light of these ablation results.

A2: Thanks for your question. We conducted additional ablation experiments to further demonstrate the contribution and effectiveness of role-aware learning (Table 5).

  • A(2.1): Clarification of Table 4. The PSNR in Table 4 is an encoder-side component reconstruction metric. For each combination, the selected components are directly passed to the decoder, while the remaining components are set to zero. Therefore, no watermark embedding, manipulation, extraction, or backup fusion is involved. It can not reflect the gain that appears to derive from the embedding and fusion stages. Therefore, the approximately 0.3 dB difference in Base+Enhancement reconstruction does not imply that the performance gains come from embedding or fusion. To avoid further confusion, we have clarified this evaluation procedure in both the caption of Table 4 and the revised manuscript (page 10).                                                                                                                                                                                            
  • A(2.2): Does role-aware learning effectively learn what information should be protected for recovery Table 4 is intended to verify whether role-aware learning successfully identifies which information is essential for recovery. With the same representation capacity, Base + Enhancement achieves similar reconstruction quality with and without role-aware learning (25.199 vs. 24.879 dB). However, the key difference lies in how recovery information is distributed. Without role-aware learning, adding Enhancement improves PSNR by 6.911 dB (17.968 → 24.879 dB). With role-aware learning, Base alone achieves 24.816 dB, while Enhancement contributes only an additional 0.383 dB. These results demonstrate that role-aware learning concentrates the primary recovery information in Base rather than simply increasing reconstruction capacity.                                                                                                                                                                                       
  • A(2.3): Contribution of role-aware learning. We extended Table 5 with an RHWG end-to-end ablation. We removed role-aware learning while keeping the same backup allocation (B×2+E), payload size (34,560), and all other settings unchanged. The full model achieves comparable WM PSNR, while improving Ref. PSNR by 4.406 dB and Rec. PSNR by 4.358 dB . These results demonstrate that, under the same payload size, role-aware learning improves recovery fidelity without compromising watermark imperceptibility.                                                                                                                                                                                            Table 5. Ablation study of the key designs in RHWG, CAWE, and RMBR.
    ModuleAblation SettingWM PSNRRef. PSNRRec. PSNR
    RHWGw/o Role-aware learning40.30920.49630.196
    CAWEFixed subband38.74423.36732.482
    RMBRAverage fusion40.00221.54230.698
    ProCastFull40.60124.90234.554

AuthorsOpenReview ↗

Rebuttal by Authors (3/3)

Q3.1: In the RMBR module, the two decoded views share the same enhancement. Does this imply that any corruption of the enhancement component will simultaneously degrade both views?

A(3.1): Yes. Enhancement's corruption may affect both decoded views. However, this is precisely why we introduce role-aware learning and unequal protection. In our design, Base is trained to retain the primary recovery information, and Enhancement carries only less complementary information. Thus, even if Enhancement is corrupted, the main content can still be recovered from Base backup. Table 4 supports this design. Without role-aware learning, Enhancement carries important recovery information (Base → Base+Enhancement: ↑6.911 dB). With role-aware learning, its contribution becomes much smaller (↑0.383 dB). At the system level, Table 5 further shows that, under the same payload and (B×2+E) allocation, role-aware learning improves Ref. PSNR from 20.496 to 24.902 dB. This supports the effectiveness of aligning information roles with unequal protection. The recovery-critical information is concentrated in the duplicated Base, while the single-copy Enhancement is used only for refinement.

Q3.2: The locator module mixes spatial information. Will extraction errors potentially affect a broader area than the directly manipulated carrier region?

A(3.2): Local extraction errors may affect neighbouring features because the extractor and Locator use spatial context. However, this does not necessarily lead to errors spreading across the predicted localization mask. We have considered this issue in our design. Specifically, the Locator uses the pixel-aligned discrepancy (∣IR−Vavg∣|I_R - V_{\mathrm{avg}}|) to identify local inconsistencies between the received image and the extracted reference. Together with pixel-level BCE and Dice supervision, this helps limit the spread of localization errors. Across three general image manipulation scenarios, ProCast achieves an average F1 of 0.9892, AUC of 0.9996, and IoU of 0.9805, suggesting that spatial spreading of localization errors is limited in practice.


Q3.3: While the OACB reversible construction is mathematically motivated, the rationale for its architecture and protection allocation is quite limited.

A(3.3): OACB and the subsequent protection allocation address two different requirements. We provide a detailed explanation below.

  • OACB architecture. Our objective is not merely to extract compact features, but to learn how recovery information should be redistributed among components with different protection priorities. Conventional non-invertible encoders may lose information during feature extraction, making it difficult to distinguish intentional component allocation from irreversible encoding loss. In contrast, OACB preserves the complete latent representation when all components are retained. Its orthogonal mixing enables information exchange across channels, while additive coupling provides nonlinear learnable transformations with an exact inverse. Consequently, information loss occurs only when components are intentionally omitted, making the learned recovery roles interpretable and compatible with unequal protection.                                                                                              
  • Protection allocation. Our protection strategy follows the learned recovery roles of different components. As shown in Table 4, Base alone achieves 24.816 dB, while Enhancement and Auxiliary provide only small additional gains (↑0.383 dB and ↑0.060 dB). Therefore, we duplicate Base for stronger protection, transmit Enhancement once, and discard Auxiliary to reduce the payload size. This ensures that more important recovery information receives stronger protection.

Q4: Since these losses serve as foundational support for the method, it is recommended to move this discussion into the main text.

A4: We have moved the detailed loss formulations from the Appendix to the main paper (page 5).


Reference

[1] A symmetric self-embedding mechanism for high-fidelity image recovery against tampering. Liu et al. TIFS 2025
[2] Robust image self-recovery against tampering using watermark generation with pixel shuffling. Kim et al. CVPRF 2026   
[3] High-fidelity face content recovery via tamper-resilient versatile watermarking. Yu et al. ACM MM 2026

If you have any additional concerns, please do not hesitate to let us know. We are more than willing to address them and sincerely appreciate your valuable feedback and support.
Reviewer_3jc4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The proposed ProCast is a proactive watermarking method that finds tampered regions in an image and restores what was originally there. The paper organizes it around three questions: a) What to protect: An invertible projection built from Orthogonal Additive Coupling Blocks splits Haar coefficients into Base, Enhancement and Auxiliary parts, using asymmetric decoding constraints. b) Where to embed: Each part's residual goes into its own Haar subspace, weighted by content-adaptive softmax. c) How to recover: Diagonal and rotational backups from other regions are fused, weighted by how reliable the predicted tamper map says they are. The model is trained only on COCO splicing. On unseen manipulations and datasets, it reports better numerical results compared to prior work.

Critical strengths
  • The what/where/how structure is a sensible way to break down the imperceptibility vs. fidelity trade-off, and Figs explain the pipeline well.
  • Tough generalization test. The model is trained on splicing only, then tested on copy-move, SDXL, HD-Painter, E4S and Qwen-Image-Edit across five datasets.
Critical weaknesses
  • No adaptive-attacker analysis. The threat model only covers attackers who don't know the scheme. An attacker who knows them could edit a region and its backups together, or forge consistent payloads. The paper also doesn't test watermark removal, regeneration/purification attacks, or attacks on the Locator. If the mappings are actually keyed, please say so.
  • The experiments are thin in places. There are no seeds, variance or confidence intervals, and some test sets are small.
Questions for authors
  1. Are the cross-region mappings keyed or secret? How does ProCast hold up if an attacker edits a region along with its known backups, or uses diffusion-based regeneration?
  2. Were all baselines retrained on COCO at 256×256 with the same degradations? If not, how does that affect fairness?
  3. Can you add quantitative Qwen-Image-Edit results and test at higher resolutions (e.g., 512 or 1024)?
4 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/2)

Thank you for the constructive comments and recognition of our paper’s contributions. We have carefully addressed your concerns, revised the manuscript, and would be glad to discuss further if you have any follow-up questions.


W1: Threat model clarification, adaptation to stronger adaptive attacks, and watermark-level attacks.

A(w1): Thank you for raising these important security concerns. Following your suggestions, we have conducted additional experiments and expanded the discussion of stronger attack scenarios in Appendix A3.3.

  • (1.1) Threat Model. Indeed, following the convention adopted by existing self-recovery methods[1-4] cited in our manuscript, we target watermark-unaware semantic manipulations. Under the more challenging threat model you describe, most, if not all, existing self-recovery methods would likely fail. This limitation is also a major motivation for improving watermark imperceptibility, making watermarks less noticeable to potential malicious attackers. To avoid confusion, we have explicitly clarified our threat model on page 3 and its limitations on page 10 in the revised manuscript.                                                                                          
  • (1.2) Stronger adaptive attacks. Inspired by the reviewer, we have also considered the possibility to extend our work under the threat model that the attacker may know the scheme. Without modifying our existing framework, we can introduce key-based scrambling into the current backup mechanism to make backup locations less predictable. Specifically, we apply key-based scrambling to the rotational backup while keeping the diagonal backup unchanged, thereby increasing the difficulty of coordinated attacks targeting multiple backups ( Appendix A3.3 ).                                                                                              We have also conducted a preliminary experiment to verify that this modification does not substantially affect the performance of the original ProCast, as reported in Table R4. Importantly, introducing key-based scrambling does not alter the core design of ProCast, as the two backups continue to serve their original purpose of providing redundant recovery information. This extension makes coordinated attacks on known backup locations more difficult. However, it does not prevent payload forgery or attacks involving a compromised key. Addressing these threats requires additional authentication mechanisms and is left for future work.                                                                                               Table R4. Comparison of ProCast with and without key-based backup scrambling.
    MethodWM PSNRWM SSIM ↑Rec. PSNR ↑Rec. SSIM ↑
    ProCast - Keyed40.9620.982034.5300.9617
    ProCast - Original40.5300.982035.3300.9644

                                                                                        

  • (1.3) Watermark-level attacks. We further compare with SOTA methods under watermark-level attacks, e.g. global regeneration attacks, in Table R5. ProCast achieves better results than the compared methods. (discussed in Appendix A3.3 ).                                                                                               Table R5. Quantitative comparison under global regeneration after regeneration-aware retraining.
    MethodWM PSNR ↑Ref. PSNR ↑Rec. PSNR ↑
    IMUGE+ [1]32.0115.98717.219
    SSEM-HIR [2]30.3317.05819.230
    ReImage [3]35.3317.95220.120
    ProCast38.4218.05821.500

W2: * No seeds, variance, or confidence intervals, and some test sets are small.*

A(w2): We have conducted additional experiments in the Appendix Table 15 .

  • (2.1): Seed and variance. We evaluate training stability using three independent runs with random seeds 83, 210, and 1203. ProCast achieves 40.5390 ± 0.104 dB in WM PSNR, 34.4959 ± 0.100 dB in Rec. PSNR, and 0.9974 ± 0.0003 in F1 (mean ± standard deviation). We additionally report bootstrap 95% confidence intervals of [40.4791, 40.5975], [34.2619, 34.7323], and [0.9968, 0.9979], respectively, to quantify test-set sampling uncertainty. These results indicate consistent performance across training runs.                                           

  • (2.2): Test-set size. Our evaluation covers 8,500 images across five datasets: COCO (5,000), ValAGE (1,000), ILSVRC (1,000), CelebA-HQ (1,000), and ADE20K (500). All baselines are evaluated on the same test sets to ensure fair comparisons.

    For reference, the original ReImage [3] evaluates on 520 CelebA images and 1,047 ILSVRC images. Thus, our evaluation scale is comparable to or larger than that of prior work, providing a broad assessment across different datasets and manipulation types.

AuthorsOpenReview ↗

Rebuttal by Authors (2/2)

Q1: Are the cross-region mappings keyed or secret? How does ProCast hold up if an attacker edits a region along with its known backups, or uses diffusion-based regeneration?

  • A(1.1): The cross-region mappings in our threat model are assumed to be secret and are not publicly disclosed. However, even the attacker does not know the backup locations, a tampered region may still overlap with its backups by chance. Unlike existing methods [1–4] that rely on a single backup, ProCast employs multiple cross-region backups to improve recovery reliability. This design reduces the risk of complete information loss, as recovery remains possible even when one backup is corrupted. Nevertheless, if a region and all its backups are corrupted simultaneously, ProCast may also fail to recover that region. This is an open challenge for watermark-based self-recovery methods rather than a specific problem for ProCast, and we have acknowledged it in the manuscript.                                                                        
  • A(1.2): Additioanl ealuation on diffusion-based regeneration attack in Table R5. For a fair comparison, we included regeneration attacks in the training of all baseline methods and evaluated their robustness against a SOTA regeneration attack [5]. ProCast outperforms the strongest baseline, but despite these improvements, the overall recovery quality remains limited across all evaluated methods [1-3]. This highlights that global regeneration remains a common challenge for existing watermark-based self-recovery approaches, rather than a limitation specific to ProCast. Improving robustness against such attacks remains an important direction for future work.                                                                                          Table R5. Quantitative comparison under global regeneration after regeneration-aware retraining.
    MethodWM PSNR ↑Ref. PSNR ↑Rec. PSNR ↑
    IMUGE+ [1]32.0115.98717.219
    SSEM-HIR [2]30.3317.05819.230
    ReImage [3]35.3317.95220.120
    ProCast38.4218.05821.500

Q2: Were all baselines retrained on COCO at 256×256 with the same degradations? If not, how does that affect fairness?

A2: Thanks for your question.

  • For all general-purpose self-recovery methods, we adopted consistent experimental settings, retraining them on COCO at 256×256 with the same degradations.
  • For face-specific methods, we retrained them on CelebA-HQ using the same resolution (256×256) and degradations, and evaluated them on the CelebA-HQ test set. This setup allows face-specific methods to benefit from in-domain training data and be evaluated under favorable conditions, supporting a fair comparison. For VeriFi [4], the authors have not released the code or pretrained weights. However, VeriFi was originally trained on facial images at 256×256 with common degradations.
  • Therefore, we compared it using the same dataset and manipulation types as reported in its original paper to ensure a fair comparison. Despite being trained only on COCO, our ProCast still outperforms these face-specific methods in facial recovery (see Table R2), demonstrating strong cross-domain generalization.

Q3: Add quantitative Qwen-Image-Edit results and test at higher resolutions?

A3: Quantitative Qwen-Image-Edit results at higher resolutions 512 in Table R3. These results demonstrate that ProCast can effectively support different input resolutions.                                                                                               Table R3. Quantitative comparison under Qwen-Image-Edit manipulations.

MethodWM PSNR↑WM SSIM↑Rec. PSNR↑Rec. SSIM↑
ReImage-25636.61010.943333.74330.9284
ProCast-25641.06690.975835.47310.9547
ProCast-51242.35740.987335.67860.9588

Reference

[1] Learning to immunize images for tamper localization and self-recovery. Ying et al. TPAMI 2023
[2] A symmetric self-embedding mechanism for high-fidelity image recovery against tampering. Liu et al. TIFS 2025
[3] Robust image self-recovery against tampering using watermark generation with pixel shuffling. Kim et al. CVPRF 2026   
[4] High-fidelity face content recovery via tamper-resilient versatile watermarking. Yu et al. ACM MM 2026
[5] Visible image watermarks are provably removable using generative ai. Zhao et al. NeurIPS 2024

If you have any additional concerns, please do not hesitate to let us know. We are more than willing to address them and sincerely appreciate your valuable feedback and support.

Thank you for your responses and additional experiments. I have carefully reviewed all the responses. However, I will maintain my original score, as there seem to be many open problems and future work that need to be resolved: (1) the proposed method, as well as the baselines, fails against the diffusion-based regeneration attack released three years ago; and (2) when a region and all its backups are corrupted, the proposed method may fail.

AuthorsOpenReview ↗

Response to Reviewer 3jc4

We thank the reviewer for the further comments. We want to emphasize that our primary goal is to improve the trade-off between watermark imperceptibility and recovery fidelity, rather than to develop a method resistant to all stronger attacks.

Actually, these two problems are the same in nature: recovering content when the embedded backup information is substantially destroyed. The fundamental principle of all existing self-recovery methods is to embed backup information into images. Therefore, this is an open challenge for the entire field of watermark-based self-recovery, rather than a limitation specific to ProCast.

  • Diffusion-based regeneration. This attack falls outside the conventional threat model for watermark-based self-recovery. The regeneration attack (NeurIPS, 2024) used in our additional experiments can remove nearly 100% of the embedded watermark information, making self-recovery extremely challenging. Nevertheless, ProCast still achieves the best performance among the existing SOTA methods [1-3] under this attack.                                                                                          

  • Corruption of all backups. This concern assumes an even more extreme scenario. Under such a condition, neither traditional nor deep learning-based self-recovery methods can recover the original content from the embedded backup information. This assumption presupposes that all embedded recovery information has been completely destroyed, which is equivalent to imposing a failure condition on the self-recovery mechanism. Compared with single-backup methods, ProCast already takes a further step by employing multiple cross-region backups to reduce the risk of simultaneous corruption.

Overall, ProCast addresses a specific and important problem (trade-off) under a clearly defined threat model, with extensive experiments validating its core contributions. Even under diffusion-based regeneration attacks beyond this threat model, ProCast still outperforms the evaluated state-of-the-art methods. We think these broader challenges should be distinguished from our contributions to improving the imperceptibility–recovery trade-off.

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

This paper proposes ProCast, which is a proactive framework for image self-recovery against tampering. It combines role-aware hierarchical watermark generation, content-adaptive watermark embedding, and reliability-aware multi-view blind recovery to address “what to protect,” “where to embed,” and “how to recover.”

Critical strengths
  • The motivation to improve the trade-off between watermark imperceptibility and recovery fidelity is clear, and the framework is coherently organized.

  • The evaluation covers multiple datasets and tampering types, with encouraging PSNR and SSIM results.

Critical weaknesses
  • The experimental evidence for the central contribution of role-aware learning is incomplete. Table 4 supports the intended recovery roles of the components, but the comparison with and without role-aware learning reports only reference PSNR. Unlike the CAWE and RMBR ablations in Table 5, it does not report WM PSNR or Rec. PSNR. Its contribution to the final trade-off between watermark imperceptibility and recovery fidelity therefore remains unclear.

  • The baseline coverage could be better justified. For example, WRAE is evaluated in ReImage, while Watermark Anything is relevant to the localization comparison. Including these methods where applicable, or briefly explaining their exclusion, would clarify the scope of the evaluation.

  • The face recovery comparisons with RecoverMark and VeriFi, as well as the Qwen-Image-Edit results, are qualitative. (The stated lack of an available VeriFi implementation is acknowledged.) Additional quantitative results, where feasible, would strengthen these comparisons.

  • There are minor grammatical issues. For example, the Table 3 caption states, “While we achieve a better trade-off between the two objectives and additionally supports self-recovery.” Here, “supports” should be “support,” and the sentence beginning with “While” is incomplete.

Questions for authors

See weaknesses

2 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/2)

Thank you for your constructive comments and recognition of our paper's contributions. We have carefully addressed your concerns and revised the manuscript accordingly. We would be happy to discuss any follow-up questions.


Q1: The experimental evidence of role-aware learning is incomplete.

  • A(1.1): Contribution of role-aware learning. We extended Table 5 with an end-to-end ablation of RHWG. Specifically, we removed role-aware learning while keeping the backup allocation, payload size, and all other settings unchanged. As shown in Table 5, the full model achieves comparable WM PSNR while improving Ref. PSNR by 4.406 dB and Rec. PSNR by 4.358 dB. These results demonstrate that role-aware learning improves recovery fidelity without compromising watermark imperceptibility under the same payload size.                                                                                               Table 5. Ablation study of the key designs in RHWG, CAWE, and RMBR.
    ModuleAblation SettingWM PSNRRef. PSNRRec. PSNR
    RHWGw/o Role-aware learning40.30920.49630.196
    CAWEFixed subband38.74423.36732.482
    RMBRAverage fusion40.00221.54230.698
    ProCastFull40.60124.90234.554

                                                           

  • A(1.2): Clarification of Table 4. Table 4 is designed to examine how role-aware learning distributes recovery information among the learned components. For each combination, the selected components are directly passed to the decoder, while the remaining components are set to zero. No watermark embedding, manipulation, extraction, backup transmission, or recovery fusion is performed. Therefore, WM PSNR is not evaluated in this analysis, and Rec. PSNR is not applicable. To avoid confusion, we have revised the caption of Table 4.

Q2: Additional Comparison with W-RAE and Watermark Anything

  • A(2.1): Comparisons with W-RAE. W-RAE did not release its code and models. We thus included the W-RAE and W-RAE* reproduction results provided by ReImage [1]. Identical testing conditions are ensured for a fair comparison (details in Appendix Table 12). The resulting comparison is reported in Table R1.                                                                                                            Table R1. Comparison with W-RAE on the valAGE dataset.

    MetricW-RAEW-RAE*ProCast (Ours)
    WM PSNR↑38.0429.2741.01
    WM SSIM↑0.960.760.982
    WM LPIPS↓0.010.180.012
    SD Inpaint
    Rec. PSNR↑21.5823.6135.39
    Rec. SSIM↑0.560.630.9593
    Rec. LPIPS↓0.360.340.0285
    SDXL
    Rec. PSNR↑18.5323.6336.61
    Rec. SSIM↑0.480.630.95
    Rec. LPIPS↓0.420.340.028
    Splicing
    Rec. PSNR↑21.6123.0836.13
    Rec. SSIM↑0.570.630.95
    Rec. LPIPS↓0.360.350.03
  • A(2.2): Add comparison with Watermark Anything for the localization comparison in Table 3 as shown below.                                                    

    MetricWatermark AnythingProCast (Ours)
    VenueICLR'25—
    Recovery❌✅
    Avg. Imperceptibility
    PSNR↑38.6941.47
    SSIM↑0.97010.9787
    LPIPS↓0.01650.0153
    HD-Painter (COCO)
    F1↑0.900.99
    AUC↑0.99670.9999
    IoU↑0.830.98
    Copy-Move (COCO)
    F1↑0.1170.983
    AUC↑0.7730.999
    IoU↑0.0980.972
    SDXL (ADE20K)
    F1↑0.94350.9947
    AUC↑0.99690.9999
    IoU↑0.89650.9895
    E4S (CelebA-HQ)
    F1↑0.950.95
    AUC↑0.980.99
    IoU↑0.900.92
AuthorsOpenReview ↗

Rebuttal by Authors (2/2)

Q3: Add quantitative comparisons.

A3: We have added quantitative comparisons with RecoverMark and VeriFi in Table R2, and quantitative results for Qwen-Image-Edit in Table R3. Both tables have been included in Appendix A3.2.

  • A(3.1): Quantitative face recovery comparisons with RecoverMark and VeriFi. ProCast outperforms both face-specific methods in PSNR and SSIM. Notably, ProCast is trained on COCO (a general-purpose image dataset) and evaluated on CelebA-HQ (a face dataset), whereas the competing methods are trained on face datasets, demonstrating stronger cross-domain generalization and more reliable recovery.                                                                                               Table R2. Quantitative comparison with face-specific manipulation methods on CelebA-HQ.

    MethodVenueHD-Painter PSNR↑HD-Painter SSIM↑E4S PSNR↑E4S SSIM↑
    RecoverMarkCVPR'2623.400.76922.110.703
    VeriFiACM MM'2628.340.89225.120.839
    ProCast (Ours)34.550.94635.190.960
  • A(3.2): Quantitative comparisons with the SOTA method ReImage under Qwen-Image-Edit manipulations at different resolutions. ProCast consistently outperforms ReImage in watermark imperceptibility and recovery quality.                                                                                               Table R3. Quantitative comparison under Qwen-Image-Edit manipulations.

    MethodWM PSNR↑WM SSIM↑Rec. PSNR↑Rec. SSIM↑
    ReImage-25636.61010.943333.74330.9284
    ProCast-25641.06690.975835.47310.9547
    ProCast-51242.35740.987335.67860.9588

Q4: Minor grammatical issues.

A4: Table 3 caption has been revised, and we carefully proofread the manuscript.


References:

[1] Robust image self-recovery against tampering using watermark generation with pixel shuffling. Kim et al. CVPRF 2026  

If you have any additional concerns, please do not hesitate to let us know. We are more than willing to address them and sincerely appreciate your valuable feedback and support.

Comments on the paper

2 comments

Dear Reviewers,

Thank you for your time and effort in reviewing this paper!

The authors have now posted their responses. I would like to kindly encourage you to review their rebuttal and engage in the discussion, particularly if any of your concerns remain unresolved.

Your feedback is greatly appreciated and will help us reach a fair and well-informed decision.

Best regards,
Area Chair, ICLR 2027

AuthorsOpenReview ↗

General Response

Dear Reviewers, ACs, and SACs,

We deeply appreciate the insightful and valuable comments provided by all reviewers


We thank the reviewers for recognizing the significance of our research in addressing the trade-off between watermark imperceptibility and recovery fidelity, a fundamental yet insufficiently explored challenge in image self-recovery. ProCast addresses this challenge through role-aware recovery representation learning, which guides the overall self-recovery framework. We believe ProCast offers a promising direction for advancing image self-recovery.

Overall, we are encouraged by the reviewers' positive feedback, which highlights:

  • The motivation to improve the trade-off between watermark imperceptibility and recovery fidelity is clear and well justified. (Reviewer zznq)
  • The proposed what/where/how framework is coherent and well organized, with clear illustrations of the overall pipeline. (Reviewers zznq, 3jc4)
  • The method demonstrates strong generalization from splicing-only training to diverse unseen manipulations. ( Reviewer 3jc4)
  • The experimental evaluation is extensive and demonstrates consistent system-level improvements across multiple datasets and manipulation types. (zznq,Nmnq)

To address the reviewers' main concerns, we have conducted several additional experiments and analyses, including:

  • Table 5: added an ablation study to demonstrate the contribution of role-aware learning. (Reviewers zznq, Nmnq)
  • Table R1 and Table 3: extended the comparisons with additional baseline methods. (Reviewer zznq)
  • Tables R2 and Tables R3: provided quantitative comparisons to support the qualitative results. (Reviewers zznq, 3jc4)
  • Tables R4 and Tables R5: Evaluated under stronger attacks and provided further analysis. (Reviewer 3jc4)
  • Tables R6: added a parameter-efficiency comparison to examine whether the performance gains primarily arise from increased model capacity. (Reviewer Nmnq)
  • Table 15 (Appendix): added multi-seed evaluation to assess performance stability. (Reviewer 3jc4)
  • Clarified the purpose and implementation of Table 4. (Reviewers zznq, Nmnq)
  • Clarified the threat model (Reviewer 3jc4) and further explained the conceptual novelty and module-design rationale (Reviewer Nmnq).

All revisions in the paper are highlighted in blue. We sincerely appreciate the reviewers' constructive suggestions and remain committed to continually improving our work.

We address each reviewer's comments point by point below. We welcome further discussion and look forward to continued engagement. Thank you!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.