acceptodds
Under review as a conference paper at ICLR 2027

RemTraceNet: Few-Shot Forensic Detection of Invisible Watermark Attacks

Abstract

Removing an invisible watermark and concealing the forensic evidence are distinct objectives: successfully disrupting the embedded watermark does not imply that the removal process is forensically undetectable. When verification fails, removal traces can provide complementary evidence for provenance and ownership verification, whereas their absence leaves the cause of the failure ambiguous. Existing methods are typically evaluated by watermark suppression and perceptual quality, while forensic stealth is rarely considered. We therefore study watermark-attack-specific few-shot forensics: for each known pipeline, a specialist can separate its outputs from paired clean and unattacked watermarked controls. Separate Attack-vs-Clean and Attack-vs-Watermarked evaluations prevent watermark-presence shortcuts. Image-aligned and prompt-matched controls are used for post-hoc and generator-integrated schemes, respectively. In this work, we introduce RemTraceNet, which fuses constrained residuals, local relations, FFT/Haar statistics, and block-DCT evidence at native resolution. Across 23 removal pipelines and 10 watermark configurations, we evaluate native and inputs. With 100 attacked training images per pipeline, the three-seed TPR@1%FPR, macro-averaged over attacks and watermark configurations, ranges from 82.75% to 88.17% across resolutions and control types. Under the condition of same labels and protocol, RemTraceNet outperforms retrained SRNet, ZhuNet, and SiaStegNet baselines by 10.80–15.68 percentage points. Extensive experimental results show that erasing a watermark and erasing evidence of its removal are distinct challenges, and that removal traces remain learnable under limited supervision.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
invisible watermarking, watermark attacks, watermark removals, watermark attacks forensics, forensic detection

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_AYthOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents RemTraceNet, a few-shot forensic detection framework integrating complementary spatial and frequency-domain features for detecting residual traces introduced by invisible watermark removal attacks.

After careful examination, I have serious concerns regarding the integrity and quality of this submission, and I recommend a clear rejection.

Critical strengths

This paper explores a meaningful and underexplored forensic problem and provides extensive experiments across diverse attack pipelines and watermark configurations, demonstrating consistent improvements over retrained steganalysis baselines.

Critical weaknesses
  • The manuscript is missing the mandatory AI usage statement required for this year’s ICLR submission.
  • Figure 1 appears to be an AI-generated architecture diagram.
  • Line 173: The learned 5!×!5 Bayar filter fixes its center coefficient ...
  • The loss components in Equation (5) are merely described in Table 1, without explicit mathematical formulation. This presentation style is non-standard and unlikely to originate from human writing.
  • The work also exhibits the stereotypical seed 42 issue common in AI-generated manuscripts.
Questions for authors

I am puzzled by the quality of this submission. I wonder whether this work has undergone internal review by experienced researchers prior to submission, and why the authors consider this manuscript to satisfy ICLR’s acceptance standards.

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

This paper examines an interesting question in invisible watermark forensics: even when watermark verification fails after an attack, the attack pipeline may leave detectable traces. The authors formulate an attack-specific few-shot detection task using triplets of clean, unattacked watermarked, and attacked images. By treating both clean and watermarked images as negative examples, the training protocol aims to reduce the risk of detecting watermark presence rather than attack traces. The proposed RemTraceNet combines residual features, local relations, FFT/Haar statistics, and block-DCT evidence. It also uses global-local aggregation and a task-specific training objective. The experiments cover 23 attack pipelines and 10 watermark configurations, five of which are unseen during training. The authors evaluate the model at two image resolutions and compare it with three retrained steganalysis baselines. Selected ablation studies also support the benefits of several components and training strategies. The results suggest that attack-related traces can be detected with limited training data. Such traces may provide additional forensic evidence when direct watermark verification fails, potentially supporting provenance and copyright verification.

Critical strengths
  1. A meaningful forensic perspective. The paper explores a meaningful direction for invisible watermark forensics. Even when direct watermark verification fails, traces left by the processing pipeline may still provide useful evidence. This offers a complementary way to investigate watermark removal and has potential value for provenance and copyright verification.

  2. Broad experimental coverage. The authors design a triplet-based training protocol using clean, unattacked watermarked, and attacked images. Treating both clean and watermarked images as negative controls helps reduce the risk of detecting watermark presence rather than attack traces. The experiments cover 23 attack pipelines and 10 watermark configurations, including five unseen during training. Evaluations across two resolutions, multiple training sizes, and three random seeds, together with ablations, provide useful empirical evidence for the proposed approach.

Critical weaknesses
  1. Gap between the research motivation and the actual detection task The paper aims to provide additional forensic evidence when watermark verification fails. However, the proposed detector only identifies whether an image has undergone a specific attack pipeline, regardless of whether the watermark was actually removed. The authors also acknowledge this limitation in Section 4.1. The model may therefore learn general processing artifacts rather than traces specific to watermark disruption. A crucial experiment is missing: applying the same processing pipeline to unwatermarked images. Without this comparison, it remains unclear how the detected traces differ from those captured by conventional processing-history detectors. Experiments do not clearly establish whether RemTraceNet provides any watermark-removal-specific evidence beyond general processing artifacts. Furthermore, the method requires a separate detector for each known attack pipeline and does not evaluate generalization to unseen attacks. Although this is consistent with the stated task, it limits the practical applicability of the approach.

  2. Unclear novelty and insufficient comparisons The research motivation overlaps with prior work, such as The Forensic Cost of Watermark Removal (Evennou and Kijak, 2026, arXiv:2604.25491). Such or similar works are scarcely discussed or cited. RemTraceNet also combines several established forensic techniques. Although the ablations support their effectiveness, the additional contribution over existing methods remains unclear. The main baselines, SRNet, ZhuNet, and SiaStegNet, were originally designed for steganalysis. More relevant manipulation detectors, such as Bayar and Stamm (2018), which is already cited in this work, are missing. Moreover, Table 4 shows that the performance gains are largely concentrated on conventional distortions. RemTraceNet performs worse than SiaStegNet on VAE-based compression and achieves only a modest improvement on diffusion regeneration. These results make it difficult to establish a clear advantage over existing forensic methods, particularly for more complex attacks.

Questions for authors
  1. What exactly does RemTraceNet learn? Does RemTraceNet learn general traces left by an attack pipeline, or traces specifically related to watermark disruption? Could the authors test the detector on unwatermarked images processed by the same attack pipeline? If the model mainly learns general processing traces, what additional value does it provide for watermark removal forensics compared with existing processing-history detectors?
  2. What is the advantage over existing forensic detectors? Can the authors compare RemTraceNet with existing image manipulation or processing-history detectors under the same training and evaluation protocol? Such a comparison would help clarify whether the proposed architecture and training strategy offer meaningful improvements over existing forensic methods, especially for advanced attacks.
Reviewer_MSBHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RemTraceNet, a few-shot forensic detection framework for identifying traces left by invisible watermark-removal operations. The method combines residual, relational, spectral, and block-DCT features with global-local fusion and Top-k aggregation. Experiments across 23 attack pipelines and 10 watermark configurations show promising improvements over existing steganalysis baselines.

Critical strengths
  • The paper addresses a relevant problem by investigating whether watermark-removal operations leave detectable forensic traces even when the watermark is no longer recoverable.

  • The evaluation covers 23 processing pipelines, 10 watermark configurations, and two image resolutions. The inclusion of unseen watermark configurations and different training sizes is useful.

  • The combination of spatial, residual, and frequency-domain representations is reasonable for forensic detection. The reported improvements over existing steganalysis baselines are encouraging.

Critical weaknesses
  • The novelty of the work is not entirely clear to me. Recent papers such as [1] and [2] have already explored forensic traces left by watermark-removal attacks. The second paper also investigates attack-specific detection. I think these works should be discussed, and the authors need to explain more clearly what is new in their formulation and approach.

  • Most of the components used in RemTraceNet, including constrained convolutions, SRM filters, FFT/Haar features, DCT statistics, and feature fusion, are already established in image forensics. While combining them may be useful, I am not convinced that the architecture offers a substantial methodological contribution. It would help to explain what is fundamentally different from existing forensic detection architectures.

  • One of my main concerns is whether the model actually learns traces related to watermark removal or simply recognizes image-processing artifacts. Section 4.1 mentions that a positive label does not necessarily mean the watermark was successfully removed. Since operations such as JPEG compression, Gaussian blur, and denoising also leave detectable artifacts on ordinary images, the current experiments do not clearly separate these two possibilities.

  • The generalization setting is quite restrictive. The authors train a separate detector for every known attack pipeline, but in practice, we may not know which operation was used to remove the watermark. I would like to see how the model performs on unseen attacks or when an image undergoes multiple processing operations. The experiments are also limited to MS-COCO, making it difficult to judge performance on other image distributions.

  • The presentation could also be improved. Several labels and descriptions in Figure 1 are too small to read comfortably. Similarly, the image panels in Figure 2 are quite small, making it difficult to observe the forensic differences. Figure 3 also has crowded labels. I noticed a few typographical and formatting issues, such as "while We" in Section 1, "uesing" in Section 3.1, and the unusual "5!×!5" notation in the filter description. These should be corrected.

[1] The Forensic Cost of Watermark Removal (2026) [2] Removing the Watermark Is Not Enough (2026)

Questions for authors
  1. What is the main novelty of RemTraceNet compared with existing watermark-removal forensic methods, particularly [1] and [2]? Since most of the architectural components are already used in image forensics, I would like to understand what is fundamentally new in the proposed approach.

  2. How can the authors confirm that the model is learning watermark-removal traces rather than general image-processing artifacts? For example, if the same transformations are applied to clean images without watermarks, would the model produce similar detection scores? Also, how does the performance change when only successful watermark-removal attacks are considered?

  3. How would the proposed framework handle an unknown or unseen attack pipeline? Since a separate model is trained for each known attack, it would be useful to see cross-attack evaluation results. I am also curious how the model performs when multiple processing operations are applied to the same image.

  4. How much of the reported improvement comes from the proposed architecture rather than the additional loss functions? Could the authors compare the models using the same training objective and report results across multiple random seeds? This would help better understand the actual contribution of the architecture.

  5. Why does the performance drop significantly for generator-integrated watermarks under diffusion-regeneration attacks at 512×512? Could this be related to the differences between clean and watermarked image realizations, or are there other reasons? Some further analysis would help explain this behavior.

Comments on the paper

1 comment
Area_Chair_hz3rOpenReview ↗

Discussion Phase

As the discussion period is rapidly coming to a close (Nov 18 AoE), I would like to remind the reviewers to please participate in discussion and acknowledge any comments posted by the authors. A secondary, important reminder: full disclosure of any AI usage is required and the reviewer must include their own human-written assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.