acceptodds
Withdrawn from ICLR 2027

From Global Verification to Fine-Grained Tamper Localization: Exploiting Spatial Reliability in Robust Watermark Decoders

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Robust image watermarking has become an effective tool for content authentication, yet existing systems are primarily designed for global message recovery and provide little information about where image integrity has been compromised. In this work, we show that globally trained watermark decoders already contain a previously overlooked, message-conditioned spatial integrity representation. By probing intermediate features of frozen pretrained watermark decoders, we find that local consistency with the authentic embedded message progressively emerges together with message recovery: intact regions retain strong support for the embedded watermark, whereas manipulated regions exhibit degraded local consistency. Counterfactual experiments with shuffled and random messages further show that this spatial separation is specifically tied to the authentic watermark rather than generic editing artifacts or message-independent feature instability. Building on this observation, we introduce a lightweight decoder-side tamper localization framework that extracts spatially distributed watermark evidence from intermediate decoder features, measures its consistency with the embedded message to construct an unreliability map, and combines this active watermark cue with RGB appearance for pixel-level localization. Importantly, the original watermark encoder and global decoder remain entirely frozen, requiring no modification to the deployed watermarking pipeline. Our findings reveal that robust watermark decoders encode substantially richer spatial information than their global objectives explicitly require, providing a new route for extending existing watermark systems from global verification to fine-grained tamper localization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
digital watermark, generative model security

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_nhBgOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a localized tampering detection method by appending and training a localization branch onto a frozen image watermarking decoder. The core mechanism leverages the disagreement between spatially predicted watermark bits and the ground truth to identify tampered regions. In essence, the approach operates similarly to "anomaly localization within the decoder", where tampered areas are treated as anomalies that the pre-trained watermarking decoder cannot correctly interpret.

Critical strengths

The paper successfully demonstrates that a supervised branch module can utilize spatial message disagreement from a frozen watermarking decoder to localize tampering manipulations.

Critical weaknesses
  1. There is a major scientific concern regarding whether the provided evidence fully supports the claimed representation phenomenon. The authors state that "globally trained watermark decoders already contain a previously overlooked, message-conditioned spatial integrity representation." However, the current empirical evidence does not yet clearly support this specific claim.

  2. Significant omissions in the methodology section affect the paper's reproducibility, readability, and overall interpretation.

Questions for authors

Questions Related to the Weaknesses:

  1. Can the authors clarify how much of the localization performance stems from the claimed "spatial integrity representation" versus how much is simply local decoding failure being successfully leveraged through supervised learning? Does the proposed method reliably identify content manipulation and representation, or does it primarily isolate areas that are structurally inconsistent with the underlying expected watermark?

  2. Significant mathematical and architectural details are missing. For example. (1) Equation 3: Five loss terms are listed, but their mathematical definitions are missing. (2) Architectures: The exact architectures for the probe, local message head, and refinement network are currently unspecified. (3) Features and Resolution: Which specific feature stage is extracted from the decoder, and what is its resolution in the final system? (4) Hyperparameters and Training: Details regarding the probe's capacity, receptive field, and key training configurations and steps have been omitted.

Other questions:

  1. Does the localization branch mistake benign watermark degradation (e.g., standard compression or minor noise) for intentional tampering?
  2. Have the authors considered using random or shuffled controls to ensure the method is explicitly identifying tampering rather than generic editing artifacts?
  3. The text mentions the probe as "lightweight." To support this claim, please provide quantitative evidence such as parameter counts, computational complexity, or runtime measurements.
Reviewer_i8QFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether watermark decoders trained for global message recovery contain spatial information useful for tamper localization. The authors analyze intermediate decoder features using clean only message probes and introduce a localization branch that combines message-conditioned watermark disagreement with RGB information. Experiments on RegionBase, Stable Diffusion inpainting, and splicing report competitive localization results while keeping the original watermark encoder and global decoder frozen.

Critical strengths

S1: The main strength is the investigation of spatial information within watermark decoders originally trained for global verification. This is a useful question, particularly because the approach does not require modifying the pretrained encoder or global decoder.

S2: The clean-only layer-wise probing experiments provide interesting evidence of spatially varying message support in frozen decoder features. The improvement over RGB-only localization also suggests that the extracted watermark-reliability cue provides complementary information for tamper localization.

Critical weaknesses

W1: The main concern is whether the experimental construction sufficiently supports the claimed tamper-localization capability. Equation (4) preserves the original watermarked pixels outside the manipulation mask while inserting edited pixels inside it. If those inserted pixels do not retain the same watermark, the task may become substantially easier than localization after direct editing of a watermarked image.

W2: A second concern is the interpretation of the spatial reliability analysis. The random-message counterfactual has an expected disagreement of 0.5 by construction, while the final reliability head also receives manipulation-related supervision. Stronger controls are needed to establish what spatial information is already present in the pretrained decoder and how much is learned by the added branch.

Questions for authors
  1. In Eq. (4), the manipulated image is constructed by combining the edited image with the watermarked image using the ground-truth mask. If the inserted region does not contain the original watermark, the method may primarily identify where watermark evidence is missing. Were the edited images generated from the watermarked images or the original images? Could the authors report results when editing is applied directly to watermarked images without mask-based compositing?
  2. The interpretation of Fig. 2(c) needs further justification. For an independently sampled uniform random binary message, the expected disagreement in Eq. (1) is exactly 0.5, regardless of the local predictions. Thus, the near-zero spatial gap is expected and does not independently establish message-conditioned integrity information. Although the authors acknowledge this limitation, a stronger control is needed to distinguish watermark-specific evidence from generic editing-induced feature disruption.
  3. In Section 3.2, each spatial probe predicts the embedded message at every feature location, but the message-sampling protocol is unclear. Were messages independently randomized across training images and evaluation samples? If a fixed message was reused, the probes could learn a message prior rather than recover watermark information from the features. The authors should also clarify the receptive fields of S1–S8, since predictions from large receptive fields may not represent genuinely local watermark evidence.
  4. The claim that useful spatial reliability already exists in the pretrained decoder is difficult to separate from the training of the final localization branch. In Eq. (3), the local message head is optimized using manipulation-related gap and alignment losses, even though the original decoder remains frozen. What happens when the clean-only message probe is frozen and used to construct the reliability map for localization? This experiment would help clarify whether the localization performance comes from the pretrained decoder or is largely learned by the additional branch.
  5. The comparison in Table 1 raises questions about baseline fairness. The RGB-only variant in Table 3 achieves F1 = 0.86, substantially higher than the strongest reported passive baseline in Table 1 (F1 = 0.401), although the evaluation setting for Table 3 is not explicitly stated. Were the passive methods retrained or fine-tuned using the same training data, and how were their thresholds selected? Also, Watermark Anything is included in Table 2 but omitted from the localization comparison despite providing spatial watermark detection. A comparison under a clearly specified protocol would be useful.
  6. Table 2 evaluates global watermark bit recovery under benign distortions, but the corresponding localization performance is missing. JPEG compression, resizing, and blur may weaken local watermark evidence without actual tampering. Could the authors report localization performance when manipulated images undergo these transformations, together with false-positive rates on untampered watermarked images?
  7. Several important methodological details are missing from Section 3.4. Equation (3) introduces the message, gap, alignment, and negative losses, but their mathematical definitions are not provided. The architecture of the local message head, the selected decoder stage, and the localization network are also insufficiently specified. These details are important for reproducing the results and understanding which components contribute to the reported improvements.
Reviewer_JegzOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  • SUMMARY: This paper investigates the spatial information hidden in pretrained watermark decoders originally trained for global message recovery. The authors show that intermediate decoder features contain message-conditioned spatial reliability that can distinguish intact from tampered regions. Based on this observation, they propose a lightweight localization framework that combines local watermark disagreement with RGB features to predict tampered regions, while keeping the original encoder and decoder frozen. Experiments demonstrate competitive tamper localization performance across multiple editing scenarios.

  • ASSESSMENT: I find the observation interesting, but have several concerns. In particularly, the proposed approach of freezing a "global" watermarking scheme likely would be inferior to a dedicated scheme on tamper detection. Nonetheless, the proposed approach would still be useful to get a baseline or quick preliminary solution for further enhancement.

Critical strengths
  1. Novel perspective. The paper presents an interesting perspective that pretrained global watermark decoders already contain spatial information useful for tamper localization.

  2. Am promising approach to get ``baseline'' solution. This observation suggests a quick solution that extend existing robust watermarking systems for tamper detection, which can serve as baseline for comparison.

Critical weaknesses
  1. Why Adopt a Global Watermarking Method? It is unclear why extending an existing robust watermarking system for tamper detection would be more effective than using a method designed specifically for this task. Consider the well-known spread-spectrum watermark, which is commonly used for tracing. Its detector aggregates watermark evidence across the image and recovers the global message when the aggregate exceeds a threshold. The local watermark evidence can also be examined: weak evidence in a particular region may indicate tampering there. By analogy, one could freeze such a scheme and learn to detect local tampering, following the approach in this paper. However, block-based methods can be viewed as global watermarking schemes with a blockwise repetition structure designed to provide local evidence for tamper detection. Fridrich [1] is one example. This raises the concern that learning a localizer on top of a frozen global watermark may not outperform methods that build in local evidence.

[1] Fridrich “Image Watermarking for Tamper Detection”, ICIP 1998.

  1. Dependence on reliable watermark messages. The method assumes access to the correct embedded message at inference time, which may not always be available in practice. If the globally decoded message is used instead, decoding errors could distort the unreliability map and degrade localization accuracy. Moreover, severe benign distortions may weaken local watermark evidence even when the correct message is available, potentially causing false positives.

  2. Unverified scalability with message length. The method relies on predicting the full watermark message at each spatial location. Although the paper reports an additional experiment with a 100-bit backbone, it does not systematically evaluate how local prediction reliability changes as message length increases. Global message recovery does not guarantee reliable local recovery, particularly for longer messages.

Questions for authors
  1. Relation to prior work: How does the proposed method relate to prior watermarking methods for tamper localization, including fragile and semi-fragile watermarking and blockwise spread-spectrum methods such as [1]? Please clarify the differences in their objectives and localization mechanisms.

  2. Advantage over purpose-designed methods: What evidence shows that learning a localizer on top of a frozen global watermark offers an advantage over methods designed specifically for tamper detection, particularly block-based methods?

  3. Generality across architectures: Do the observed spatial-reliability properties also hold for other watermarking architectures, such as the CNN-based MBRS, the INN-based FIN, and Transformer-based methods? Evaluating or discussing this would clarify how broadly the findings apply.

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

Related papers

Discussion (0)

Sign in to comment.