acceptodds
Withdrawn from ICLR 2027

Localized Multi-Object Watermarking

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

As images are increasingly reused and recomposed, tracing the provenance of individual objects requires watermarks that remain recoverable after partial theft. Beyond conventional global image watermarking, localized watermarking has recently emerged to protect individual objects within an image. However, localized watermarks remain vulnerable to object-level partial theft, where cropping, transformation, and recomposition of a watermarked object may preserve only a small portion of the embedded signal. Furthermore, the problem becomes more challenging when a composite image contains multiple objects, requiring independent message extraction and localization. To address these challenges, we propose MoMark, a localized multi-object watermarking framework that embeds independent messages into localized regions and jointly recovers their messages and spatial support. We employ a two-stage training strategy: single-object message extraction and localization, followed by multi-object training with partial-support objectives. At inference, MoMark pools messages over local patches and separates or merges watermarked regions by message similarity, without requiring ground-truth masks. Evaluated on complex segmentation masks with multiple independently watermarked objects, MoMark improves object-level message extraction and localization across geometric distortions, non-geometric distortions, and a variety of partial-theft settings.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Localized Image Watermarking, Multi-object Watermarking, Partial Image theft, Digital Provenance

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

The paper formulates localized multi-object watermarking: each object instance carries an independent message, and the extractor must recover each message and its spatial support from a composite image without ground-truth masks, known messages, or a known object count. MoMark uses a JND-modulated UNet embedder and an SWT+UNet extractor predicting per-pixel presence and bit logits, with mask-weighted pooling for region decoding. Training is two-stage (single-object, then multi-object plus partial-support fragment recovery with a PSNR-budget penalty); inference groups patch-level pooled messages by similarity and refines boundaries by bit-agreement reassignment. On COCO at matched ~39 dB PSNR, MoMark beats WOFA, WAM, and MaskWM on attack-average bit accuracy and instance IoU at k=3/5, generalizes from <=3 training objects to 5, survives a retention sweep down to 10% of bounding-box area, and is evaluated under instance-specific theft, inpainting/VAE/regeneration/outpainting, 64/128-bit messages, and latency.

Critical strengths
  1. The problem formulation matches the real threat. Treating the object, not the image, as the unit of reuse — and requiring joint message recovery and instance localization with unknown object count — exposes a gap that WOFA (single global message) and WAM/MaskWM (separated regions) do not cover. The returned-instance-count analysis makes this concrete: at k=5 with 10% retention, WAM returns 0.718 instances, MoMark 4.843.
  2. The evaluation is thorough and fair: 14 attack conditions at matched fidelity, retention sweeps, instance-specific theft, generative attacks with matched VAE fine-tuning, longer messages, efficiency, and full ablations. The multi-object gains are consistent (theft-average bit accuracy 0.962/0.911 at k=3/5 vs <=0.728/0.659 for the baselines).
Critical weaknesses
  1. Methodological novelty is modest relative to the formulation. The SWT front-end, per-pixel bit logits with mask-weighted pooling (as in WAM), two-stage curriculum, and similarity grouping are individually familiar; the contribution reads as a strong integration plus the fragment-supervision recipe. The transferable takeaway beyond watermarking is limited.
  2. Robustness collapses against the strongest attacks, and there is no adaptive-attacker analysis. Under object-level inpainting, bit accuracy is ~0.66 even after VAE fine-tuning; under VAE reconstruction/regeneration ~0.6; clean 128-bit accuracy at k=1 is only 0.788. An adversary who knows the scheme — perturbing SWT features to flip pooled bits, or embedding forged messages to fabricate provenance — is never considered, and provenance systems derive their value from exactly such settings.
Questions for authors
  1. What drives the clean-input degradation at 128 bits (0.788 bit accuracy at k=1 vs 1.000 for MaskWM) — capacity, training schedule, or pooling? Does longer training close it?
  2. How sensitive is decoding to the message-grouping threshold (normalized Hamming 0.25) and the 32-pixel minimum candidate size, especially when two objects carry similar messages?
  3. Can embedding and verification be keyed, so an adversary cannot forge provenance on stolen content? Is that in scope?
  4. Have you tried adaptive removal against the SWT+UNet extractor (white-box bit-flip perturbations at matched PSNR)?
Reviewer_U4qJOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes MoMark, a multi-object local image watermarking framework. MoMark adopts a two-stage training strategy, progressing from single-object watermark extraction and localization to multi-object learning with local support-region sampling. Experiments on COCO compare MoMark with WAM, MaskWM, and WOFA under geometric transformations, non-geometric transformations, and local watermark theft. The authors report strong message recovery and instance-level localization performance, including generalization from three training objects to five test objects.

Critical strengths
  1. The training strategy is well-motivated. Training on single objects first, followed by multi-object and partial-support-region training, helps the model progressively learn message recovery, localization, and local robustness.

  2. The paper evaluates both geometric and non-geometric transformations and achieves competitive multi-object localization performance compared with existing methods.

Critical weaknesses
  1. The algorithmic novelty is not sufficiently convincing. The paper fails to clearly explain the core methodological innovations of MoMark compared with WAM and MaskWM.

  2. The fairness of baseline comparisons needs to be improved. WOFA uses 30-bit watermarks, whereas WAM, MaskWM and MoMark use 32-bit watermarks. The message lengths are not standardized across methods, which may affect the fairness of the comparisons.

  3. The experimental section is somewhat disorganized and overly focused on reporting numbers already presented in the tables, with insufficient analysis. Additionally, it refers to tables outside the main text.

  4. The perceptual quality evaluation is insufficient. Although PSNR and chroma loss are useful metrics, they cannot fully capture visual artifacts in complex textures, fine object boundaries, and high-frequency details. Additional metrics, such as SSIM and FID, should be included. The experimental section lacks an evaluation of image quality metrics.

Questions for authors
  1. Why do the baseline methods use different watermark message lengths (30 bits for WOFA versus 32 bits for WAM, MaskWM, and MoMark)? Could the authors provide additional experiments with same message lengths to ensure a fair comparison?

  2. Could the authors include additional image quality evaluation experiments using metrics such as SSIM and FID?

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

This paper investigates whether it is possible to embed different watermark messages into different objects within an image and, when those objects are cropped, resized, or pasted into other images, detect them and extract their respective messages. The authors’ proposed MoMark method first predicts the locations of watermarks and the message information at each location, then separates the different objects through weighted pooling, message clustering, and boundary adjustment. Training consists of two stages: first, learning single-object localization and message extraction; then, learning multi-object extraction and message recovery when only a small portion of an object remains. Experiments were primarily conducted on the COCO dataset and included standard image processing tasks, copy-and-paste operations, cropping, scaling, occlusion, and some generative edits. The results show that MoMark outperforms the compared methods in most multi-object tests and can handle more objects than were used during training.

Critical strengths
  • The research question has practical significance.
  • The method design aligns clearly with the problem to be solved. Using message similarity to distinguish connected objects and merge disconnected regions of the same origin is a reasonable approach. The inclusion of dedicated training for fragment recovery directly addresses the challenge of extracting watermarks from objects that have been heavily cropped. Together, these design elements form a relatively comprehensive processing workflow.
  • The experimental results demonstrate excellent performance. The method maintains its advantage even when the number of objects in the test set is increased to five from a maximum of three in the training set, indicating that the method possesses a certain degree of generalization capability regarding the number of objects.
Critical weaknesses
  • The omission of the directly relevant MuST results in an incomplete assessment of the contribution. MuST: Robust Image Watermarking for Multi-Source Tracing (Wang et al., AAAI 2024) has already investigated the localization and message extraction of multiple watermark materials after cropping, resizing, and recombination. MoMark’s main plagiarism experiments are highly relevant to this scenario, yet the paper neither discusses nor compares this method. While MoMark does feature different designs for message grouping and fragment training, it is necessary to explain what problems these designs solve relative to prior work and what practical benefits they provide.

  • The ablation experiments do not clearly elucidate the contributions of individual components. Table 12 does not standardize embedding distortion; for example, the PSNR of the full model is 37.2 dB, while removing multi-object message supervision results in 44.2 dB. The watermark strengths differ significantly between the two, so the change in accuracy cannot be entirely attributed to this training component. Furthermore, the paper does not sufficiently compare the effects of uniform pooling versus weighted pooling, or of using only connected components versus incorporating message grouping. The current results indicate that the overall approach is effective, but they are insufficient to determine the primary source of the improvement.

  • The explanation of instance-level evaluation is incomplete, which affects the assessment of the actual attribution capability. The paper needs to more clearly explain how predicted objects are matched with ground-truth objects, as well as how false negatives, false positives, duplicate predictions, and misattributions are scored.

Questions for authors
  • What is the most significant new contribution of this paper compared to MuST? Please provide additional discussion and, whenever possible, make comparisons under the same message length, watermark strength, and object scale.

  • How exactly were the multi-object samples and evaluation results obtained? Do the objects in the second stage come from the same COCO image, or were they extracted from different images and then recombined? How were the backgrounds and placement positions selected during testing? How are predicted objects matched with ground-truth objects, and how are false negatives and false positives scored? Additionally, how does the message configuration used for TPR relate to the independent multi-message extraction experiment?

  • Could you provide key ablation results at the same PSNR? Additionally, we would like to compare uniform pooling with weighted pooling on the same extractor, as well as the effects of using connected components alone versus incorporating message grouping. This would provide a clearer understanding of the actual benefits contributed by each component.

  • Does the inpainting process restore the original image pixels outside the editing region? Please explain the mask handling, prompt, and key sampling parameters. Table 3 evaluates objects that were not selected for editing; therefore, it is necessary to distinguish whether the watermark degradation stems from local editing or from the entire image being reconstructed via VAE. If both effects are present, we ask the authors to explicitly clarify this.

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

Related papers

Discussion (0)

Sign in to comment.