- 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
- 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.
- 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
- 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.
- 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
- 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?
- 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?
- Can embedding and verification be keyed, so an adversary cannot forge provenance on stolen content? Is that in scope?
- Have you tried adaptive removal against the SWT+UNet extractor (white-box bit-flip perturbations at matched PSNR)?