acceptodds
Withdrawn from ICLR 2027

SRW: FROM DISCRETE WATERMARK IDENTIFICATION TO SEMANTIC PROVENANCE RETRIEVAL

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Watermarking embeds recoverable source information in AI-generated images to support attribution and auditing after dissemination. Compression, cropping, resizing, and editing can damage this signal, while incomplete provenance records create a second obstacle: recovering a discrete identifier does not by itself retrieve a semantically equivalent source description. We introduce SRW, which reframes watermark identification as semantic provenance retrieval. A provenance payload is source metadata intentionally embedded in the image and may differ from the prompt controlling visible content. SRW selects informative payload tokens, injects their continuous representation into diffusion latents, and learns to recover it from distorted images for cosine-ranked search with calibrated rejection. In the recorded attack sweep, retrieval accuracy is 94% after JPEG quality 50, 95% after 0.5× resizing, and 94% after a 20% crop. In a separate evaluation excluding the exact source text, Recall@1 reaches 93% for paraphrases and 82% for hard semantic neighbors. Payload-shuffling summaries further support recovery of the embedded payload rather than visible content alone. Bank-size diagnostics give 98% Top-1 at 1K candidates and 92% at 10K. These results support a semantic retrieval interface beyond discrete identity lookup, with strong impulse noise, image fidelity, and joint distortion–source-absence evaluation remaining limitations.

Primary area:
generative models
Keywords:
AI-generated images, watermarking, semantic provenance retrieval, diffusion models, robust watermark recovery, image attribution, provenance tracking.

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_UGXDOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes SRW, a watermarking method for latent diffusion images that embeds a continuous representation of provenance text rather than a discrete identifier. The provenance payload is separate from the generation prompt. A frozen text encoder and learned attention readout produce a pooled embedding, which a learned injector maps to a latent perturbation. An image-side decoder recovers the embedding, and cosine similarity is used to retrieve related payloads from a candidate bank, with a rejection threshold calibrated on held-out negatives.

Experiments on Stable Diffusion 1.5 report robustness to several image transformations, retrieval of paraphrases and semantically related records when the exact source text is absent, and performance across different bank sizes. However, the paper states that the reported values are aggregate results inherited from an earlier workbook, while important protocol details, raw outputs, and quantitative ablations are unavailable or unverified.

Critical strengths

None

Critical weaknesses

W1. Missing evaluation protocols and underlying records prevent the central results from being assessed reliably. The paper explicitly states that its numerical results are inherited from an earlier experiment workbook and have not been independently reverified. Sample counts, seeds, raw predictions, and uncertainty estimates are unavailable. More fundamentally, the relevance-label and negative-pool construction underlying the retrieval and detection metrics remains unresolved. The definition of the “Random” category, several transformation settings, and the baseline adapters also require recovery or verification from the original records. Without these details, it is unclear what constitutes a successful retrieval or how the reported methods were compared. The bank size used in the attack sweep is unspecified, although Table 6 shows that retrieval accuracy depends on bank size. No quantitative ablations support the attention selection, injection mechanism, loss terms, or hyperparameter choices. Moreover, the motivating setting combines image distortion with the absence of the exact source record, but these conditions are evaluated separately. The payload-binding control and several additional robustness results are reported only qualitatively. These are substantive evidence gaps, rather than presentation issues alone. Unless the original experimental records can be recovered, assessing the headline claims would require rerunning the experiments under clearly specified protocols.

W2. The comparisons do not establish the claimed advantage of semantic payloads. HiDDeN and RingID do not directly encode semantic relationships, so their low performance in the source-absent semantic retrieval setting does not, by itself, establish an advantage over a complete identity-based retrieval system. A more informative alternative would decode an identity, retrieve its retained metadata, and use that metadata for semantic search. The paper acknowledges this alternative but does not evaluate it. Such a comparison must also specify whether source absence refers to the retrieval bank, the metadata store, or both. The practical tradeoff is similarly unclear. The paper reports a single fidelity operating point, with PSNR 25.64 dB and SSIM 0.8152, but no retrieval–fidelity curve as injection strength varies. The existing results therefore do not establish whether the semantic benefit justifies the image distortion. Finally, the related-work discussion does not address closely related semantic-watermarking approaches such as SEAL and OF-SemWat. Their objectives need not be identical to SRW’s, but the paper should explain the distinction between content-derived watermarking and independently specified provenance payloads. The principal concern is not the absence of uniformly superior results; it is that the comparisons do not isolate or substantiate the claimed contribution.

Questions for authors

Q1. Evaluation protocols and recoverability of results. For Tables 1, 2, 3, 5, and 6, please specify the query counts, bank sizes and composition, relevance-set construction R(s)R(s), negative pools for AUROC and FPR95, transformation settings, and seeds. Which details and per-query results can be recovered from the original runs, and which would require new experiments?

Q2. Identity watermark with retained metadata. How does SRW compare with an identity watermark linked to a metadata database, followed by semantic search over candidate records? Please use matched transformations and candidate banks, and distinguish removal of the source from the retrieval bank from removal of its metadata mapping.

Q3. Combined distortion and source absence. What are Recall@1 and rejection rates when the exact source record is absent and the image is also distorted, for example by JPEG 50 or the tested crop? Please define the relevance criterion and threshold-calibration protocol for this setting.

Q4. Component contributions and fidelity tradeoff. Is quantitative evidence available for attention selection versus mean pooling, the individual loss terms, and latent versus pixel injection? How do retrieval performance and PSNR/SSIM vary with injection strength λ\lambda?

Q5. Payload binding and relation to prior work. In the pA/sBp_A/s_B shuffling test, how often does the decoder favor the embedded payload sBs_B over the generation prompt pAp_A, and across how many examples? How does SRW’s independently specified provenance payload differ from the semantic information used by SEAL and OF-SemWat?

Reviewer_rJr1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SRW embeds a continuous representation of textual provenance metadata into diffusion latents and learns to recover it from distorted images for semantic retrieval. The method combines attention-based payload selection, learned latent injection, an image-side decoder, and cosine similarity search with threshold-based rejection. The formulation distinguishes the generation prompt from the embedded payload and separates exact attribution from semantic tracing. The experiments report retrieval under image distortions and, separately, retrieval when the exact source text is absent from the candidate bank, together with bank-size and threshold diagnostics.

Critical strengths
  1. A potentially useful problem formulation. Recovering provenance semantics could help trace images when exact records are unavailable but relevant descriptions remain. The paper appropriately distinguishes such evidence from proof of source identity.

  2. Preliminary evidence of feasibility. The reported results cover several image distortions and a separate evaluation with the exact source text excluded. These provide a useful starting point for assessing the proposed system.

Critical weaknesses
  1. The incremental contribution is not established by the related-work discussion and comparisons. SEAL already uses semantic information in diffusion watermarking, although for content verification, while OF-SemWat embeds and recovers text. Adding semantic retrieval to OF-SemWat is a natural adaptation to SRW's setting, without requiring a historical ID-to-record mapping, but this alternative is not evaluated. Moreover, the HiDDeN/RingID adapters in Table 3 are unspecified (Section 5.1; Appendix B), so their low recall cannot isolate watermark limitations from retrieval-interface limitations. The authors appropriately qualify these as lookup controls and acknowledge the missing ID-plus-semantic-search baseline, but these qualifications do not establish the advantage of directly recovering continuous semantic representations.

  2. Robustness is not established for the paper's central retrieval task. SRW's main task-level contribution is retrieval when the exact source text is absent. JPEG compression, resizing, and cropping are routine dissemination conditions for watermarked images, so evaluating this contribution requires testing distorted images against a bank without the exact source text. Sections 5.2 and 5.3 assess these conditions separately. Consequently, the reported results do not establish whether the proposed semantic retrieval capability remains effective under common dissemination conditions.

Questions for authors
  1. Clarification of Table 3: How are the outputs of HiDDeN and RingID converted into candidate rankings? After excluding the exact source text from the candidate bank, does the ID-to-record mapping remain accessible, and what source information is available to each method?

  2. Clarification of payload construction: Section 5.1 identifies caption/prompt datasets as payload sources and explicitly states that the source-absence evaluation uses s = p. How are p and s selected and paired during training and the other evaluations? Beyond the shuffling control in Section 5.4, which evaluations use provenance metadata distinct from the generation prompt?

Reviewer_ehbpOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes SRW, a watermark for diffusion images that carries a semantic vector instead of a bit string or key. A payload text is encoded by a frozen text encoder, and a learned head selects and pools informative tokens into one normalized vector. An injector maps this vector to a perturbation of the initial Stable Diffusion 1.5 latent. A decoder trained under simulated attacks recovers the vector from the image, and provenance is found by cosine search over a bank of text records with an optional rejection threshold

Critical strengths
  1. The framing is interesting. Moving from exact ID lookup to retrieval in a semantic space, where a slightly wrong decode still lands near a related record, is a reasonable new way to think about provenance. Separating the generation prompt from the embedded payload is also a good design choice, because it makes a binding test possible in principle.
  2. The paper is open about its limits. It states clearly that robustness holds only for moderate edits, that impulse noise breaks the method, and that several key evaluations are missing.
Critical weaknesses
  1. The results cannot be verified, and the paper says so itself. Section 5.1 states that query-level predictions, counts and seed-level uncertainty are unavailable. Appendix A says this revision does not re-verify the underlying workbook or raw experiment logs. Relevance sets and negative pools "need recovery from the logs", baseline adapters are unspecified, and per-example shuffling scores "were not retained". Appendix D summarizes a supplementary PDF that is not included. Every number is a rounded point estimate with no sample size, so I cannot tell whether 94% comes from 50 images or 5,000.
  2. The central claim is not tested against a fair baseline. The claim is that semantics should live inside the watermark. The natural alternative is an existing robust bit watermark (HiDDeN, Stable Signature, Gaussian Shading) carrying a quantized or hashed text embedding, which gives the same semantic neighborhood property with a mature carrier. An ID watermark plus a semantic database is also missing, which the paper admits. The HiDDeN and RingID numbers in Table 3 (about 1 to 3%) come from undefined adapters and are effectively a strawman.
  3. There is no quantitative evidence that SRW reads the hidden payload. Plain CLIP already reaches 86% clean retrieval from visible content (Table 1). The shuffle test (content from pA, payload sB) is the only evidence that SRW reads the hidden channel. It is reported only as a direction, with no success rate and no sample size.
  4. Generalization to unseen payloads is unclear. The paper does not say whether the bank payloads are disjoint from the training payloads, so memorization cannot be ruled out.
  5. Key evaluations are missing, and the authors acknowledge most of them:
  • The joint setting of distortion plus a missing source is untested.
  • There are no ablations of token selection, latent injection or the loss terms.
  • Regeneration and adaptive attacks, the main threat to latent watermarks, are only described qualitatively.
  • Only one threshold point is reported for calibrated rejection.
  • Only one backbone is tested.
  1. The image quality cost is significant and not compared. PSNR is 25.64 dB and SSIM is 0.815. There is no comparison with Tree-Ring or Gaussian Shading using FID or CLIP score, which are the appropriate metrics for a method that perturbs the initial latent.
Questions for authors
  1. How many images and queries are behind each number in Tables 1 to 6, and can you report variance over seeds?
  2. Are the payloads in the retrieval bank disjoint from those used to train the decoder?
  3. How does SRW compare with a bit watermark carrying a quantized text embedding, and with an ID watermark plus a semantic database?
  4. What is the quantitative success rate of the shuffle test (content from pA, payload sB) over many examples, and under distortion?
  5. How were the HiDDeN and RingID adapters built, and how are relevance sets and negative pools defined? What does "Random" mean when the exact source is excluded?
  6. How does SRW perform under diffusion regeneration and img2img attacks, and what are its FID and CLIP scores compared with other diffusion watermarks?

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

Related papers

Discussion (0)

Sign in to comment.