acceptodds
Under review as a conference paper at ICLR 2027

Alignment Is Not Enough: Preserving Message Separability for Geometrically Robust Image Watermarking

Abstract

Invisible watermarking embeds a traceable identifier in an image and must recover it after routine geometric processing in real-world use. Resampling attenuates message-bearing components, while strengthening the signal harms visual quality, so the identifier must remain separable from competing candidates under both pressures. Existing systems recover under some of these conditions but collapse under others, and we therefore design the system around separability. Our readout matches each template over its orbit of integer feature shifts, absorbing alignment variation without estimating it. Because no alignment search can recover components that interpolation has erased, the codebook excludes the spectral component that vanishes under half-bin interpolation. A two-stage embedding then places the message in the host image, first balancing residual energy against message margins, then refining perceptual quality at fixed energy while protecting geometric margins. Across nine baselines and three datasets in closed-set 64-ID identification, ours is the only system whose mean recovery exceeds 99.5% in every evaluated condition, while a synchronization-corrected baseline falls as low as 60% when geometry is composed with JPEG. Ours reaches 99.92% off-grid ID accuracy at LPIPS 0.00946, retains 99.90% after a PNG codec round trip, and accepts no wrong ID across 18 watermarked conditions at a 0.04 rejection threshold.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Geometric robustness, Message separability, Image watermarking, Perceptual embedding

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_ttvFOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the problem of recovering invisible image watermarks after geometric transformations, where rotation and scaling can cause alignment changes and interpolation-induced loss of message-bearing information. The authors argue that correcting alignment alone is insufficient because the correct watermark must remain distinguishable from competing message identities after geometric resampling.

The proposed method consists of three main components: (1) a Fourier–Mellin-inspired feature representation with orbit-based matching that provides invariance to integer circular shifts in feature space without explicitly estimating geometric alignment; (2) an interpolation-compatible codebook that excludes vulnerable Nyquist-frequency components to reduce message degradation under fractional feature shifts; and (3) a two-stage, per-image embedding optimization that first balances residual energy and message separation, then improves perceptual quality at fixed residual energy while protecting message margins under a predefined geometric transformation.

The authors evaluate the framework on a closed-set identification task involving 64 watermark IDs, using COCO, AMBER, and DIV2K datasets and nine baseline configurations. The proposed method achieves over 99.5% mean ID recovery across all evaluated transformation conditions, including rotation, scaling, translation, and JPEG compression. Additional ablation studies examine the contributions of the spectral codebook constraint and perceptual refinement with geometric-margin protection.

Critical strengths
  1. Well-motivated problem and theoretically grounded approach: The paper identifies an important limitation of alignment-based watermark recovery: geometric resampling can remove message-bearing information even when alignment is corrected. The proposed interpolation-compatible codebook is motivated by mathematical analysis of fractional shifts, providing a clear connection between the identified problem and the proposed solution.

  2. Strong experimental performance across multiple geometric transformations: The proposed method achieves over 99.5% mean ID recovery across all evaluated conditions on COCO, AMBER, and DIV2K. Particularly notable is its performance under combined rotation, scaling, translation, and JPEG compression, where it achieves 99.75% accuracy on COCO validation compared with 64.10% for WAM+SyncSeal. These results demonstrate strong robustness within the evaluated transformation settings.

  3. Meaningful ablation studies supporting the proposed components: The paper investigates the effectiveness of excluding fragile Nyquist-frequency components and the two-stage embedding strategy. The codebook ablation demonstrates improved off-grid recovery, while the refinement experiments show that perceptual distortion can be substantially reduced without sacrificing recovery when geometric margins are protected. These experiments provide useful evidence supporting the proposed design choices.

  4. Joint consideration of watermark robustness and perceptual quality: Rather than focusing exclusively on recovery accuracy, the method explicitly balances message separability, residual energy, and perceptual distortion. The two-stage optimization provides a structured approach to improving visual quality while maintaining watermark recovery under the selected geometric transformations.

Critical weaknesses
  1. Limited geometric robustness evaluation: Although the paper focuses on geometrically robust watermarking, the experiments primarily consider rotation, scaling, and limited translation under controlled synthetic transformations. Other important geometric distortions, such as perspective transformations, shearing, and substantial cropping, are not evaluated. These transformations are common in real-world image capture and editing scenarios. Therefore, the reported results do not fully establish robustness beyond the evaluated geometric settings.

  2. Limited message capacity and scalability: The main experiments consider only 64 watermark IDs, equivalent to 6 bits of information. Although Appendix E.1 extends the evaluation to 128 and 256 IDs, acceptance coverage decreases as the message space increases. For example, coverage under continuous rotation-scale transformations drops from 91.60% for the jointly trained 64-ID reference to 73.24% for the reoptimized 256-ID codebook. This raises concerns about scalability to larger message spaces and practical applications requiring substantially longer watermark identifiers.

  3. Gap between theoretical analysis and practical geometric transformations: The theoretical results establish desirable properties under a specified feature-space interpolation operator. However, actual geometric transformations involve pixel-space interpolation, which does not necessarily satisfy the same assumptions. Although the authors provide supporting empirical results, a quantitative analysis connecting the theoretical guarantees to the observed robustness under pixel-space resampling would strengthen the technical contribution.

  4. Insufficient evaluation under broader image-processing and adversarial attacks: The paper evaluates JPEG compression and its combination with geometric transformations, but does not provide comprehensive experiments involving noise, blur, brightness changes, repeated compression, or adaptive adversarial watermark removal. While these attacks extend beyond the primary geometric focus, evaluating them would help establish the practical robustness and limitations of the proposed framework.

  5. Limited justification and independent validation of the rejection threshold: The authors use a fixed threshold of 0.04 and report zero incorrectly accepted watermark IDs across the evaluated conditions. Although threshold sensitivity is discussed, the threshold was selected during development, and the authors acknowledge that all evaluation datasets were inspected during the project. Since a large score difference does not necessarily guarantee a correct prediction, validation on a strictly held-out dataset would provide stronger evidence of threshold generalization.

  6. High computational cost of watermark embedding: Although decoding is efficient, the proposed method requires per-image optimization involving 1,000 Adam steps in the first stage and 150 additional steps in the second stage. The paper does not provide a sufficiently detailed comparison of end-to-end embedding costs against encoder-based watermarking approaches. This raises concerns about computational scalability for applications requiring watermarking of large image collections.

Questions for authors
  1. Geometric transformation generalization: The paper evaluates rotation, scaling, and translation, but does not consider perspective transformations, shearing, or substantial cropping. How would the proposed method handle these transformations, particularly in real-world camera capture scenarios? Can the authors provide additional experiments or discuss the expected limitations?

  2. Message capacity and scalability: The primary evaluation considers 64 IDs, while Appendix E.1 extends the codebook to 256 IDs with decreasing acceptance coverage. What motivates the selection of 64 IDs as the main evaluation setting? How would the method scale to significantly larger message spaces, such as 16-bit or 32-bit identifiers, while maintaining robustness and perceptual quality?

  3. Comparison with higher-capacity watermarking methods: Existing watermarking approaches can support considerably longer binary payloads. Since this paper evaluates closed-set identification rather than arbitrary bit-sequence recovery, could the authors clarify how the proposed approach compares with higher-capacity watermarking methods in terms of the capacity–robustness–distortion trade-off?

  4. Theoretical guarantees versus practical transformations: The theoretical analysis establishes norm and inner-product preservation under a specified feature-space interpolation operator. However, actual geometric transformations involve pixel-space resampling. Could the authors quantify the discrepancy between ideal fractional feature shifts and the features extracted from geometrically transformed images?

  5. Rejection threshold selection: The authors use a rejection threshold of 0.04, selected during development. Could they clarify the calibration procedure and explain how this threshold generalizes to unseen datasets and transformations? In particular, how does the method handle incorrect predictions with a score margin above 0.04, and what is the resulting trade-off between identification accuracy and acceptance coverage?

  6. Robustness against other image-processing attacks: The paper evaluates JPEG compression and geometric transformations, but does not comprehensively examine noise, blur, brightness or contrast adjustments, and repeated compression. Could the authors provide additional evaluations to determine whether the proposed message separability is preserved under these common processing operations?

  7. Novelty relative to existing methods: The proposed framework builds upon established Fourier–Mellin representations, group-invariant matching, interpolation theory, and margin-based optimization. Could the authors more explicitly distinguish the methodological contributions from these existing techniques and clarify the advantages of their approach over other learned transformation-invariant watermarking frameworks?

3 comments
AuthorsOpenReview ↗

Response to Reviewer ttvF

Thank you for the detailed review. We answer each point with new measurements.

Protocol. COCO val, 5,000 hosts × 3 models unless stated. acc = top-1 ID accuracy; coverage = acceptance at η=0.04\eta=0.04; FAR = acceptance on unwatermarked images.

W1/Q1, W4/Q6: beyond the paper's conditions

Conditionsacc (%)
Noise, blur, brightness/contrast/gamma, sharpening99.6–100
JPEG 30–90, repeated JPEG, WebP; resize 0.5/0.7599.9–100
Crop-resize 0.8/0.9; occlusion 0.8; translation ±16/32 px99.7–100
Perspective ±2/4%98.0–99.9
Offgrid / full-angle / cont. scale + JPEG, noise or blur99.4–99.9

FAR is 0.09–0.41%, and wrong acceptances are ≤ 5/15,000 per tested condition. On never-used ImageNet val, noise, blur, JPEG50, resizing, crop-resize, occlusion, translation and the two composites tested give 99.7–100% acc, with no wrong acceptance. Scope (Q1). The readout is exactly invariant to integer circular feature shifts; robustness to pixel-space transforms is evaluated empirically. Shear, anisotropic scaling and mirroring lie outside the modelled group and are not handled (e.g. 18% acc at shear 10°). §4.5 will state this boundary, and camera recapture is left to future work. Regeneration (500 × 3), acc vs. best baseline: VAE-256 93.6 vs. 79.6; img2img 0.1 98.1 vs. 96.4; VAE-512 99.9 vs. 100. At stronger img2img (0.2–0.3), MBRS is ahead. Adaptive white-box removal is not evaluated.

W3/Q4: theory vs. pixel-space resampling

We measured the discrepancy directly on real resampling. (i) The single-branch residual ϵ=∥Φa(Ay)−TδΦa(y)∥\epsilon=\lVert\Phi_a(Ay)-T_\delta\Phi_a(y)\rVert: the learned correction lowers its median 1.5–2.2× vs. the base frontend, and ϵ\epsilon correlates positively with the decoded gap drop in all 57 strata (median Spearman ≈ 0.16). (ii) For the full 11-mask decoder, Prop. A.3 guarantees an unchanged prediction if Γ>2RcdH\Gamma>2R_c d_H, where dHd_H is the orbit-Hausdorff distance between feature sets (a sufficient condition). At our operating point the measured median dHd_H values (≈ 0.2–0.4) make this bound conservative. (iii) The bound is nonetheless predictive: the ratio of 2dH2d_H to the untransformed image's full-decoder true-ID gap predicts wrong-or-rejected outputs with stratified AUC 0.925 (ε alone: 0.74). The revision adds this analysis to §3.2.

W5/Q5: threshold

η\eta was set on COCO development data and frozen. On never-used ImageNet val (1,000 × 3), coverage is 99.2–99.4% clean, 93–95% offgrid and 91–93% full-angle, with no wrong acceptance. The calibration procedure also transfers: per-model thresholds calibrated on unwatermarked COCO to FAR ≤ 10−310^{-3} (η\eta = 0.029–0.055), applied unchanged to ImageNet, give FAR ≤ 0.1% on clean and 90°-rotated unwatermarked hosts; no wrong acceptance was observed across the 19 watermarked conditions. Fig. 4b–c gives the accuracy–coverage trade-off; we will add the ImageNet curve.

W2/Q2/Q3: capacity

Why 64: registry identification; a fixed set lets 32–256-bit methods compete on one task. Larger M. Appending templates to the frozen 64-ID codebook without re-embedding keeps old images at 91.6–98.9% acc across 19 conditions for M=1024M=1024 (unchanged η\eta). Re-optimised codebooks (pilot, one model) of up to 1,024 candidates, evaluated on 512 hosts carrying IDs 0–511, keep 97.7–100% acc. The coverage drop you note comes from the fixed η\eta: gaps shrink as competitors densify. At App. E.1's η=0.023\eta=0.023, the 1,024-candidate codebook reaches 86.7–100% coverage, with no wrong acceptance and 0/2,560 unwatermarked acceptances in this batch. 16/32-bit: 2162^{16} IDs is unmeasured, and 2322^{32} exceeds an exhaustive closed-set readout. Trade-off (Q3): with residuals scaled per image toward a target LPIPS of 0.01 (approximate after 8-bit quantisation), ours leads on geometry + JPEG composites: acc ≈ 99% vs. 68–87% for a 64-bit CASIAL re-implementation and 45–86% for the best of six bit-payload baselines. In all corresponding comparisons, the 95% paired-host bootstrap confidence intervals for the differences lie above zero (jo6p W1). Bit-payload methods remain preferable for arbitrary bit strings.

W6: cost

Embedding takes 12.9 s/image on an A100, far slower than an encoder; image-to-decision takes 11.5 ms. The method targets registration-time marking (see jo6p W4).

Q7: novelty

We do not claim Fourier–Mellin, max-filtering or trigonometric interpolation. Our contributions are: (1) Eq. 3 and Prop. 3.1 turn interpolation loss into a codebook constraint: the Nyquist-free subspace is the largest on which fractional shifts preserve norms and inner products (Table 2). (2) An orbit score exactly invariant to integer feature shifts, with no pose estimation. (3) Margin protection on the half-bin view during refinement (Table 3: offgrid coverage 94.1 vs. 67.4 / 68.9%).

Thank you for the response. Some of my concerns have been addressed. However, my overall rating remains unchanged, as it reflects my assessment of the paper.

AuthorsOpenReview ↗

Response to reviewer ttvF followup

Thank you for re-reading our response. We understand that the additional experiments address only part of your concerns.

In the revised paper, we have narrowed our claims rather than presenting the new results as evidence of unrestricted robustness. Section 4.5 now distinguishes exact invariance to integer circular feature shifts from empirical pixel-space robustness, and explicitly excludes shear, anisotropic scaling, mirroring, and camera recapture from the supported scope. We also present Proposition A.3 as a conservative diagnostic bound, add quality–robustness curves with operating-point-specific threshold calibration, report the embedding cost, and label the 1,024-candidate results as a pilot study. Performance with 16/32-bit identifiers and robustness to adaptive attacks remain unverified.

We recognize that limitations remain. If there is a specific technical issue that is decisive for your rating, we would appreciate the opportunity to address it. We hope the additional evidence and clarified scope make the paper’s contribution and limitations clearer, and that you might reconsider your assessment in light of the revised version.

Reviewer_f1mLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper targets invisible image watermarking for source identification. It argues that geometric robustness has been misframed as an alignment problem, whereas what actually breaks under resampling (combined with the invisibility budget) is the separability of candidate messages. It proposes (i) a Fourier–Mellin frontend with a shift-invariant orbit readout that decodes without estimating geometry, (ii) a spectral codebook constraint (excluding radial Nyquist lines) that provably preserves norms under fractional shifts, and (iii) a two-stage embedding. On a closed-set 64-ID task over COCO/AMBER/DIV2K against nine baselines, it reports >99.5% recovery at very low distortion.

Critical strengths
  1. The performance of the model is strong, and the comparison is thorough. The set of SOTA baselines and the datasets used for evaluation are both very sufficient.
  2. The motivation for the article to enhance the watermark's own resistance to attacks makes sense.
  3. The core contribution of the codebook is theoretically well supported, and the claim is verified by adequate experiments.
Critical weaknesses
  1. The figures and tables in the paper need to be discussed in more detail; for example, the meaning of the elements in Figure 1 is unclear.
  2. The quality of the watermark is poor, and the paper should show the recognition performance of the model under higher-quality conditions.
  3. The other core contribution, the two-stage refinement, is rather engineering-oriented.
Questions for authors
  1. Could the authors clarify what each element in Figure 1 represents?
  2. Could the authors provide performance under higher-quality conditions?
  3. Is there theoretical evidence that a single-stage formulation is fundamentally worse than the two-stage design, rather than merely harder to tune?
2 comments
AuthorsOpenReview ↗

Response to Reviewer f1mL

Thank you for the review. Your concerns cover what Figure 1 shows, what the method costs in visual quality, and whether stage two is principled. We answer each with new measurements.

Protocol. COCO val, 5,000 hosts × 3 frozen models for residual scaling. The single-stage comparison re-embeds 500 hosts with one model, against a two-stage control on the same platform. acc = top-1 ID accuracy; coverage = acceptance at η=0.04\eta=0.04.

W1 / Q1: Figure 1

  • (a) After rotation or resizing the image is still recognisable, but the evidence for the hidden ID A is weaker. "ID?" marks that A's score no longer clearly exceeds the alternatives.
  • (b) The rings A and B are the orbits of two candidate IDs (each template under all its feature shifts). The decoder compares the image evidence against each orbit; by Eq. 2, the margin is half the difference of the squared distances to the two orbits. Refinement without geometric protection ("edit") makes the evidence for A and B similar, so the decision becomes uncertain.
  • (c) Avoid fragile signals: the radial Nyquist component, which half-bin interpolation erases (Eq. 3), is excluded from the codebook (§3.2). Protect ID during refinement: stage two keeps the margin penalty on the half-bin view (§3.3).

We will redraw Figure 1 with a legend mapping each glyph to its symbol and section, and expand the captions of Figs. 2–4 and Tables 1–3.

W2 / Q2: higher-quality operating points

By LPIPS, our operating point (0.0095) is already below the WAM family, RivaGAN, DWT-DCT-SVD and SS (Table 1); PSNR is where we are lower. We add two higher-quality evaluations.

Matched perceptual quality. We scaled each method's residual per image toward common LPIPS targets (same 1,000 hosts, conditions and thresholds; matching is approximate after 8-bit quantisation). At target LPIPS 0.005, about half our default distortion, ours keeps 96.5–97.3% acc on geometry + JPEG composites, against 42–70% for a CASIAL re-implementation and 32–75% for the best of six baselines. In all corresponding comparisons, the 95% paired-host bootstrap confidence intervals for the differences lie above zero (full table with jo6p W1).

Higher PSNR. We scale our stored residual, y=x+αwy=x+\alpha w, with no re-embedding:

Target PSNRacc cleanacc full-angle
38 dB99.798.4
40 dB98.695.2
42 dB94.988.0

Accuracy degrades gracefully. Gaps shrink with energy, so coverage at the fixed η\eta decreases (e.g. 61% clean at 40 dB) and η\eta should be recalibrated to the operating point. Unwatermarked FAR stays at 0.24%. The revision adds these quality–robustness curves.

W3 / Q3: is two-stage fundamentally better than single-stage?

We do not claim a theoretical result here, but the empirical comparison clearly favours the two-stage design. We ran a single-stage objective for the same 1,150 steps, with energy, both margin penalties and λp(LPIPS+1−SSIM)\lambda_p(\mathrm{LPIPS}+1-\mathrm{SSIM}) but no norm pinning, at fixed settings (lr 0.05, 500 hosts, one model):

PSNR (median)LPIPS ↓Offgrid cov.Full-angle cov.
Two-stage35.70.009093.892.6
Single, λp\lambda_p=135.70.038491.486.6
Single, λp\lambda_p=0.25 / 4 / 1635.8 / 35.4 / 34.90.042 / 0.039 / 0.03991.6 / 85.4 / 63.688.4 / 79.6 / 54.6

At a similar median PSNR (35.70 vs. 35.73 dB for λ=1), single-stage LPIPS is about 4× higher (0.0384 vs. 0.0090) and coverage is lower. Varying λp\lambda_p changes LPIPS only slightly (0.042 → 0.038), and larger λp\lambda_p costs coverage. This holds for this setting, not in general. One possible explanation, which we have not tested, is that the joint objective trades perceptual quality against residual norm, a trade-off that fixed-norm refinement removes.

We agree that the schedule is engineering in character. The paper's theoretical content is in Props. 3.1–3.2. In the tested controls, retaining the half-bin margin preserves higher offgrid coverage at comparable LPIPS (94.1% vs. 67.4 / 68.9% for clean-view / on-grid margins, Table 3).

Revisions: a redrawn Figure 1 and captions, the quality–coverage curve, the single-stage comparison, and contribution 2 restated as an empirical design.

Reviewer_f1mLOpenReview ↗

Official Comment by Reviewer f1mL

Dear author, thank you for your reply. Your response has addressed some of my concerns, but due to the significant decrease in performance at higher PSNR, I think the model's effectiveness may mainly depends on low-quality watermarks. Therefore, I maintain my rating.

Reviewer_jo6pOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies image watermark identification under geometric transformations, emphasizing message separability after resampling. It combines correlation over template orbits in a Fourier–Mellin representation, a spectrally constrained codebook, and two-stage per-image embedding that determines residual strength before refining perceptual quality at fixed raw energy. Experiments compare nine baseline configurations on 64-ID identification and examine selective decoding, message-set expansion, and color adaptation.

Critical strengths

(1) Clear motivation based on message separability The distinction between geometric misalignment and resampling-induced information loss is well motivated. The true-ID margin against the strongest competitor provides a coherent objective linking representation, codebook design, and embedding.

(2) Mechanistic justification and controlled codebook comparisons The analysis of radial Nyquist attenuation under half-bin interpolation motivates the spectral constraint. The endpoint ablation controls initialization, frontend, optimization budget, host images, messages, and perturbations, providing targeted evidence for this design.

(3) Well-controlled refinement ablations Refinement variants share stage-one residuals and match raw energy, update counts, and margin-penalty counts and weights. Experiments on 512 hosts and three frozen models help isolate the contributions of perceptual refinement and geometric margin regularization.

Critical weaknesses

(1) Limited coverage of specialized geometric baselines Despite including WAM-RST and WAM+SyncSeal, the comparison could better cover methods targeting geometric robustness. Adding CASIAL and, where compatible, SynTag would strengthen the evaluation.

(2) Unequal task adaptation Although all methods perform 64-ID identification, the proposed codebook is explicitly optimized for this task, whereas baselines use random candidate pools and hard-bit Hamming decoding. Matched visual quality and stronger baseline adaptation would improve comparability.

(3) Insufficient scalability validation Expansion is tested only up to 256 IDs through codebook reoptimization and image re-embedding. Scalability to 1024+ IDs under fixed quality and false-acceptance requirements, together with historical-image compatibility, remains unverified.

(4) Incomplete computational-cost reporting Embedding requires 1,150 Adam updates per image, while the reported 15.60 ms median measures only clean ID readout. Embedding latency, complete verification cost, and computational scaling with ID count should be reported.

Questions for authors

(1) How are the initial templates sampled, normalized, and spectrally projected? How do different initializations affect the separability of the codebook?

(2) How are t1t_1, t2t_2, and η\eta selected? What is the relationship between the actual embedding interval and the full-window decoding interval? Can distributions be reported for different distortion and capacity levels on independent data?

(3) Given fixed PSNR/LPIPS and full-library false acceptance rate, can the system support more than 1,024 IDs? How are samples that cannot be reliably embedded within the quality budget handled?

(4) After adding IDs, can recognition reliability be maintained without re-embedding old images? Is it necessary to update old templates or manage codebook versions?

(5) At different levels of distortion, which IDs are most prone to confusion? Can error acceptance evaluations be supplemented for composite distortion, unknown sources, and large codebooks?

1 comment
AuthorsOpenReview ↗

Response to Reviewer jo6p

Thank you for the detailed review; we ran the requested experiments.

Protocol. COCO val, 5,000 hosts × 3 models unless stated. acc = top-1 ID accuracy; coverage = acceptance at η=0.04\eta=0.04; CA = fraction of test images accepted with the correct ID. For CA we reuse R7 thresholds, calibrated per method/arm (per seed for ours) to a pooled empirical FAR budget of 10−310^{-3} on unwatermarked hosts 1000–1999; tests use hosts 0–999.

W1: CASIAL and SynTag

SynTag embeds during generation and does not fit our post-hoc protocol; §2 will discuss it. CASIAL has no public code. We re-implemented it (64-bit payload, 50k steps; 41.0 dB, 99.9% clean bit acc vs. 40.8 / 99.4 reported). Its native operating point has about 4× our LPIPS (0.038 vs. 0.0093), so we also compare at matched perceptual distortion. We scaled each residual per image toward LPIPS targets of 0.005 / 0.01 / 0.02 on the same hosts, conditions and thresholds; matching is approximate after 8-bit quantisation. Geometry + JPEG composites at target LPIPS 0.01 (acc % / CA):

ConditionOursCASIALBest of 6 baselines
Full-angle + JPEG7099.1 / 0.8876.6 / 0.3945.3 / 0.20
Offgrid + JPEG5099.1 / 0.9067.8 / 0.3275.8 / 0.27
Trans + JPEG7098.9 / 0.8887.0 / 0.6385.6 / 0.35
Ongrid + JPEG7099.0 / 0.8986.7 / 0.6386.3 / 0.33

CASIAL uses fixed pool/soft_ml decoding. "Best of 6" selects the best method/arm among WAM, WAM-RST, WAM+SyncSeal, MBRS, TrustMark and RivaGAN, separately for acc and CA. Across 3 LPIPS levels × 4 composites × 2 opponents × 2 metrics, all 48 pointwise 95% paired-host bootstrap confidence intervals for the differences lie above zero.

CASIAL is stronger on single geometric conditions and at its own native strength. CASIAL uses geometric augmentation, whereas our exact readout invariance applies to integer circular feature shifts; these comparisons do not isolate the causal contribution of either mechanism. Natural extensions: a larger orbit group (e.g. mirrored orbits) and augmentation-trained frontends. The revision adds this same-protocol table and scopes Table 1's statements to its nine baselines.

W2: fairer adaptation of the nine baselines

We added (a) RM(1,5) maximum-distance codewords, (b) soft ML decoding and (c) per-image PSNR matching. On composites ours still leads (acc, best arm): full-angle + JPEG70 99.7 vs. 62.2 (95.1 vs. 47.1 at 40 dB), and offgrid + JPEG50 99.7 vs. 91.3 (95.9 vs. 92.4). At matched LPIPS, against the six strongest, the lead persists (W1). On single conditions the adapted baselines are competitive, and §4.2 will make this scope explicit.

W3, Q3, Q4: scalability and historical images

Without re-embedding (Q4). We freeze the 64 templates and append new ones. At M=1024M=1024, old images keep 91.6–98.9% acc across 19 conditions, with 1 wrong acceptance in 285,000 decisions; these results use the unchanged η=0.04\eta=0.04. Old templates never change, so versioning is append-only. Re-optimised (W3; pilot, one model). Codebooks of up to 1,024 candidates, evaluated on 512 hosts carrying IDs 0–511, keep 97.7–100% acc with LPIPS ≤ 0.012. At App. E.1's η=0.023\eta=0.023, the 1,024-candidate codebook reaches 86.7–100% coverage, with no wrong acceptance in this batch. Missing an embedding target does not guarantee abstention.

W4: cost (median)

Embedding takes 12.9 s/image on an A100 (stage one 10.4 s, stage two 2.5 s), which suits registration-time marking. Image-to-decision takes 11.5 ms at batch 1 and 1.1 ms batched; the paper's 15.60 ms was measured on an RTX 3090. Each extra codeword adds ≈1.4 µs (6.4 ms at M=4096M=4096, batched).

Q1: initialisation

Templates are initialised as Gaussian, projected onto Eq. 4's support and unit-normalised after every update (100 warm-start updates at ∣kr∣≤10|k_r|\le10). Over 3 seeds, μ+=0.014±0.003\mu_+=0.014\pm0.003 vs. 0.186±0.0080.186\pm0.008 for random codebooks; coverage is 99.25 ± 0.31%, and offgrid 93.3 ± 1.0%.

Q2: t1t_1, t2t_2, η\eta

t1=0.045t_1=0.045 and t2=0.06t_2=0.06 are embedding targets for the single-feature margin γm\gamma_m on the clean and half-bin views; η=0.04\eta=0.04 thresholds the pooled gap Γ\Gamma over the full orbit and 11 masks. All were set on COCO development hosts and frozen. On never-used ImageNet val, coverage is 99.2–99.4% clean, 93–95% offgrid and 91–93% full-angle, with no wrong acceptance. COCO-calibrated thresholds give FAR ≤ 0.1% on clean and 90°-rotated unwatermarked ImageNet hosts.

Q5: confusable IDs and FAR

Per-ID accuracy is near-uniform: the worst ID scores 99.6% clean and 97.6% full-angle. A permutation test finds no significant per-ID accuracy difference under clean, JPEG70, offgrid or trans+J70 (pp = 0.09–0.42). Unwatermarked FAR is 0.09–0.41% across 40+ new conditions. No wrong ID was accepted under noise, blur, photometric changes, compression, resizing, cropping or similarity composites. Transformations outside the modelled group are discussed with ttvF.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We thank the reviewers for specific, constructive reviews. Detailed answers are in each thread; this note summarises the new evidence and the planned revisions.

Protocol. The submitted system is evaluated with its three frozen models and η=0.04\eta=0.04 unchanged. Variants are labelled: matched-quality comparisons, re-optimised codebooks (pilot), single-stage embedding, and FAR-calibrated thresholds.

New evidence

  • Robustness at matched perceptual quality (jo6p, ttvF, f1mL). We scaled each method's residual per image toward LPIPS targets of 0.005 / 0.01 / 0.02 (matching is approximate after 8-bit quantisation), keeping the same 1,000 hosts, conditions and FAR-calibrated thresholds. On geometry + JPEG composites, ours leads a CASIAL re-implementation and the best of six strong baselines (WAM, WAM-RST, WAM+SyncSeal, MBRS, TrustMark, RivaGAN; best selected per metric). At target LPIPS 0.01, ours reaches 98.9–99.1% acc, vs. 67.8–87.0% for CASIAL and 45.3–86.3% for the best baseline; correct acceptance is 0.88–0.90 vs. 0.32–0.63 and 0.20–0.35. All 48 pointwise 95% paired-host bootstrap confidence intervals for the differences lie above zero.
  • Fairer baselines (jo6p). Among the original nine baselines, ours still leads on these composites after codeword and decoder adaptation: full-angle + JPEG70 gives 99.7% vs. 62.2% at native strength and 95.1% vs. 47.1% at 40 dB.
  • Broad processing (ttvF). Noise, blur, photometric changes, JPEG/WebP, resizing, cropping, translation and mild perspective give 98.0–100% acc, with unwatermarked FAR 0.09–0.41%.
  • Held-out data (ttvF, jo6p). On never-used ImageNet val at unchanged η\eta, no wrong ID was accepted across the watermarked conditions. COCO-calibrated thresholds give FAR ≤ 0.1% on clean and 90°-rotated unwatermarked ImageNet hosts.
  • Capacity (jo6p, ttvF). Appending templates without re-embedding keeps old images at 91.6–98.9% acc across 19 conditions with 1,024 candidates. A re-optimised 1,024-candidate codebook, evaluated on 512 hosts carrying IDs 0–511, keeps 97.7–100% (pilot).
  • Design and theory (f1mL, ttvF). At similar PSNR, two-stage embedding gives about 4× lower LPIPS than a single-stage objective. The ratio of 2dH2d_H to the untransformed image's full-decoder true-ID gap predicts wrong-or-rejected outputs with stratified AUC 0.925.

Planned revisions

  1. A same-protocol comparison table (native, matched PSNR, matched LPIPS) including CASIAL. Table 1's statements will be scoped to its baselines.
  2. §4.5 will state the transformation scope: invariance is exact for integer circular feature shifts and evaluated empirically in pixel space, and shear, anisotropic scaling and mirroring lie outside the modelled group.
  3. Quality–robustness curves, with η\eta recalibrated per operating point, and Prop. A.3 presented as a conservative, diagnostic bound.
  4. The two-stage schedule presented as an empirically motivated design centred on half-bin margin protection, with embedding cost (12.9 s/image) reported for registration-time use.
  5. Figure 1 explained element by element in the introduction, and clearer captions.

We upload a revised PDF with these changes highlighted in blue. Section, table and appendix numbers in our replies refer to the submitted version.

Area_Chair_q3fjOpenReview ↗

ICLR Review Discussion – Follow-up on Author Rebuttal

Dear Reviewers,

Thank you for your time and effort in reviewing this submission.

The authors have submitted their rebuttal. Could you please take a moment to review their responses and assess whether your concerns have been adequately addressed? If you have any remaining concerns or new questions, please feel free to raise them in the discussion.

If you have already responded to the authors, please kindly disregard this reminder.

Thank you again for your valuable contributions to the review process.

Best,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.