acceptodds
Under review as a conference paper at ICLR 2027

Ephemeral Spherical Modulation: Watermarking Where the Diffusion Channel Cannot Look

Abstract

Diffusion watermarks written into the initial noise must keep it standard Gaussian while the message survives generation, image processing and inversion, and existing constructions are compared pipeline by pipeline under assumed noise models. We ask what the pipeline observes: under ideal secret frames, the same Gaussian latent for every prior-preserving watermark, so watermarks differ only in how their keys read one channel. The prior then reduces to a Gram invariant of the frame-sharing pattern. Without a noise model, the decoder reads the channel through a uniformly random consistent frame, and on acutely recovered blocks equal magnitudes minimize expected sign errors among profiles with no coordinate above a quarter of the block energy. Equal magnitudes and an exact prior force a separate frame for every block of every image: ephemeral spherical modulation (ESM), whose frames are regenerated from a key and per-image nonce, with half-normal modulation as its width-one case. On SD1.5 and SD2.1-base, ESM recovers 114/128 and 110/128 complete 512-bit messages under JPEG50, versus 69/128 and 74/128 for half-normal modulation. Channel samples from one watermark's images reproduce another's recovery rates without fitted parameters, and the tested crop leaves recovered blocks nearly orthogonal to their sources, putting every readout of the family at chance. These results concern payload recovery with the correct key and nonce, not calibrated detection, synchronization or cryptographic security.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Nonce-regenerated blockwise rotations enable spherical diffusion watermarking with ideal Gaussian-prior guarantees, a conditional decoding law, and improved JPEG recovery over a matched half-normal control.

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_z3YDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies watermarks that hide a multi-bit message in the initial noise of a diffusion model. This noise is a long list of numbers sampled from a standard Gaussian distribution. Each message bit is copied into many of these numbers and decides whether they are positive or negative. A secret key then hides this sign pattern by randomly flipping signs or rotating groups of numbers. The paper shows that, if these flips and rotations are truly random and never reused across noise blocks or images, all such watermarks give the model the same kind of noise. As a result, image edits such as JPEG compression damage the noise equally for all of them. What separates a strong watermark from a weak one is therefore only how many message bits it loses from that same damage. Using this view, the paper introduces ESM, which splits the noise into small blocks, gives all numbers in a block the same size, and rotates each block with its own secret rotation. On two Stable Diffusion models, ESM recovers the full 512-bit message after JPEG compression much more often than two prior watermarks, Gaussian Shading and Spherical Watermark.

Critical strengths
  1. The theory gives a clean picture of watermarking methods that are usually compared only by running them. It explains why Spherical Watermark's shared rotation makes the noise detectably non-Gaussian (Proposition 1), and its formula reproduces ESM's recovery rate from the measured noise of each image (Section 5.3). The authors also test the theory in a creative way. Running one method's decoder on the noise recovered from another method's images gives nearly the same success rate as running it on its own images (Section 5.1).
  2. After JPEG compression and resizing, ESM recovers most full messages. Under JPEG50, it recovers 114 and 110 of 128 messages, against at most 46 for the prior methods. The ablation (Table 2, Appendix C.3) also explains the importance of each part, with most of the gain coming from giving all numbers in a block the same size.
Critical weaknesses
  1. The baselines miss some initial-noise watermarks. PRC [1] carries 512-bit messages, and T2SMark [2] tests up to 1024 bits. PRC also proves its noise cannot be distinguished from Gaussian noise for a fixed key, which this paper leaves to future work. The paper also does not compare with other initial-noise watermarks such as WIND [3], SEAL [4], Tree-Ring [5], and Gaussian Shading++ [6]. They can be compared using a standard metric such as detection at a fixed false-positive rate.
  2. The evaluation is not sufficient to support practical use. Each setting has 128 images from 16 prompts and 8 keys. The edits are JPEG, resize, and crop. Other standard tests such as noise, blur, brightness, and rotation are missing, and diffusion regeneration uses only 32 images on SD1.5.
  3. The paper is hard to follow. For example, the abstract uses "Gram invariant" and Figure 1 uses "fiber" and "ideal secret frames" before the reader knows what they mean.

References

[1] Sam Gunn, Xuandong Zhao, and Dawn Song. An Undetectable Watermark for Generative Image Models. ICLR 2025.

[2] Jindong Yang, Han Fang, Weiming Zhang, Nenghai Yu, and Kejiang Chen. T2SMark: Balancing Robustness and Diversity in Noise-as-Watermark for Diffusion Models. NeurIPS 2025.

[3] Kasra Arabi, Benjamin Feuer, R. Teal Witter, Chinmay Hegde, and Niv Cohen. Hidden in the Noise: Two-Stage Robust Watermarking for Images. ICLR 2025.

[4] Kasra Arabi, R. Teal Witter, Chinmay Hegde, and Niv Cohen. SEAL: Semantic Aware Image Watermarking. ICCV 2025.

[5] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-Rings Watermarks: Invisible Fingerprints for Diffusion Images. NeurIPS 2023.

[6] Zijin Yang, Xin Zhang, Kejiang Chen, Kai Zeng, Qiyi Yao, Han Fang, Weiming Zhang, and Nenghai Yu. Gaussian Shading++: Rethinking the Realistic Deployment Challenge of Performance-Lossless Image Watermark for Diffusion Models. arXiv:2504.15026, 2025.

Questions for authors
  1. When the paper applies an idealized error-correcting code to its measured errors, ESM's advantage shrinks, and in one setting it is no longer significant (Appendix C.2). Does ESM still clearly outperform the other methods when a standard error-correcting code is built into the message?
3 comments
AuthorsOpenReview ↗

Response to Reviewer z3YD (1/2)

Response to Reviewer z3YD (1/2)

Thank you for the review, and for describing the theory as "a clean picture" of methods "usually compared only by running them" and the transplant test as creative. We ran new experiments on PRC, a built-in Reed–Solomon code, 13 distortions, detection at a fixed FPR and a crop control; the full tables are in the General Response (R1–R8).

W1. Missing baselines: PRC (512 bits; noise provably indistinguishable for a fixed key), T2SMark (up to 1024 bits), WIND, SEAL, Tree-Ring, Gaussian Shading++. Compare, e.g., by detection at a fixed FPR.

PRC: now evaluated, and cited in the revised PDF (official code, 512-bit payload, our pipeline, both backbones, n=128; R4).

  • JPEG50: PRC 106 / 107 vs ESM 114 / 110 (SD1.5 / SD2.1). JPEG30: 86 / 85 vs 97 / 98.
  • ESM is ≥ PRC in every cell of the five main channels; the paired intervals include 0, so we claim parity or better, not superiority.
  • On the 13 new distortions (SD2.1), PRC trails or ties ESM in 12 and is one image ahead on median 3 (32 vs 31).
  • PRC's own codeword carried by ESM's modulation rises from 21/32 to 31/32 at JPEG50 (SD2.1, 32 requests). The two are complementary.

Undetectability under a fixed key is no longer left open. New Prop. J.2 (R3): if HMAC-SHA256 and ChaCha20 are PRFs, the key is secret, the Gaussian draw is fresh per image and nonces are unique, then ESM latents for any number of images are computationally indistinguishable from latents built with truly random frames, which in real arithmetic are exactly N(0,I)\mathcal N(0,I) for every message (Prop. 1(i)); we do not bound floating-point sampler effects. PRC builds undetectability into its code; ESM obtains it from per-image frames keyed by a unique public nonce (transport and verification in R2) and adds an exact recovery law (Thm. 3).

T2SMark (Yang et al., NeurIPS 2025).

  • How it works: watermark-carrying coordinates are sampled from the tails (∣z∣>τ\vert z\vert > \tau) with signs set by the bits, and each bit is read as the sign of a projection. The central zone is sampled from the truncated normal. A 16-bit session key is recovered from a first stage.
  • Evaluation: its main setting is 256 bits, and its capacity study reports bit accuracy up to 1024 bits. Its limitations section notes that embedding the session key in the truncated tails "introduces subtle distributional anomalies, which can be detected".
  • On capacity, ESM is evaluated at 1024 bits in App. D on complete messages: under JPEG50 it recovers 47/64 complete 1024-bit messages vs 14/64 for HN, whereas T2SMark's capacity study reports bit accuracy. The revised PDF will add T2SMark to §2; it is not among the methods run for R1–R8.

Tree-Ring, WIND and SEAL answer a different question.

  • Tree-Ring writes a key pattern into the Fourier spectrum of the noise and tests for its presence (zero-bit).
  • WIND generates the noise from a hash of a secret salt and one of N indices, adds a Fourier group identifier, and identifies an image by searching the indices.
  • SEAL seeds noise patches from a locality-sensitive hash of the semantic embedding and counts matching patches.

They answer "is this ours / which index?", not "which 512-bit message?", so they carry no arbitrary payload to compare on. On your suggested metric, detection at a fixed FPR, ESM already reaches the ceiling on clean, JPEG50, JPEG30 and resize (128/128, below), so on these channels a zero-bit scheme can at most tie it while carrying no payload. The revised PDF will discuss them in §2.

Gaussian Shading++ is a deployment layer around Gaussian Shading. It carries a per-image seed in a separate PRC-coded channel next to GS's payload channel and adds signatures for third-party verification. Its modulation is GS's, which our GS baseline implements with the authors' code (App. A.4). It also shows how ESM's per-image nonce can be carried in band.

Detection at a fixed FPR, now reported (R2). New Prop. J.1: under ideal frames, a wrong nonce or an unwatermarked image gives exactly Bin(512,1/2)\mathrm{Bin}(512,1/2) agreement with any candidate, so thresholds are analytic. For the implementation, 640,000 wrong-nonce decodes per backbone match this (FPR 0.0092–0.0100 against an exact tail of 0.0095). At aggregate FPR 10−310^{-3}, with up to 10610^6 candidate nonces, ESM is detected in 128/128 images on clean, JPEG50, JPEG30 and resize on both backbones (95% Clopper–Pearson lower bound 97.2%). HN and GS also reach 128/128, so this metric is saturated for the stronger payload watermarks on these channels. Complete 512-bit recovery is the endpoint that separates them, which is why the paper uses it.

AuthorsOpenReview ↗

Response to Reviewer z3YD (2/2)

Response to Reviewer z3YD (2/2)

Q1. With an idealized ECC, ESM's advantage shrinks and in one setting is not significant (App. C.2). Does ESM still clearly outperform when a standard ECC is built into the message?

Yes. We built a standard Reed–Solomon code into the message and ran generation and inversion on both backbones (R1): RS(128,64), ×16 repetition, the same 512 net bits, standard hard-decision decoding, n=128.

backbone, channelESM+RSHN+RSESM − HN (points)
SD1.5 JPEG5011996+18.0 [+6.2, +32.0]
SD2.1 JPEG50124107+13.3 [+1.6, +28.9]
SD1.5 JPEG3011673+33.6 [+16.4, +52.3]
SD2.1 JPEG3011870+37.5 [+18.0, +57.0]

All four intervals exclude 0, including SD2.1 JPEG50, the backbone and channel of the non-significant C.2 cell. ESM also beats the rotated-HN control, which shares its frames, in all four (R1). Against GS and SW, our extension of App. C.2's main-run RS analysis (validated image by image on the embedded SD1.5 run) keeps every JPEG interval above 0 up to t=16t=16 with their native decoders, e.g. ESM − GS +10.9 [+3.1, +21.1] (SD1.5) and +10.9 [+2.3, +21.9] (SD2.1) at JPEG50, t=16t=16. With idealized soft-assisted decoding (the GMD criterion, which a real errors-and-erasures RS decoder matched in every embedded cell), the largest net payload that keeps ≥ 99% JPEG50 exact recovery is 448 bits for ESM vs 128 for HN and 128 for GS on SD1.5, and 320 vs 16 and 256 on SD2.1 (R1); SW never reaches it.

W2. Evaluation is not sufficient: 128 images (16 prompts × 8 keys); only JPEG, resize and crop; no noise, blur, brightness or rotation; regeneration uses only 32 SD1.5 images.

  • Noise, blur, brightness, rotation and more (R5). We added 13 new channels on both backbones (n=32 per cell): JPEG70/10, Gaussian noise σ=0.02/0.05, blur r=1/2, median 3/5, brightness ×1.3/×0.7, contrast ×1.3 and rotation 2°/5°.
    • Among ESM, HN, GS and SW, no method exceeds ESM's count in any of the 26 cells. Examples (ESM / HN / GS / SW, SD1.5): JPEG10 14 / 4 / 4 / 0; blur r=2 32 / 24 / 19 / 12; noise σ=0.05 25 / 13 / 13 / 13.
    • With n=32 several intervals are wide, so we read the table as consistent support. Intervals excluding 0 include blur r=2 vs HN on both backbones and noise σ=0.05 vs HN on SD1.5.
    • PRC (SD2.1) trails or ties ESM in 12 of the 13 and is one image ahead on median 3 (all intervals include 0).
  • Rotation and the crop are synchronization failures (R5, R6). Rotation defeats every method tested, PRC included (0/32), though ESM keeps the highest bit accuracy (68–72% at 2°). When the paper's crop is pasted back at its original position without rescaling, ESM recovers 123/128 and 125/128 complete messages, against 0/128 for HN and SW. So for ESM the crop as tested fails through rescaling and misalignment, not through the 36% of the image that is lost. Geometric resynchronization is orthogonal to the modulation, and §6 of the revised PDF will scope it explicitly as a synchronization problem.
  • Sample size. The ESM, HN, GS and SW comparisons are paired (same key, prompt, payload, Gaussian draw and nonce), with a crossed key × prompt bootstrap, and both prespecified primary intervals exclude 0. The direction of the effect holds across:
    • two backbones;
    • an independent 1,024-bit study (47/64 vs 14/64);
    • the independent embedded-RS executions;
    • a re-execution on identical inputs (SD1.5 JPEG50: ESM 114 vs HN 64);
    • 15 of 16 key × backbone groups (one tie).
  • Regeneration.
    • At strength 0.1, ESM keeps 24/32 vs 16/32 (interval [−12.5, 62.5], not significant at n=32), and its bit-accuracy gain, +0.641 [+0.036, +1.532], excludes 0.
    • At 0.3, the bit-accuracy difference is significant: +3.326 [+1.312, +5.359], 98.09% vs 94.76%, i.e. about 2.7× fewer bit errors. The regeneration study keeps the submission's SD1.5, n=32-pair design; at that size the bit-level difference already excludes 0 at both strengths, while the paired PSNR, SSIM and LPIPS intervals all contain 0, so the attack distorts both methods alike (App. C.10).

W3. Hard to follow: the abstract uses "Gram invariant", and Fig. 1 uses "fiber" and "ideal secret frames", before they are defined. The revised PDF will fix each instance:

  • The abstract will say "the inner products of blocks that share a secret rotation" and "idealized, independent secret rotations".
  • Fig. 1's caption will define the fiber ("the set of rotations that map the codeword to the transmitted block, which the decoder cannot tell apart") and ideal frames ((A1)).
  • An "ESM at a glance" paragraph, Algorithm 1 and the pipeline figure will come before the theory.
  • Every term will be defined at first use, with a notation table (General Response 3/3).

If a point still seems insufficient, please tell us which one, and we will respond to it before the discussion ends. If these results address the weaknesses you listed, we would appreciate it if you would reconsider your rating.

Thank you for the response. It addressed my error-correction concern. However, the new results show that ESM performs about the same as PRC, and I still find the paper somewhat hard to follow. Therefore, I keep my score.

Reviewer_a7xtOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper develops a framework for initial-noise watermarks in latent diffusion models, representing each scheme through a code and a pattern of shared orthogonal frames. Under independent Haar-distributed frames and a generation–processing–inversion channel independent of the key and message, it characterizes Gaussian prior preservation through the Gram matrices of blocks sharing a frame. It also derives angular decoding-error laws and establishes conditions under which equal-magnitude profiles improve raw sign recovery. These results motivate Ephemeral Spherical Modulation (ESM), which combines equal-magnitude coordinates and chi-distributed radii with a separate frame for each block and image. The implementation regenerates frames from a secret key and a per-image nonce. Experiments on SD1.5 and SD2.1-base compare ESM with half-normal modulation (HN) and adapted baselines using 512-bit payloads. ESM substantially improves complete-message recovery under JPEG compression. Ablations, decoder-transplant experiments, coding analyses, and distribution diagnostics investigate the source and limitations of these gains. The paper explicitly focuses on payload recovery with the correct key and nonce, rather than end-to-end detection or cryptographic security.

Critical strengths

S1. A useful theoretical characterization that yields a concrete design principle. Proposition 1 reduces Gaussian prior preservation to a condition on the Gram matrices of frame-sharing blocks, covering several frame-sharing configurations within one framework. Together with Theorem 4, it explains both the decoding benefit of equal magnitudes and the distributional problem introduced by sharing frames under this construction. ESM follows the resulting design principle, with approximately 50 ms of CPU overhead per encoding or decoding operation (Table 25). The distinction between exact prior preservation under ideal Haar frames and its finite-key implementation remains important. Nevertheless, the theoretical characterization is a substantive contribution beyond moment matching or empirical classifier tests.

S2. A carefully controlled primary comparison with informative ablations. The study uses paired requests, shared payloads and Gaussian draws, a prespecified primary contrast, and crossed bootstrap intervals over keys and prompts. The widen-then-equalize ablation supports magnitude equalization as the main source of improvement: the reported gains are +14.8 points [−2.3, 34.4] for frame widening and +24.2 points [10.9, 38.3] for equalization. The interleaving control further supports this interpretation. The paper also reports negative results and execution discrepancies rather than omitting them. These choices make the ESM–HN comparison substantially more interpretable and support the paper’s mechanistic contribution.

Critical weaknesses

W1. The practical significance of the payload-recovery gain remains uncertain. The evaluation supplies the correct per-image nonce externally, while a deployment would need a reliable mechanism for recovering or retrieving it. An incorrect nonce reconstructs incorrect frames, but nonce handling, calibrated detection, and candidate-search false positives are not evaluated. These omissions are acknowledged and do not invalidate the payload-channel analysis; they limit its implications for a complete watermarking system. Robustness also has clear limits: all methods recover 0/128 messages after the tested crop, and ESM recovers 3/32 under regeneration strength 0.3. Moreover, the headline complete-message recovery advantage depends on the coding configuration. With the idealized Reed–Solomon code at t=8t=8 and 384 net bits, the SD2.1-base JPEG50 advantage over HN falls to +4.7 points [−3.9, 14.1], while the reported bit-accuracy gains are 0.63 and 0.65 points. The results therefore establish a useful improvement under the tested configuration, but not yet its magnitude under practical coding and verification protocols.

W2. Missing comparisons, adapted baselines, and limited evaluation scale weaken the empirical positioning. PRC (Gunn, Zhao, and Song, ICLR 2025) is directly relevant to initial-noise payload recovery and undetectability, yet is neither cited nor evaluated. A comparison would also clarify the distinction between ESM’s ideal-frame distributional guarantee and computational undetectability with a finite key. Gaussian Shading and Spherical Watermark are evaluated through adapters at a common 512-bit operating point, rather than as complete published systems. This is useful for controlled component comparisons, but the paper does not establish that the adapters reproduce the original methods at their default settings. Consequently, the reported gains should not be interpreted as demonstrated superiority over the full baseline systems. The study uses 128 requests per backbone, a limited distortion set, and a small regeneration test. Several exact-recovery intervals are wide. Likewise, “44 of 46 transplant intervals contain zero” does not by itself establish equivalence, although the small observed bit-accuracy differences are informative. Fidelity diagnostics also remain inconclusive: FID uses a small generated reference set, and the SD2.1-base CLIP classifier distinguishes ESM from Gaussian-noise images with AUC 0.582 [0.551, 0.622]. The proposed explanation involving separate executions is plausible but untested. These limitations affect the breadth of the empirical conclusions more than the central theoretical contribution.

Questions for authors

Q1. PRC comparison and guarantee scope. How does ESM’s exact Gaussianity under ideal Haar frames relate to computational undetectability under a finite key? Can PRC be compared under the same pipeline, payload size, and channels, reporting both exact recovery and bit accuracy?

Q2. Nonce handling and verification. What verification setting is intended for images of unknown origin? If the nonce is carried in-band, what is the resulting end-to-end recovery rate? If verification searches NN stored nonces, how are detection scores calibrated, and what are the TPRs at aggregate FPRs of 1% and 0.1% as NN varies?

Q3. Baseline validation. Have the GS and SW adapters been validated against their original implementations at the published default settings? Were the LAW/LAW-M reconstructions in App. C.5 similarly checked? Please distinguish validated reproductions from component-level adaptations.

Q4. Interpretation of crop failure. The evaluated crop also rescales the retained region. Can a control preserve the original scale and restore the original canvas alignment, for example by padding at the known crop offset? This would help separate information loss from geometric misalignment.

Q5. Execution variability and distribution diagnostics. What explains the 69-versus-64 HN recovery discrepancy on identical latents, and how many independent executions were performed? Can interleaved ESM and Gaussian-noise control generations under matched execution conditions test the proposed explanation for the SD2.1-base classifier result?

5 comments
AuthorsOpenReview ↗

Response to Reviewer a7xt (1/4)

Response to Reviewer a7xt (1/4)

Thank you for the review, and for recognizing the Gram-invariant characterization as "a substantive contribution beyond moment matching or empirical classifier tests" and the controlled design with its reported negative results. We answer every point below; the full tables are in the General Response (R1–R8).

W1(a) & Q2. Nonce handling, calibrated detection, candidate-search false positives; TPR at aggregate FPR 1% / 0.1% as NN varies; end-to-end rate with an in-band nonce.

Exact null (new Prop. J.1, proof in [Rev: App. J.1]).

  • Under (A1), suppose the decoding frames are independent of the image's latent. This is the case for a wrong nonce, and for an unwatermarked image, whose latent never depended on the key. Then agreement with any candidate message, even one chosen after seeing the image, is exactly Bin(512,1/2)\mathrm{Bin}(512,1/2).
  • Proof idea: a Haar frame is invariant under sign flips of any subset of coordinates; bits use disjoint coordinates; and the soft sum is odd. So the decoded sign vector is uniform, whatever the channel did. The probability is over the secret frames.
  • For the implemented ChaCha20 frames, the statement holds up to the PRF advantage and the normal sampler's numerics. We checked it directly: over 640,000 wrong-nonce decodes per backbone, ESM's χ2\chi^2 goodness-of-fit p≥0.135p\ge0.135 in all ten model × channel cells, and the empirical FPR is 0.0092–0.0100 at τ = 283 (exact tail 0.0095) and 0.00063–0.00095 at τ = 292 (exact tail 0.00084).
  • The GS adapter's null deviates slightly, as the proposition predicts. Its 16:16 tie rule makes its decision non-odd, so a null bit is 1 with probability 0.430. Its tail FPRs are still at nominal.

TPR for a log of NN issued nonces. τ(N, α) is the smallest ss with N⋅P(Bin(512,1/2)≥s)≤αN\cdot P(\mathrm{Bin}(512,1/2)\ge s)\le\alpha:

αN=1N=1N=103N=10^3N=106N=10^6
10−210^{-2}283305320
10−310^{-3}292311324

ESM's lowest correct-nonce agreement over 128 images is 422 / 495 under JPEG50 (SD1.5 / SD2.1) and 373 / 464 under JPEG30. So for both α, all three NN and both backbones, TPR is 128/128 on clean, JPEG50, JPEG30 and resize. The 95% Clopper–Pearson lower bound is 97.2% per cell.

HN and GS are also at 128/128. Detection at a fixed FPR is therefore saturated on these channels, which is why the paper keeps complete 512-bit recovery as the endpoint that separates modulations. Under the crop, TPR is ≤ 6.2% at N=1N=1 and 0 for N≥103N\ge10^3 for every method, consistent with the synchronization diagnosis (W1(c)).

Cost. A full candidate decode costs about 50 ms (Table 25). A 16-of-128-frame partial decode is about 6–7× cheaper than a full decode in the same timing harness; it has the same exact null and a TPR ≥ 98.4% at N=106N=10^6, α=10−3\alpha=10^{-3} under JPEG50/JPEG30. Large logs therefore call for bounded candidate sets (per user or per time window), or for transporting the nonce with the image.

In-band nonce. The deployments we target carry the nonce in provenance metadata, as an IV travels with a ciphertext, with the log test above as the fallback. We report no end-to-end in-band rate, because no in-band header was run in the discussion period. Where in-band transport is needed, the established route is Gaussian Shading++'s: the per-image seed is protected by a pseudorandom code in a separate channel. Prop. 1 explains why that is needed: a key-fixed sign mask is exactly Gaussian only without repetition, and any repeated header bit reintroduces a fixed sign pattern across images.

W1(b). "An incorrect nonce reconstructs incorrect frames." That is precisely the property that makes verification calibratable. With an incorrect nonce, ESM decodes to an exactly uniform 512-bit string under (A1), and empirically indistinguishable from one for the implementation, so a wrong candidate never accumulates evidence beyond the binomial tail. Because frames are never shared across images, ESM also exposes none of the cross-image inner products that a shared frame reveals (Prop. 1(ii)).

AuthorsOpenReview ↗

Response to Reviewer a7xt (2/4)

Response to Reviewer a7xt (2/4)

W1(e). The advantage depends on the coding configuration: with idealized RS at t=8t=8 (384 net bits), the SD2.1 JPEG50 advantage over HN falls to +4.7 [−3.9, 14.1]; the BA gains are only 0.63 / 0.65 points.

The cell you cite is one of 24 JPEG cells in Table 9; all 12 at t=2t=2 and 11 of 12 at t=8t=8 exclude 0. In the extended sweep for App. J.3 (t up to 31), the SD2.1 JPEG50 ESM–HN interval includes 0 for every t≥4t \ge 4.

Because the concern is practical coding, we tested a real code embedded in the payload on both backbones (R1). For SD2.1 JPEG50, the backbone and channel of the cell you cite, ESM+RS recovers 124/128 vs 107/128 for HN+RS, i.e. +13.3 [+1.6, +28.9]. All four ESM − HN intervals and all four ESM − rot.HN intervals in R1 exclude 0. Three further points (details in R1):

  • Prediction. Predictions of the embedded layout from the main run's channel samples have intervals containing every observed ESM and HN JPEG count on both backbones. The SD2.1 prediction of the gap (+9.4) is smaller than the observed +13.3, so we take the predicted +9.4, rather than the observed +13.3, as the expected SD2.1 gap.
  • Soft decoding. A real errors-and-erasures decoder on the SD1.5 embedded run gives ESM − HN +15.6 [+3.9, +30.5]. With idealized soft-assisted RS, the largest net payload with ≥ 99% JPEG50 exact recovery is 448 bits for ESM vs 128 for HN (SD1.5), and 320 vs 16 (SD2.1). GS, which gains most from soft information, comes within the noise of ESM at t=8t=8 with soft-assisted RS: +4.7 [−0.8, 11.7] and +3.9 [−1.6, 12.5].
  • BA. 0.63 / 0.65 points correspond to HN's bit-error rate of 0.81% / 0.77% falling to 0.18% / 0.13% with ESM. That is why outer codes keep the gap.

W1(c) & Q4. All methods recover 0/128 after the crop. Can a control that preserves scale and restores alignment separate information loss from misalignment?

We ran this control (R6). With the paper's 410² crop pasted back at its original offset, without rescaling:

  • ESM recovers 123/128 (SD1.5) and 125/128 (SD2.1);
  • HN and SW recover 0/128;
  • GS recovers 10–13/128 (81–90 with its soft decoder).

ESM − HN is +96.1 and +97.7 points, with lower interval bounds ≥ 85.9. About 36% of the image area is lost, yet ESM still recovers 96–98% of complete messages. For ESM the crop as tested thus fails through rescaling and misalignment, as §5.3 inferred from the angles, not through lost area. Among the four methods tested, ESM is also the most robust to this content loss.

Rotation by 2° / 5°, also new, defeats every method we ran, including PRC on SD2.1. It is a synchronization failure of the same kind, which §6 of the revised PDF will state as orthogonal to the modulation.

W1(d). ESM recovers 3/32 at regeneration strength 0.3. HN recovers 0/32 there. The per-bit evidence is clear even where complete messages are not:

  • At 0.3 the bit-accuracy difference is significant: +3.326 [+1.312, +5.359] (98.09% vs 94.76%). ESM's per-bit error rate is about 2.7× lower, and that is the quantity an outer code converts.
  • At 0.1, ESM keeps 24/32 vs 16/32 at matched PSNR/SSIM/LPIPS, interval [−12.5, 62.5], so this is not significant at n=32.

The regeneration study keeps the submission's SD1.5, n=32-pair design (the embedded code was not run under regeneration); at that size the bit-level difference above already excludes 0. §6 scopes removal attacks such as regeneration explicitly.

W2(d). "44 of 46 transplant intervals contain zero" does not establish equivalence. The revised PDF will say "agreement within sampling error", not equivalence, and will report the largest |Δ| with its interval per cell. The evidence that carries weight is:

  • bit-accuracy differences are at most 0.49 points in every cross-construction cell;
  • the HN readout reproduces all 1,280/1,280 logged HN outcomes;
  • on identical channel samples the order HN < rotated HN < ESM holds in every column (e.g. 47.7 / 63.3 / 89.1% on SD1.5 ESM images).

W2(c). 128 requests per backbone, a limited distortion set, a small regeneration test; several wide intervals.

  • Distortions. We added 13 channels (noise, blur, median, brightness, contrast, JPEG70/10, rotation) on both backbones, n=32 per cell (R5). Among ESM, HN, GS and SW, no method exceeds ESM's count in any of the 26 cells; PRC is covered in part 3/4. At n=32 we read these cells as consistent support, not as 26 separate significance claims.
  • Design. The primary contrast is paired and prespecified, with a crossed key × prompt bootstrap. Both primary exact-recovery intervals exclude 0 ([17.97, 52.34] and [11.72, 46.88]).
  • Consistency. The direction of the effect holds across two backbones, an independent 1,024-bit study (47/64 vs 14/64, App. D), the earlier pilot (26 vs 12 of 32), the independent embedded-RS executions (R1), a re-execution on identical inputs (114 vs 64) and 15 of 16 key × backbone groups (one tie, Table 37).
AuthorsOpenReview ↗

Response to Reviewer a7xt (3/4)

Response to Reviewer a7xt (3/4)

W2(a) & Q1. PRC is not cited or evaluated. How does exact Gaussianity under ideal Haar frames relate to computational undetectability under a finite key? Can PRC be compared in the same pipeline, reporting exact recovery and BA?

We will cite and discuss PRC (Gunn, Zhao & Song, ICLR 2025) in §2 and App. H.

(1) Theory: new Prop. J.2 (R3) [Rev: App. J.2, full proof]. The assumptions are that HMAC-SHA256 and ChaCha20, as used, are PRFs; the key is secret; each image's Gaussian draw is fresh; and nonces are unique (a counter, or random 96-bit values with collision probability ≤ q2/297q^2/2^{97}).

Under these assumptions, an efficient adversary that adaptively picks messages and the public nonces and observes qq latents, or images generated from them, has at most the PRF advantage of distinguishing ESM from ESM with a truly random function. With a random function every frame is Haar and fresh, so Prop. 1(i) makes each latent exactly N(0,I)\mathcal N(0,I) in real arithmetic. We do not bound the floating-point difference between normal samplers.

The two notions are layered:

  • Prop. 1 is an exact, information-theoretic statement in the ideal model, and it comes with a converse: for an equal-magnitude code, any frame sharing is singular.
  • Prop. J.2 transfers that statement to a finite key under standard symmetric-key assumptions. Nonce uniqueness is necessary, because a repeated nonce re-creates the lattice of Prop. 1(ii). That is the structure the Gram-grid detector catches with AUC 1.000 on a synthetic shared-frame construction (Table 18).

PRC obtains undetectability through its code and needs no per-image state. ESM obtains it through per-image frames, and in addition provides results that PRC does not address: a channel-free recovery law (Thm. 3) and the optimality of equal magnitudes among spread profiles (Thm. 4).

(2) Experiment: PRC in our pipeline (official codec, 512-bit payload, n=128 per backbone; full table in R4).

  • Exact recovery. ESM's count is ≥ PRC's in every cell of the five main channels. At JPEG50 it is 114 vs 106 (SD1.5) and 110 vs 107 (SD2.1); at JPEG30, 97 vs 86 and 98 vs 85. The JPEG differences have intervals that include 0, so on exact recovery we claim parity or better, not superiority.
  • Bit accuracy. PRC never returned a wrong message; it abstains (22/128 erasures at SD1.5 JPEG50, 42/128 at JPEG30). ESM always returns a message, with bit accuracy 99.82% / 99.60%, which an outer code can use. Because PRC abstains, bit accuracy is not comparable across the two decoders, so we compare exact recovery and erasures.
  • PRC's native setting (fp32, guidance 3, 100-step decoder inversion; SD2.1, 32 pairs): both methods recover 32/32 on all non-crop channels.
  • 13 new distortions (SD2.1, n=32): PRC trails or ties ESM in 12 and is one image ahead on median 3 (32 vs 31); all intervals include 0.

(3) The two are complementary. PRC's own codeword placed on ESM's modulation and decoded by the official PRC decoder (SD2.1, 32 requests) rises from 21/32 to 31/32 at JPEG50 and from 15/32 to 28/32 at JPEG30 (R4). ESM is a modulation layer that can carry PRC's undetectable code.

W2(b) & Q3. Were GS, SW and LAW/LAW-M validated against the originals at their published settings? Please distinguish validated reproductions from component-level adaptations.

The revised PDF will label every baseline explicitly [Rev: App. A.4, J.8]:

  • GS adapter. A component-level adaptation of the authors' code: the official ChaCha20 tiling, bit packing, XOR and sign assignment, at 512 bits × 32 repetitions. The revised PDF will label it as such, distinct from a reproduction of the repository's own configuration. Its auxiliary soft decoder is stronger than the original (68 / 79 vs 46 / 45 JPEG50 messages); ESM's 114 / 110 exceed both.
  • SW: checked against the SW authors' own code in its shipped configuration (512-bit message, N=31; R8). The author code recovers 32/32, 20/32 and 15/32 (clean / JPEG70 / JPEG50). Our adapter at the same fp32 / sampled-posterior numerics recovers 109/128, 47/128 and 32/128 (Table 15). The author code is thus the stronger SW implementation, and the revised PDF will adopt it as SW's reference in §5.4 and the Table 3 SW entries for the settings it covers (SD1.5; clean, JPEG70, JPEG50). The paired, per-request analyses (Table 5, R5–R7) keep the adapter, because they need ESM's requests. ESM's margin holds at the same numerics: 115/128 (89.8%) vs 15/32 (46.9%) at JPEG50, and 121/128 (94.5%) vs 20/32 (62.5%) at JPEG70 (different requests, so descriptive).
  • LAW / LAW-M. Paper-based exploratory reconstructions whose fidelity is unverified (as App. C.5 states). They support no headline claim and will be labelled accordingly.

The headline ESM–HN contrast involves no adapter, and PRC uses its official code.

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

This paper proposes ESM, an image watermarking method for diffusion models that embeds messages into the initial noise using equal-magnitude encoding within blocks, secret random rotations specific to each block and image, and repeated embedding. During decoding, the method reconstructs the rotations using a secret key and an externally supplied per-image nonce, reverses the rotations on the recovered noise, and extracts the message through soft aggregation. The reported experimental results show promising performance.

Critical strengths

The paper provides a unified analysis of amplitude allocation and rotation sharing in initial-noise watermarking. It uses Gram invariants to characterize conditions for preserving the joint Gaussian distribution and recovery angles to explain decoding performance and failure modes. Its main value lies in making watermarked noise more consistent with the Gaussian prior and recovery performance easier to analyze.

Critical weaknesses
  1. The reliance on an external nonce leaves the verification setting unclear. Experiments provide the correct image–nonce association without evaluating how the nonce is obtained for an unknown or modified image. If image retrieval is required, it may already identify the corresponding generation record. The authors should clarify what additional evidence watermark decoding provides and account for the differing information requirements when comparing with SW.

  2. The robustness gains over SW are not sufficiently isolated. ESM changes rotation sharing, padding, and decoding simultaneously. Existing ablations mainly explain improvements over HN. A direct comparison that preserves SW’s other components and changes only the rotation-sharing scheme is needed to establish the contribution of the central algorithmic change.

  3. The presentation is difficult to follow. Abstract concepts precede a sufficiently concrete explanation of the method. The connections between the ideal random-rotation assumptions, finite-key implementation, and empirical recovery performance should be made clearer.

Questions for authors
  1. How can a verifier obtain the correct nonce for an unknown or modified image? If retrieval is required, what additional verification capability does watermark decoding provide beyond locating the corresponding generation record?

  2. How does performance change when SW’s padding, binary encoding, and decoder are retained, and only the rotation-sharing scheme is modified?

  3. What mechanism explains the robustness gains from independent rotations? Can the improvements be attributed to reduced error correlation, rather than changes to padding or decoding?

4 comments
AuthorsOpenReview ↗

Response to Reviewer ugTx (1/2)

Response to Reviewer ugTx (1/2)

Thank you for the review, and for recognizing the "unified analysis of amplitude allocation and rotation sharing" and the use of recovery angles to explain performance and failure modes. Full tables are in the General Response (R1–R8); here we answer each point.

W1 & Q1. How does a verifier obtain the nonce; what does decoding add beyond retrieval; do the comparisons with SW account for differing information?

(a) What the nonce is and why it exists.

  • The nonce is a public, unique 12-byte value per image, like an IV in authenticated encryption (§3.3). All secrecy is in the 32-byte key.
  • It is required by the theory, not chosen for convenience. Prop. 1(iii) proves that equal magnitudes give an exactly Gaussian latent only if every block of every image has its own frame.
  • For an equal-magnitude code, any frame reuse, across images or across a key as in SW, puts block inner products on a lattice. That makes the latent singular w.r.t. N(0,I)\mathcal N(0,I) (Prop. 1(ii)).

(b) Information requirements.

  • HN and GS: in our main run, ESM, HN and the GS adapter receive the same key and the same per-image nonce for every request (App. A.1, A.4). The ESM–HN and ESM–GS comparisons are therefore information-matched.
  • SW: we tested this directly. Giving SW the same per-image nonce and per-block frames as ESM, with everything else unchanged, puts it on an equal information footing with ESM. The gap persists: at JPEG50, ESM recovers 114 / 110 vs SW per-block 31 / 27 (SD1.5 / SD2.1; R7a). The 12-byte nonce is therefore not the source of ESM's advantage. On SW's own channel samples, ESM's readout is predicted to recover 86.7% / 91.8% (Table 5, a descriptive column), while SW's observed rate is 27.3%.

(c) How a verifier obtains the nonce.

  1. With the image. The nonce travels in provenance metadata, as an IV travels with a ciphertext. No database is needed.
  2. From the provider's log, if metadata is stripped. The verifier tests the image against the issued nonces, preferably a bounded set (per user or per time window). The new Prop. J.1 calibrates this test. Under ideal frames, a wrong nonce or an unwatermarked image gives exactly Bin(512,1/2)\mathrm{Bin}(512,1/2) agreement with a candidate message, so the threshold for a log of size NN is analytic. 640,000 wrong-nonce decodes per backbone confirm that the implementation matches this law. ESM is detected in 128/128 images on clean, JPEG50, JPEG30 and resize for N≤106N\le10^6 at aggregate FPR 10−310^{-3} (R2). This is the same identify-by-search structure that WIND uses over its N noise indices.
  3. In band. Gaussian Shading++ protects its per-image seed with a pseudorandom code in a separate channel, and ESM's 96-bit nonce could be carried the same way. A plain key-fixed sign mask would be exactly Gaussian only without repetition, so a robust header needs such a pseudorandom code. We did not run an in-band header in the discussion period.

(d) What decoding adds beyond retrieval. Similarity retrieval is a separate system. It requires storing image features, and its false-match rate is uncalibrated. Watermark decoding is a keyed test whose false-positive rate is known in closed form (Prop. J.1). It reads the 512-bit payload, e.g. a signed identifier, out of the image content after JPEG, resizing, noise, blur and filtering (Table 3, R5), and the log holds only 12-byte nonces. The image itself thus provides the mark, the payload and an exact pp-value.

W3. Presentation: abstract concepts precede the method; the links between ideal rotations, the finite key and empirical recovery should be clearer.

The revised PDF will put the method first ("ESM at a glance", Algorithm 1 and the pipeline figure before Prop. 1) and define every term at first use (General Response 3/3). A bridge paragraph will make your three links explicit:

  • Ideal → finite key: Prop. J.2 (R3).
  • Finite key → measurement: held-out sampler tests against Gaussian noise (AUC 0.480 [0.460, 0.501], MMD p=0.992p=0.992; Table 18).
  • Ideal-frame theory → empirical recovery: the angular law predicts the implemented key's complete-message recovery within 3.04 points over ten settings, with no fitted parameter (e.g. 89.04% predicted vs 89.06% observed, SD1.5 JPEG50).
AuthorsOpenReview ↗

Response to Reviewer ugTx (2/2)

Response to Reviewer ugTx (2/2)

W2 & Q2. Isolate the change from SW: keep SW's padding, binary encoding and decoder, and change only rotation sharing.

We ran exactly this experiment (R7a). "SW per-block" keeps SW's 512 padding bits, XOR mixing, 31-fold repetition, χ radii (identical draws per request) and its native padding-XOR + majority decoder. Only the key-fixed frame is replaced by an independent frame per block and per image. The outcome is consistent with the paper's theory:

  • The prior becomes exact. With one frame per block and independent χ radii, SW's latent is exactly Gaussian under (A1) (Prop. 1(i)), and the lattice of Prop. 1(ii) disappears.
  • Recovery stays within sampling error. At JPEG50 SW per-block recovers 31/128 (SD1.5) and 27/128 (SD2.1), against 35/128 for shared-frame SW. All eight paired intervals contain 0, and the JPEG50 point differences (−3.1 and −6.2 points) are under a tenth of the +64.8-point ESM − SW per-block gap. ESM recovers 114 and 110.

This is the division of labour the paper proves. The frame pattern governs the prior (Prop. 1). By Prop. 2(i) every prior-preserving code sees the same channel, so recovery is governed by the code and readout (Thms. 3–4). For the prior the frame pattern is forced, for robustness it is not the lever, and ESM takes the only exact pattern.

The abstract makes the same claim. The decoder wants equal magnitudes, and "equal magnitudes and an exact prior force a separate frame for every block of every image" (abstract). Thm. 4(a) gives frame widening a channel-dependent sign.

The experiment also shows that ESM's components transfer to other constructions. Its per-block, per-image frames make SW exactly Gaussian under (A1), with recovery within sampling error (above). Its modulation, carrying PRC's own codeword, lifts PRC from 21/32 to 31/32 at JPEG50 and from 15/32 to 28/32 at JPEG30 (R4).

The ESM–SW gap (+64.8 points [+47.7, +81.2] against SW per-block at JPEG50, both backbones) therefore lies in the code and readout:

  • not in the frame pattern, whose paired effect is within sampling error (above);
  • a soft decoder gives SW 37 (SD1.5) and 40 (SD2.1) messages, against 35 native (Tables 27, 32), and SW per-block 39 and 33 (the R7a run; its table lists native-decoder counts), so soft decoding alone does not close the gap;
  • the remainder: SW's padding-mixed binary code with hard majority. Each padding bit is XORed into 31 message copies on average, so one padding error corrupts many copies at once. On SW's own channel samples, ESM's readout is predicted to recover 86.7% / 91.8% (Table 5, descriptive), against SW's observed 27.3%;
  • and ESM's equal-magnitude code is admissible only with per-block, per-image frames (Prop. 1(iii)). The frames enable; the equal-magnitude soft readout delivers.

Q3. What mechanism explains the gains from independent rotations; can they be attributed to reduced error correlation rather than padding or decoding?

Two mechanisms, both now measured.

(1) Equal magnitudes. With frames held fixed, equalizing the magnitudes adds +24.2 points [10.9, 38.3] of JPEG50 exact recovery, and every key group is positive (Table 2). Theory:

  • Thm. 4(b): equal magnitudes maximize the coherent soft sum at unchanged noise.
  • Thm. 4(c): among spread profiles (no coordinate carries more than ¼ of the block energy), they minimize the expected number of sign errors on acute blocks. At the median JPEG50 angle this is 0.18 vs 0.26 per coordinate.

(2) Reduced error correlation, which comes from where the copies sit rather than from the rotation itself (R7b).

  • The variance of the number of erroneous copies per bit, relative to the independent (Poisson-binomial) prediction from per-block error rates, is 1.007 for ESM on both backbones (JPEG50, CIs containing 1). It is 1.204–1.223 for HN, 1.137–1.138 for SW and 1.027–1.033 for GS.
  • ESM's copies of a bit sit in 32 distinct blocks. HN's copies sit in one latent column and SW's in one column block, consistent with SW's recovery staying within sampling error when only its frame pattern changes (R7a).
  • The paper's interleaver control isolates the same effect: a random interleaver accounts for 7–14 of HN's widening gain, leaving 1.5–1.9 points for rotation. ESM still beats interleaved HN by 24–30 points on SD1.5 and 19–25 on SD2.1 (App. C.3).

So the gain comes from equal magnitudes plus a layout whose repetitions are independent. Padding and hard decoding are additional losses specific to SW. The rotations are what make equal magnitudes exactly Gaussian under (A1).

We are happy to clarify anything before the discussion closes. If these results resolve your concerns about the verification setting and the source of the gains, we would appreciate it if you would reconsider your rating.

Thank you to the authors for their detailed response. I will read it carefully and provide feedback before the end of the weekend.

Thank you for the timely and detailed response, including the additional experimental results. My main concerns have been addressed, and I have therefore decided to maintain my positive rating.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General response (1/3): outer codes and nonce verification

General response (1/3): outer codes and nonce verification

We thank the reviewers. We are encouraged that they found the Gram-invariant theory a substantive contribution "beyond moment matching or empirical classifier tests" (a7xt), the channel transplant a creative test of it (z3YD), and the primary comparison carefully controlled (a7xt).

The shared requests were outer codes, nonce handling, PRC and other baselines, more distortions, and presentation. Beyond answering them, the new results show that ESM's components transfer to other constructions. Its per-block, per-image frames make SW exactly Gaussian under (A1), with SW's recovery staying within sampling error (R7a), and its modulation lifts PRC's own code from 21/32 to 31/32 at JPEG50 (R4). We put the main new results here once and point to them from the individual threads. Except the SW author-code check (R8) and PRC's native-setting arm (R4), all runs use the paper's pipeline and paired key × prompt design. Counts are complete 512/512-bit recoveries, and intervals are 95% crossed (key × prompt) bootstrap unless stated otherwise.

R1. A standard error-correcting code embedded in the payload, both backbones (z3YD-Q1, a7xt-W1). RS(128,64) over GF(2⁸), ×16 repetition, 512 net bits, hard-decision RS decoder, n=128 [Rev: App. J.3]:

backbone, channelESM+RSrot.HN+RSHN+RSESM − HNESM − rot.HN
SD1.5 JPEG5011910896+18.0 [+6.2, +32.0]+8.6 [+2.3, +16.4]
SD1.5 JPEG301168773+33.6 [+16.4, +52.3]+22.7 [+9.4, +38.3]
SD2.1 JPEG50124109107+13.3 [+1.6, +28.9]+11.7 [+3.9, +21.9]
SD2.1 JPEG301189070+37.5 [+18.0, +57.0]+21.9 [+9.4, +35.9]
  • Consistency with App. C.2. We read the embedded layout off the main run's channel samples, a different execution. The resulting predictions have intervals that contain every observed ESM and HN JPEG count on both backbones. For example, ESM on SD1.5 JPEG50 is predicted at 97.7% [92.2, 100] and observed at 93.0%. As an implementation check, the real decoder met its bounded-distance criterion on all 1,920 SD1.5 image × arm × channel cells, and the real errors-and-erasures decoder never selected a miscorrection.
  • Soft-assisted decoding (real errors-and-erasures RS, no oracle; SD1.5 embedded run): 122 / 109 / 102 at JPEG50, i.e. ESM − HN +15.6 [+3.9, +30.5].
  • Payload at a reliability target. Using idealized soft-assisted RS on the main run (n=128 per cell), the largest net payload with ≥ 99% JPEG50 exact recovery is:
ESMHNGS
SD1.5448128128
SD2.132016256

GS gains most from soft information. With soft-assisted RS at t=8t=8, ESM − GS is +4.7 [−0.8, 11.7] (SD1.5) and +3.9 [−1.6, 12.5] (SD2.1).

  • Why codes keep the gap. ESM's +0.63 / +0.65 BA points mean that HN's JPEG50 bit-error rate of 0.81% / 0.77% falls to 0.18% / 0.13% with ESM.

R2. Nonce verification with an exactly known false-positive rate (ugTx-W1/Q1, a7xt-W1/Q2, z3YD-W1) [Rev: App. J.1].

  • New Prop. J.1. Under ideal frames (A1), a wrong nonce or an unwatermarked image yields decoding frames that are independent of the image. Every decoded bit is then a fair coin, and agreement with any candidate message is exactly Bin(512,1/2)\mathrm{Bin}(512, 1/2). For the implemented ChaCha20 frames this holds up to the PRF advantage and the numerics of the normal sampler.
  • Empirical check. 640,000 wrong-nonce decodes per backbone match the ideal law: ESM's empirical FPR at the 283-agreement threshold is 0.0092–0.0100 (exact tail 0.0095), and χ2\chi^2 goodness-of-fit gives p≥0.135p\ge 0.135 in every channel.
  • Thresholds. A verifier holding NN issued nonces flags an image when agreement ≥ τ, where N⋅P(Bin≥τ)≤αN\cdot P(\mathrm{Bin}\ge\tau)\le\alpha. At α=10−3\alpha=10^{-3}, τ = 292 / 311 / 324 for N=1N = 1 / 10310^3 / 10610^6.
  • Detection power. ESM's lowest correct-nonce agreement over clean, JPEG50, JPEG30 and resize on both backbones is 373/512. So TPR is 128/128 in every such cell for N≤106N\le10^6 and α=10−2\alpha=10^{-2} or 10−310^{-3} (95% Clopper–Pearson lower bound 97.2%).
    • HN and GS also reach 128/128. Detection at a fixed FPR is saturated on these channels, and complete-message recovery is what separates the methods.
    • Under the tested crop, TPR is at most 6.2% at N=1N=1 and 0 for N≥103N\ge10^3 for every method.
  • Cost. A candidate decode costs about 50 ms (Table 25). A partial decode that regenerates 16 of the 128 frames is about 6–7× cheaper than a full decode in the same timing harness; it has the same exact null and a TPR ≥ 98.4% at N=106N=10^6, α=10−3\alpha=10^{-3} under JPEG50/JPEG30. Large logs therefore call for per-user or per-time-window candidate sets, or for transporting the nonce with the image. What is exact is the threshold, not the cost.
AuthorsOpenReview ↗

General response (2/3): finite key, PRC, more distortions, crop control, SW author code

General response (2/3): finite key, PRC, more distortions, crop control, SW author code

R3. Finite-key undetectability (a7xt-Q1, z3YD-W1) [Rev: App. J.2, new Prop. J.2]. Assume:

  • HMAC-SHA256 and ChaCha20, as used to derive the frames, are PRFs;
  • the key is secret;
  • the Gaussian amplitude draw is fresh for each image;
  • nonces are unique.

Then ESM latents for any number of images, including adaptively chosen messages and the images generated from them, are computationally indistinguishable from latents built with a truly random function. In real arithmetic its frames are ideal (A1), and Prop. 1(i) makes those latents exactly N(0,I)\mathcal N(0,I) for every message. We do not bound the floating-point difference between the implemented and the exact normal sampler.

R4. PRC (Gunn et al., ICLR 2025; official code, 512-bit payload, our pipeline, n=128 per backbone) [Rev: App. J.4]. Exact recoveries, with PRC decoder erasures in parentheses:

cleanJPEG50JPEG30resizecrop
SD1.5 PRC126 (2)106 (22)86 (42)122 (6)0
SD1.5 ESM128114971250
SD2.1 PRC128 (0)107 (21)85 (43)127 (1)0
SD2.1 ESM128110981280
  • ESM's count is ≥ PRC's in every cell of these five channels. The JPEG intervals include 0, so we claim parity or better:
    • SD1.5: +6.2 [−6.2, +19.5] (JPEG50) and +8.6 [−4.7, +23.4] (JPEG30);
    • SD2.1: +2.3 [−10.9, +14.8] and +10.2 [−3.9, +25.0].
  • The failure modes differ. PRC never returned a wrong message: it abstains. ESM always returns a message with high bit accuracy, which an outer code can use.
  • In PRC's native setting (SD2.1, 32 pairs), both methods recover 32/32 on all non-crop channels.
  • On the 13 distortions of R5 (SD2.1, n=32), PRC trails or ties ESM in 12 and is one image ahead on median 3 (32 vs 31). All ESM − PRC intervals include 0.
  • Complementarity. PRC's own codeword carried by ESM's modulation (SD2.1, 32 requests) rises from 21/32 to 31/32 at JPEG50 (+31.2 [+3.0, +65.6]) and from 15/32 to 28/32 at JPEG30 (+40.6 [+6.2, +78.1]).

R5. Thirteen new distortions (z3YD-W2, a7xt-W2) [Rev: App. J.5]. Saved images were re-inverted, n=32 per cell. Exact recoveries, ESM / HN / GS / SW:

distortionSD1.5SD2.1
JPEG q7032/24/18/1829/22/16/12
JPEG q1014/4/4/010/1/1/0
noise σ=0.0231/23/18/1931/24/19/12
noise σ=0.0525/13/13/1322/15/13/8
blur r=132/32/32/3132/30/30/26
blur r=232/24/19/1232/24/17/10
median 331/27/24/2331/27/22/19
median 528/24/18/1726/20/15/12
bright ×1.332/32/32/2832/29/28/23
bright ×0.732/31/31/3132/31/28/27
contrast ×1.332/31/32/2832/28/24/27
rotate 2°0/0/0/00/0/0/0
rotate 5°0/0/0/00/0/0/0
  • Among these four methods, no method exceeds ESM's count in any of the 26 cells.
  • With n=32 many intervals are wide, so we read the table as consistent support, not as 26 significance claims. Intervals excluding 0 include:
    • blur r=2 vs HN on both backbones (+25.0 [+3.1, +50.0] and +25.0 [+6.2, +46.9]);
    • noise σ=0.05 vs HN on SD1.5 (+37.5 [+9.4, +68.8]);
    • most compression, noise and filter cells vs GS and SW.
  • Rotation defeats every method tested, PRC included (PRC run on SD2.1); ESM keeps the highest bit accuracy (68–72% at 2°).

R6. Crop control (a7xt-Q4) [Rev: App. J.6]. The paper's 410² crop was pasted back at its original offset, without rescaling, with a mean-grey fill (n=128):

ESMHNGS (soft dec.)SW
SD1.5123010 (90)0
SD2.1125013 (81)0

An 8-pixel-grid-snapped 416² box gives 124 / 125 for ESM, again with HN and SW at 0. About 36% of the image area is lost, and ESM still recovers 96–98% of complete messages. So for ESM the crop as tested fails through rescaling and misalignment, not through lost area.

R8. SW author code (a7xt-Q3) [Rev: App. J.8]. We checked our SW adapter against the SW authors' own code in its shipped configuration: 512-bit message, N=31, fp32 SD1.5, sampled VAE posterior, 2 signatures × 16 captions. It recovers 32/32 clean, 20/32 at JPEG70 and 15/32 at JPEG50; our adapter at the same numerics recovers 109/128, 47/128 and 32/128 (Table 15). The author code is therefore the stronger SW implementation, and the revised PDF will adopt it as SW's reference in the §5.4 / Table 3 SW entries for the settings it covers (SD1.5; clean, JPEG70, JPEG50); the paired analyses (Table 5, R5–R7) keep the adapter, which shares ESM's requests. ESM's margin holds at the same fp32 / sampled-posterior numerics: 115/128 (89.8%) vs 15/32 (46.9%) at JPEG50, and 121/128 (94.5%) vs 20/32 (62.5%) at JPEG70 (different requests, so descriptive).

AuthorsOpenReview ↗

General response (3/3): mechanism and paper changes

General response (3/3): mechanism and paper changes

R7. What drives recovery (ugTx-W2/Q2/Q3) [Rev: App. J.7].

(a) SW with per-block, per-image frames. We kept everything of SW (padding, XOR mixing, 31 repetitions, its χ₁₂₈ column radii and native decoder) and changed only the frame pattern, from one key-fixed frame to an independent frame per block and image (n=128):

SWSW per-blockΔESM
SD1.5 JPEG503531−3.1 [−15.6, +8.6]114
SD1.5 JPEG301922+2.3 [−7.0, +14.1]97
SD2.1 JPEG503527−6.2 [−19.5, +6.2]110
SD2.1 JPEG302220−1.6 [−11.7, +7.0]98

This makes SW's latent exactly Gaussian under (A1): one frame per block, independent χ radii (Prop. 1(i)). Recovery stays within sampling error of shared-frame SW: all eight paired intervals, including clean and resize, contain 0, and the JPEG50 point differences are under a tenth of the ESM − SW per-block gap below. This fits the theory: the frame pattern governs the prior (Prop. 1), the code and readout govern recovery (Thms. 3–4), and ESM takes the only exact pattern. The remaining gap, ESM − SW per-block +64.8 points [+47.7, +81.2] at JPEG50 on both backbones, comes from code and readout. ESM's readout on SW's own channel samples is predicted to recover 86.7% / 91.8% (Table 5, descriptive) against SW's observed 27.3%, and ESM's equal-magnitude code is admissible only with per-block, per-image frames (Prop. 1(iii)). The frames enable; the equal-magnitude soft readout delivers.

(b) Dependence among the repetitions of a bit. For each image and bit we counted the erroneous copies. We compared their variance with the Poisson-binomial variance implied by that image's per-block error rates (one rate per copy and 128-bit group); a ratio of 1 means independent copies. JPEG50, n=128:

ESMHNGSSW
SD1.51.007 [0.987, 1.031]1.223 [1.155, 1.307]1.027 [1.005, 1.050]1.138 [1.079, 1.213]
SD2.11.007 [0.984, 1.029]1.204 [1.134, 1.298]1.033 [1.010, 1.056]1.137 [1.077, 1.222]

ESM's copies err independently, as Thm. 3 predicts. Under JPEG50 the parameter-free cap-law prediction matches the observed number of erroneous copies within 0.1% on both backbones.

The independence comes from the layout: ESM puts a bit's 32 copies in 32 distinct blocks, HN in one latent column, SW in one column block (kept fixed in (a)). The paper's interleaver control shows the same effect: a random interleaver accounts for 7–14 of HN's widening gain, leaving 1.5–1.9 points for rotation, and ESM still beats interleaved HN by 24–30 points (SD1.5) and 19–25 points (SD2.1) (App. C.3).

Paper changes. [Rev: X] marks where an item will appear in the revised PDF. We will upload it before the discussion closes, with changes in blue, and post a note here when it is up. Besides App. J.1–J.8 (R1–R8), the revised PDF will make these changes:

  1. Method first (ugTx-W3, z3YD-W3) [Rev: §3]. An "ESM at a glance" paragraph opens §3: a key and a public nonce generate 128 block rotations from Householder factors; each block gets equal-magnitude signs at a χ radius; each bit is repeated 32× across distinct blocks; the decoder sums the 32 copies. Algorithm 1 and Fig. 4(a) move from App. A to before Prop. 1.
  2. Abstract without undefined terms (z3YD-W3) [Rev: abstract]. "Gram invariant" becomes "inner products of blocks that share a secret rotation"; "ideal secret frames" becomes "idealized, independent secret rotations".
  3. Definitions at first use, notation table (ugTx-W3, z3YD-W3) [Rev: §3, Fig. 1, App. A]: frame, frame pattern, prior-preserving, (A1)/(A2), channel, Gram invariant, fiber, recovered angle, acute block, spread profile, cap law, readout. Fig. 1's caption defines the fiber.
  4. Bridge paragraph (ugTx-W3) [Rev: §3.3] linking (A1), Prop. J.2, the sampler tests (Table 18) and the angular law, which predicts the implemented key's recovery within 3.04 points in ten settings with no fitted parameter (Fig. 3).
  5. Related work and labels (z3YD-W1, a7xt-W2a/W2b/Q3) [Rev: §2, App. H, A.4, §5.4, Table 3]: PRC, T2SMark, WIND, SEAL; "component-level adaptation" (GS), "author code" (SW, its reference where R8 applies), "exploratory reconstruction" (LAW/LAW-M). The empty heading G.1 is removed.
  6. Transplant and classifier (a7xt-W2d/W2e) [Rev: §5.1, App. C.1, C.9]: "agreement within sampling error" with the largest |Δ| per cell; the interleaved control and the scope of Prop. J.2 in C.9.
  7. Scope (all) [Rev: §6]. Rotation and the rescaled crop defeat every method we tested; both are synchronization failures, orthogonal to the modulation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.