acceptodds
Under review as a conference paper at ICLR 2027

Multi-Bit Watermarking with Semantic Integrity for Binary-Token Visual Generative Models

Abstract

The widespread deployment of large-scale image generation models raises urgent questions about provenance, attribution, and content authenticity. We present a training-free multi-bit watermarking framework for binary-token visual autoregressive models. The key idea is to treat the model's latent binary variables as redundant watermark carriers: each payload bit is assigned to many token-bit positions, and generation is steered only through positions that can be modified at low semantic cost. We investigate two visual autoregressive models with different generation mechanisms: Infinity, where selected uncertain bits at the finest scale are forced after sampling, and BiGR, where Bernoulli marginals are softly biased during bidirectional masked prediction. Extraction is purely feed-forward: the image is re-encoded through the frozen tokenizer and the payload is recovered by majority voting over the carrier sets. The framework supports variable-length payloads (16, 32, and 64 bits) and, for Infinity, achieves near-perfect bit accuracy on COCO-prompted images with negligible degradation to image quality as measured by FID, KID, and CLIP Score. Building on the multi-bit capacity, we introduce a semantic integrity mechanism: the watermark payload encodes a DINOHash perceptual fingerprint of the generated image, enabling post-hoc detection of semantic tampering. Because the watermark resides in latent token space, it can persist through pixel-level edits, while the perceptual hash of a tampered image diverges from the recovered payload—revealing manipulation.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
watermarking, generative models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_m7AtOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a training-free multi-bit watermarking method for image generators that use binary tokens. Specifically, it modifies bits with high model uncertainty in Infinity and biases sampling probabilities in BiGR to embed 16–64 bits of information. The watermark is extracted through tokenizer re-encoding and majority voting. The authors further use the image’s DINOHash perceptual fingerprint as the watermark payload, following a process of generating an image, computing its fingerprint, and regenerating the image with the fingerprint embedded to detect semantic tampering. Experiments evaluate watermark recovery, image quality, and robustness on both models, and semantic tampering detection on Infinity. The method performs well on images without attacks, but its robustness to common image processing remains limited.

Critical strengths

It can embed and recover information in Infinity and BiGR without retraining the generator or training an additional watermark decoder. This shows that binary latent representations can serve as watermark carriers.

Infinity achieves 99.77% bit accuracy for 32-bit watermarks, and all watermark bits are correctly recovered in 95.18% of images. The image-quality metrics change little, supporting the basic effectiveness of the method.

Carrier-selection ablations support the uncertainty-guided strategy. The authors also evaluate transformed negative samples and compare different content within the same tampering geometry, helping explain the effects of detection-threshold shifts, watermark damage, and perceptual-hash changes.

Critical weaknesses

At the threshold used in the paper, JPEG compression at quality 75 causes a 48.1% false-alarm rate on untampered images, while color jitter causes a 12.6% false-alarm rate (Table 16). Meanwhile, without additional processing, the detection rate for object insertion covering 5% of the image area is only 52.8% (Table 4). These results show that fingerprint differences may arise from watermark recovery errors rather than actual semantic changes, substantially limiting the usefulness of this application.

At comparable image quality, TrustMark achieves better payload recovery and robustness than the proposed method (Tables 18–19). Center cropping also raises the false-positive rate of Infinity’s watermark detection to 33.7%. Although TrustMark relies on a pretrained watermarking network while the proposed method requires no training, the authors still need to explain which specific scenarios make the convenience of training-free integration worth the loss in robustness.

Questions for authors

Does semantic tampering detection assume that a valid watermark has already been confirmed? Algorithm 2 requires a known reference payload to determine whether a watermark is present, whereas Section 5.1 directly compares the extracted payload with the current image’s perceptual fingerprint. How are these two steps connected? If arbitrary images are accepted as inputs, please report the proportion of unwatermarked real photographs and generated images that are incorrectly classified as carrying a valid watermark. The existing known-payload detection results do not directly answer this question.

How much semantic tampering detection performance remains after controlling false alarms? Could the existing experimental data be used to select a single threshold that maintains a low false-alarm rate on images that have undergone only ordinary processing, without semantic tampering, and report the detection rates for the three tampering areas at that threshold? This result would directly affect my assessment of the method’s practical usefulness.

3 comments

Thank you for the review and for the two concrete questions. Since submission we have revised the Infinity instantiation and re-run every Infinity experiment; all Infinity numbers in the revised paper come from this re-run. The changes are:

  • Carrier layout. In the submission each payload bit occupied a single bit-plane (bit k in plane k mod 32, and at 64 bits only one half of the token grid). Each bit's carriers are now spread over all 32 bit-planes and the whole 64×64 token grid, and a secret key permutes the bit-planes within every token (App. A.2, App. C). Every Infinity run uses the key. The number of carriers per bit, the share of forced carriers and the uncertainty-guided selection are unchanged.
  • Integrity threshold. τ is chosen on 1,000 held-out images that are not used in any evaluation (τ = 6, with 0.9% false alarms there and a Clopper–Pearson 95% upper bound of 1.57%), instead of τ = 8, which was not chosen on held-out images.
  • Issuance needs one generation: the finest scale of the same generation is overwritten with the hash and decoded once more, so the image is not regenerated.
  • Evaluation. Attacked images are rounded to 8 bits before they are read, FID and KID use the standard FID Inception network (old and new FIDs are not comparable), TrustMark was re-run on the new images, and our side of the BitMark comparison was re-run with BitMark's released attack code and sampling temperature (BitMark's own numbers are taken from its paper).
  • PGD refinement is no longer used for Infinity, because every payload is read back correctly without it, and its appendix was removed.
  • Carrier-selection ablation (App. J). With the new layout every arm except the smallest ratio recovers every payload, so the arms are now compared by detection score and by per-image PSNR to the unwatermarked counterpart; uncertainty-guided selection keeps the highest detection score and changes the image least.
  • BiGR. PGD refinement, used for BiGR only, now acts on the saved 8-bit image; with it, BiGR's perfect-match rate is 100%, 99.84% and 98.86% at 16, 32 and 64 bits (Section 6.2, App. E).

Below we answer each point with the revised results; table and appendix numbers refer to the revised paper.

AuthorsOpenReview ↗

Answers to raised issues

(1) False alarms caused by watermark recovery errors This was correct for the submitted version: under JPEG the payload was misread by 6.69 bits on average, which produced most of the 48.1% false alarms. In the revision the payload error is at most 0.04 bits under every transformation (Table 12). We attribute this to where the carriers sit. In the submission all carriers of a payload bit lay in one bit-plane (and at 64 bits in one half of the image), so a transformation that disturbs that plane or region hit every carrier of the bit at once (64-bit bit accuracy of 82.3% under noise and 89.6% under JPEG). Spread over all 32 bit-planes and the whole token grid, with the same number of carriers per bit, each bit now loses only a little margin. At the held-out threshold τ = 6, false alarms on untampered images are 1.0% under JPEG, 1.3% under color jitter, 1.5% under noise and 0.6% under blur, and 0.62% without processing (Table 4). These rates are at τ = 6 and the submission's at τ = 8; since a lower threshold flags more images, the comparison understates the reduction. Detection of 5% insertions without processing is 59.3% (submission: 52.8% at a 2.34% clean false-alarm rate).

(2) TrustMark and center-crop false positives The 33.7% came from the submitted layout. Center crop now has an FPR@θ of 1.4%, and every transformation of Table 2 stays within 0.5–1.4%. We attribute the change to the carrier layout. In the submission all votes of a payload bit came from one bit-plane, so the white padding (36% of the image) cast the same value, the white token's value in that plane, on every padded carrier of the bit, pulling watermarked and unwatermarked images alike toward a fixed bit pattern; this moved unwatermarked scores past θ. With votes from all 32 bit-planes, the padding now pulls every bit toward the same average value instead. With the re-run, ours and TrustMark are on par on our suite (Tables 14–15): both keep at least 99.87% bit accuracy and 100% TPR@θ under every transformation, image quality matches, and 100% of our payloads are recovered against 99.93% of TrustMark's.

The two methods rest on different deployment assumptions (App. L): our watermark is written by the generator's own sampler, needs no training and no post-processing step, and its reader needs the carrier key, whereas TrustMark's decoding is public by design. We have not evaluated attacks on the key, so this is a design property rather than a demonstrated advantage. Our standard suite does not include crop-and-resize, flips or rotation; in the zero-bit comparison (Table 18) our scheme does not survive them. Regeneration attacks are not evaluated. We leave both to future work.

(3) How detection and integrity verification connect; unwatermarked inputs (Question 1) The two uses are separate. Algorithm 2's score S tests for a known payload; this is the provenance use, for example a known model or user ID. In the integrity protocol the payload is the image's own hash and is not known in advance, so the verifier runs extraction only (Algorithm 2 without a reference payload) to obtain Ĥ, recomputes H' = H(x') from the received image, and flags the image when HD(Ĥ, H') ≥ τ (Section 5.1).

As you note, the known-payload results of Table 1 do not answer this question, and the paper does not report how often unwatermarked images pass the integrity check. The closest measurement is in App. I: the payload read from one issued image, compared with the hash of a different image, differs in 29.45 bits on average, and τ = 6 flags 99.98% of such pairs. An unwatermarked image is in a similar position, since the bits read from it are unrelated to its hash, but this is an indirect argument. We have not measured the rate on unwatermarked real photographs or generated images and leave it to future work.

(4) One threshold for all ordinary processing (Question 2) Table 12 reports this for the single threshold τ = 6, which is chosen on 1,000 held-out images before any evaluation (App. I). Over no processing, noise, blur, color jitter and JPEG, the false-alarm rate on untampered images is at most 1.5% (noise), and detection is 59.9–60.7%, 84.1–84.6% and 91.7–92.9% at 5%, 10% and 15% area. These rates hold at the strengths listed in Table 9; the suite contains no rescaling or brightness change.

Grayscale conversion and our two crops are the exception: at τ = 6 they flag 41.9%, 46.5% and 51.5% of untampered images, because DINOHash itself moves by 5.1–5.7 bits while the payload is intact. Raising τ does not remove the problem: at τ = 8 they still flag 18.7–25.6% of untampered images, while detection at 5% area without processing falls to 40.9% (App. H). The limiting factor for these transformations is the invariance of the perceptual hash, not the robustness of the payload. We have not reported a full threshold sweep under processing and leave it to future work.

I appreciate the author's detailed response and supplementary experiments. The revised carrier design, independent threshold calibration, and baseline comparison significantly alleviated my main concerns. Although the applicability of integrity detection to some common transformations and watermark-free inputs remains limited, I am raising my rating to 3.

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

This paper proposes a training-free multi-bit watermarking framework for binary-token visual generative models. It embeds 16-, 32-, or 64-bit messages by modifying latent binary tokens during generation and recovers them through tokenizer re-encoding and majority voting. The method is implemented on Infinity and BiGR using uncertainty-guided bit forcing and probability biasing, respectively. The authors further propose a semantic integrity mechanism that embeds a DINOHash fingerprint to detect content manipulation. Experiments on COCO and ImageNet demonstrate high watermark detection accuracy with limited impact on image quality. The paper also evaluates robustness against common image transformations and semantic tampering, with additional comparisons to TrustMark and BitMark.

Critical strengths
  1. The method requires no model retraining and supports multiple payload lengths. Its implementation on two different binary-token generators demonstrates some flexibility across generation architectures. The Infinity results are particularly strong, achieving 99.77% bit accuracy for 32-bit payloads with little degradation in image quality.
  2. Embedding a perceptual fingerprint rather than an arbitrary message is an interesting application of multi-bit watermarking. The authors also provide controlled experiments to distinguish semantic changes from watermark damage.
Critical weaknesses
  1. Although semantic tampering detection is a central contribution, Appendix I reports substantial false-alarm rates under benign transformations, including 48.1% for JPEG compression and 61.3% for grayscale conversion. This significantly limits the reliability of the proposed integrity verification in realistic settings.
  2. The comparison with TrustMark shows that the proposed method has lower payload recovery accuracy and weaker robustness, particularly against cropping. The zero-bit comparison with BitMark also reveals large gaps under several transformations. While the training-free design is valuable, the paper does not sufficiently establish when its approach would be preferable to existing alternatives
Questions for authors
  1. Given the high false-alarm rates under common transformations in Appendix I, how do the authors envision deploying the semantic integrity mechanism reliably?
  2. The proposed method underperforms TrustMark in robustness and payload recovery. Beyond avoiding training, what practical advantages justify using generation-time watermarking instead of post-hoc watermarking?
  3. Semantic integrity is evaluated only on Infinity. Is there evidence that the proposed protocol can work reliably on BiGR, given the embedding-induced hash drift discussed in Appendix J?
2 comments

Thank you for the review. Since submission we have revised the Infinity instantiation and re-run every Infinity experiment; all Infinity numbers in the revised paper come from this re-run. The changes are:

  • Carrier layout. In the submission each payload bit occupied a single bit-plane (bit k in plane k mod 32, and at 64 bits only one half of the token grid). Each bit's carriers are now spread over all 32 bit-planes and the whole 64×64 token grid, and a secret key permutes the bit-planes within every token (App. A.2, App. C). Every Infinity run uses the key. The number of carriers per bit, the share of forced carriers and the uncertainty-guided selection are unchanged.
  • Integrity threshold. τ is chosen on 1,000 held-out images that are not used in any evaluation (τ = 6, with 0.9% false alarms there and a Clopper–Pearson 95% upper bound of 1.57%), instead of τ = 8, which was not chosen on held-out images. Issuance needs one generation: the finest scale of the same generation is overwritten with the hash and decoded once more, so the image is not regenerated.
  • Evaluation. Attacked images are rounded to 8 bits before they are read, FID and KID use the standard FID Inception network (old and new FIDs are not comparable), TrustMark was re-run on the new images, and our side of the BitMark comparison was re-run with BitMark's released attack code and sampling temperature (BitMark's own numbers are taken from its paper).
  • PGD refinement is no longer used for Infinity, because every payload is read back correctly without it, and its appendix was removed. Carrier-selection ablation (App. J). With the new layout every arm except the smallest ratio recovers every payload, so the arms are now compared by detection score and by per-image PSNR to the unwatermarked counterpart; uncertainty-guided selection keeps the highest detection score and changes the image least.
  • BiGR. PGD refinement, used for BiGR only, now acts on the saved 8-bit image; with it, BiGR's perfect-match rate is 100%, 99.84% and 98.86% at 16, 32 and 64 bits (Section 6.2, App. E).

Below we answer each point with the revised results; table and appendix numbers refer to the revised paper.

AuthorsOpenReview ↗

Answers to the raised issues

(1) False alarms of the integrity check, and deployment At the threshold τ = 6, chosen on 1,000 held-out images, false alarms on untampered images are 1.0% under JPEG at quality 75, 1.5% under noise, 0.6% under blur and 1.3% under color jitter (Table 12 in App. H; submission: 48.1%, 72.8%, 18.1% and 12.6%). These rates are at τ = 6 and the submission's at τ = 8; since a lower threshold flags more images, the comparison understates the reduction. In the submission these false alarms came from payload read errors; the payload error is now at most 0.04 bits under every transformation. We attribute this to where the carriers sit. In the submission all carriers of a payload bit lay in one bit-plane (and at 64 bits in one half of the image), so a transformation that disturbs that plane or region hit every carrier of the bit at once (64-bit bit accuracy of 82.3% under noise and 89.6% under JPEG). Spread over all 32 bit-planes and the whole token grid, with the same number of carriers per bit, each bit now loses only a little margin.

At these strengths (Table 9; the suite contains no rescaling or brightness change), the check therefore keeps false alarms at 0.6–1.5% with detection of 59.9–60.7%, 84.1–84.6% and 91.7–92.9% at 5%, 10% and 15% pasted area. Grayscale conversion (41.9%) and our two crops, which replace 36% of the image with white padding (46.5% and 51.5%), are not covered: there DINOHash itself moves by 5.1–5.7 bits while the payload is intact, and raising τ to 8 still flags 18.7–25.6% while detection at 5% area falls to 40.9% (App. H). A fingerprint more invariant to grayscale conversion and cropping is left to future work (Limitations and Future work).

(2) TrustMark and BitMark; practical advantages With the re-run, TrustMark no longer has higher payload recovery or robustness on our suite. Our payload is recovered on 100% of images against 99.93% for TrustMark (Table 14), and under all seven transformations, cropping included, both keep at least 99.87% bit accuracy, 100% TPR@θ and FPR@θ of at most 1.4% (Table 15). Image quality matches.

The two methods rest on different deployment assumptions (App. L). Beyond needing no training, our watermark is written by the generator's own sampler, so there is no separate network and no post-processing step, and reading its payload requires the carrier key, whereas TrustMark's decoding is public by design and it ships with a removal network. We have not evaluated attacks on the key, so this is a design property, not a demonstrated advantage.

BitMark is zero-bit. App. M compares against it, as the closest comparable prior work, in the zero-bit setting both methods support, using a minimal zero-bit adaptation of our scheme; the multi-bit payload our method is built around is not exercised there. Under BitMark's stronger attack suite the gaps under strong noise, strong color jitter, crop-and-resize and flips remain (Table 18). We have not evaluated our multi-bit watermark under crop-and-resize, flips or rotation, and leave this to future work.

(3) Semantic integrity on BiGR BiGR's generation structure makes the semantic-integrity pipeline challenging to adapt, so we omitted it and evaluate the protocol on Infinity only. App. I explains why: the protocol relies on the watermarked image staying close to the unwatermarked original, so that the embedding-induced hash drift stays small (1.37 bits for Infinity, whose embedding overwrites high-uncertainty bits once, at the final scale), whereas BiGR's bias is applied at every step of the iterative masked-prediction loop and would inflate that drift. We have not measured the drift on BiGR and leave a BiGR version to future work.

Reviewer_jC9dOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a training-free multi-bit watermark for binary-token image generators, tested on Infinity and BiGR. Each payload bit is spread over a large group of token-bit positions. For Infinity, the most uncertain sampled bits at the finest scale are overwritten with the payload bit. For BiGR, the Bernoulli probabilities are biased toward the payload bit at every refinement step. Extraction re-encodes the image with the frozen tokenizer and takes a majority vote per group. The paper also proposes semantic integrity: a 64-bit DINOHash of the image is embedded as the payload, and a mismatch between the extracted and recomputed hash flags tampering. Infinity reaches 99.77% bit accuracy at 32 bits with almost no change in FID or CLIP score. Tamper detection reaches 52.8% to 91.0% for pasted objects covering 5% to 15% of the image, at a 2.34% false alarm rate.

Critical strengths
  1. The evaluation is careful and honest. It uses large samples, attacked negatives, bootstrap intervals, a well designed equal geometry control (Appendix H), and a carrier selection ablation that supports the uncertainty argument. The authors also report comparisons they lose, against TrustMark and BitMark.
  2. The method is simple and training free. It decodes an explicit payload on bitwise generators, where prior work (BitMark, IndexMark) only does zero-bit detection.
Critical weaknesses
  1. No advantage over existing methods is shown. TrustMark, applied post hoc to the same images, matches the quality and is more robust under every attack (Table 19). BitMark is much stronger in the zero-bit setting under noise (99.6 vs 41.0), color jitter (99.8 vs 25.7), random crop (98.8 vs 2.6) and flips (Table 22). The paper does not make clear when its method should be preferred.
  2. Regeneration and purification attacks are missing. The SD2.1-VAE and CtrlRegen+ rows are left as placeholders, although this is the main threat to a watermark living in the tokenizer's latent bits.
  3. The security claim is untested. The main argument over TrustMark is that the watermark is keyed and built into the generator. However, all experiments use a fixed bit-plane assignment (k mod 32, Appendix A), not the pseudorandom key of Appendix C. No removal, forgery, copy or carrier estimation attack is evaluated. BiGR may be especially exposed, since it pushes every bit of a token group in the same direction.
  4. Robustness to geometric edits is weak. Random crop drops bit accuracy to 68.9%, and center crop pushes the false positive rate to 33.7% (Table 2).
  5. Semantic integrity fails under ordinary processing. Untampered images are flagged 48.1% of the time under JPEG at quality 75, 72.8% under noise, 61.3% under grayscale, and almost always under crop (Table 16). Detection of 5% edits is only 52.8%.
  6. The integrity evaluation is limited. Only pasted objects are tested, not inpainting, object removal or generative editing. There is no tamper detection baseline, such as EditGuard, SEAL, or a signed hash stored in metadata or a database. The scheme relies on the 64-bit payload, which decodes perfectly on only 60.4% of Infinity images, and no error correction is applied. It is evaluated on Infinity only and needs two generations per image.
  7. There are smaller issues. BiGR's FID moves from 23.34 to about 19.3 with watermarking, which signals a distribution shift that is only loosely explained. The ablation table reports FID around 64 against 31 in Table 3. This is probably due to the smaller reference set, but it should be stated.
Questions for authors
  1. How does the watermark perform under VAE re-encoding, diffusion purification and recent removal attacks?
  2. What are the results with the pseudorandom keyed assignment of Appendix C? Can an attacker estimate the carriers by averaging watermarked images, or copy a watermark onto another image?
  3. In what setting is this method preferable to TrustMark, given that TrustMark is more robust at the same quality?
  4. Would error correcting codes on the 64-bit payload bring the integrity false alarm rate under JPEG to a usable level?
  5. How does semantic integrity perform on realistic edits, and against a simple signed hash baseline?
  6. What explains the BiGR FID shift? Are paired comparisons available?
3 comments

Thank you for the detailed review and the concrete questions. Since submission we have revised the Infinity instantiation and re-run every Infinity experiment, so several numbers in your review have changed; all Infinity numbers in the revised paper come from this re-run. The changes are:

  • Carrier layout. In the submission each payload bit occupied a single bit-plane (bit k in plane k mod 32, and at 64 bits only one half of the token grid). Each bit's carriers are now spread over all 32 bit-planes and the whole 64×64 token grid, and a secret key permutes the bit-planes within every token (App. A.2, App. C). Every Infinity run uses the key. The number of carriers per bit, the share of forced carriers and the uncertainty-guided selection are unchanged.
  • Integrity threshold. τ is chosen on 1,000 held-out images that are not used in any evaluation (τ = 6, with 0.9% false alarms there and a Clopper–Pearson 95% upper bound of 1.57%), instead of τ = 8, which was not chosen on held-out images. Issuance needs one generation: the finest scale of the same generation is overwritten with the hash and decoded once more, so the image is not regenerated.
  • Evaluation. Attacked images are rounded to 8 bits before they are read, FID and KID use the standard FID Inception network (old and new FIDs are not comparable), TrustMark was re-run on the new images, and our side of the BitMark comparison was re-run with BitMark's released attack code and sampling temperature (BitMark's own numbers are taken from its paper).
  • PGD refinement is no longer used for Infinity, because every payload is read back correctly without it, and its appendix was removed.
  • Carrier-selection ablation (App. J). With the new layout every arm except the smallest ratio recovers every payload, so the arms are now compared by detection score and by per-image PSNR to the unwatermarked counterpart; uncertainty-guided selection keeps the highest detection score and changes the image least.
  • BiGR. PGD refinement, used for BiGR only, now acts on the saved 8-bit image; with it, BiGR's perfect-match rate is 100%, 99.84% and 98.86% at 16, 32 and 64 bits (Section 6.2, App. E).

Below we answer each point with the revised results and state plainly what we have not done; table and appendix numbers refer to the revised paper.

AuthorsOpenReview ↗

Answers to the raised issues and questions [1/2]

(1) When our method should be preferred; TrustMark and BitMark With the re-run, TrustMark is no longer more robust on our suite. Under all seven transformations both methods keep at least 99.87% bit accuracy, 100% TPR@θ and FPR@θ of at most 1.4% (Table 15), image quality matches (Table 14), and our payload is recovered on 100% of images against 99.93% for TrustMark. On this suite neither method dominates.

The two methods rest on different deployment assumptions (App. L). Our watermark is written by the generator's own sampler, needs no training and no post-processing step, and reading the payload requires the carrier key. TrustMark is a post-processing step whose decoding is public by design, and it ships with a watermark-removal network. We have not evaluated attacks on the key, so this is a design property, not a demonstrated security advantage (point 3). TrustMark is also trained against crop-and-resize (App. L), which our standard suite does not include; we have not evaluated our multi-bit watermark under crop-and-resize, flips, rotation or regeneration, and leave this to future work.

BitMark is a zero-bit scheme: it carries no payload and only tests whether an image was produced by a watermarked model. To the best of our knowledge no multi-bit scheme operating on Infinity's binary tokens exists, so App. M compares against BitMark, the closest comparable prior work, in the zero-bit setting that both methods support, using a minimal zero-bit adaptation of our scheme; the multi-bit payload that the method and the paper are built around is not exercised there. Under BitMark's own, stronger attack suite, BitMark is clearly stronger (Table 18; our rows were re-run with BitMark's released attack code and sampling temperature, and our noise and color rows are lower than in the submission). With the like-for-like real-photo null: strong noise 22.0% vs 99.6%, strong color jitter 16.9% vs 99.8%, crop-and-resize 1.3% vs 98.8%, and flips 0.5% and 0.2% vs 100% and 78.8%; both degrade under rotation (3.8% vs 20.1%).

(2) Regeneration and purification attacks We did not evaluate these attacks. At the time of writing we considered them too expensive to run, since they require auxiliary diffusion models, so the SD2.1-VAE and CtrlRegen+ rows remain deferred and we leave these attacks to future work. The Limitations and Future-work paragraphs state this.

(3) Security claim and the keyed assignment Every Infinity experiment in the revision uses the keyed assignment: a key-derived permutation of the 32 bit-planes for each token (App. A.2, App. C). The key hides which bit-plane of a token carries which payload bit; the counts, and at 64 bits the checkerboard split of the tokens between the two halves of the payload, are public (App. C).

We have not evaluated removal, forgery, copy, averaging or carrier-estimation attacks, and leave them to future work. BiGR uses no key and its assignment is public (App. C), so we make no security claim for the BiGR instantiation.

(4) Geometric edits The cited 68.9% and 33.7% come from the submitted layout. In the revised Table 2, random crop keeps 100% bit accuracy and center crop has an FPR@θ of 1.4%. These crops keep the remaining content in place and pad with white (Table 9). We attribute the change to the carrier layout. In the submission all votes of a payload bit came from one bit-plane, so the white padding (36% of the image) cast the same value, the white token's value in that plane, on every padded carrier of the bit, pulling watermarked and unwatermarked images alike toward a fixed bit pattern; this cost watermarked images whole bits and moved unwatermarked scores past θ. With votes from all 32 bit-planes, the padding now pulls every bit toward the same average value instead. Our multi-bit evaluation does not include crop-and-resize, flips or rotation; in the zero-bit comparison our scheme does not survive them (Table 18). We have not evaluated them for the multi-bit watermark and leave this to future work.

AuthorsOpenReview ↗

Answers to the raised issues and questions [2/2]

(5) Semantic integrity under ordinary processing; error-correcting codes At the held-out threshold τ = 6, false alarms on untampered images are now 1.5% under noise, 0.6% under blur, 1.3% under color jitter and 1.0% under JPEG at quality 75, against 72.8%, 18.1%, 12.6% and 48.1% in the submission (Table 12). These rates are at τ = 6 and the submission's at τ = 8; since a lower threshold flags more images, the comparison understates the reduction. In the submission those false alarms came from payload read errors, for example 6.69 wrong bits on average under JPEG. The payload error is now at most 0.04 bits under every transformation, without error-correcting codes, so the limiting factor is the hash rather than the payload (App. H). We attribute this to where the carriers sit. In the submission all carriers of a payload bit lay in one bit-plane (and at 64 bits in one half of the image), so a transformation that disturbs that plane or region hit every carrier of the bit at once (64-bit bit accuracy of 82.3% under noise and 89.6% under JPEG). Spread over all 32 bit-planes and the whole token grid, with the same number of carriers per bit, each bit now loses only a little margin. These rates hold at the strengths listed in Table 9; the suite contains no rescaling or brightness change.

Grayscale conversion (41.9%) and our two crops (46.5% and 51.5%) still flag many untampered images. There DINOHash itself moves by 5.1–5.7 bits while the payload is intact. Raising τ to 8 still leaves 18.7–25.6% and lowers detection at 5% area to 40.9% (App. H).

Detection of 5% insertions is now 59.3% at a 0.62% clean false-alarm rate (submission: 52.8% at 2.34%). The tamper signal averages 7.57 bits at 5% area against 1.37 bits on clean images (Table 4), which leaves under five bits of headroom below τ (App. I).

(6) Scope of the integrity evaluation

  • Payload: every embedded fingerprint is read back without error from all 5,000 evaluation images (App. I), and the 64-bit payload is recovered on all 10,000 images of Table 1.
  • Cost: issuance is one generation. The finest scale of the same generation is overwritten and decoded once more, which adds 0.84 s to the 2.85 s of generation per image (App. I).
  • Edit types: we test pasted objects only. At the time of writing we considered inpainting, object removal and generative editing too expensive to run at a scale large enough for meaningful results, and we leave them to future work. The equal-geometry control (App. G, Table 11) gives a partial view: at identical paste geometry, a displaced copy of the image's own content is detected in 31.9%, 59.2% and 74.5% of images at 5%, 10% and 15% area, and a blurred region in 5.5%, 13.8% and 22.1%.
  • Baselines: we have not compared against EditGuard, SEAL or a signed hash. We agree that EditGuard and SEAL would be worthy comparisons; we have no such results at present and leave them to future work.
  • Model: BiGR's generation structure makes the integrity pipeline challenging to adapt, so we omitted it (App. I explains why).

(7) BiGR FID shift and ablation FID With the standard FID network, BiGR's FID moves from 22.33 to 17.8–18.1 while KID stays within its subset noise (Table 3). The watermarked and unwatermarked BiGR images share class labels and seeds, so the CLIP Score drop of 0.17–0.21 in Table 3 is the mean per-image paired difference; App. K shows matched pairs, which often differ in composition. We have not done a further paired analysis and leave it to future work. App. B.6 discusses the FID change as a distribution shift caused by the bias changing which tokens are committed early, and presents this as a hypothesis. The ablation FIDs are computed on 1,000 images, and Table 13 now states that they are not comparable with the 10,000-image FIDs of Table 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.