acceptodds
Under review as a conference paper at ICLR 2027

ByteMark: Multi-Bit Watermark for Bit-wise Image Generators

Abstract

Watermarking is a central tool for provenance tracing of AI-generated content, yet existing methods for bit-wise image generators are zero-bit: they can only distinguish between watermarked or clean (non-watermarked) data. Many real-world deployments demand more, such as tracing a generated image back to the specific model, user, or API key that produced it. We introduce ByteMark, the first multi-bit watermarking scheme for bit-wise image generators. Since bit-wise generators require an equal number of zeros and ones to achieve high fidelity, ByteMark encodes each message bit in the transitions within tokens, by inducing the next generated bit to be either repeated or flipped based on the value of the message bit. Because the rule is relative to the preceding bit rather than to a fixed value, it embeds the message without disrupting the bit distribution. We further extend this mechanism to payloads with more than a thousand bits by embedding different parts of a message into many partitions of a given image. We show that ByteMark preserves the visual fidelity of generated images, is highly robust to a wide range of watermark removal attacks, and that it exhibits radioactivity, i.e., new models trained on ByteMark-ed images inherit and produce the watermark. These properties enable ByteMark to provide fine-grained provenance tracing in bit-wise image generation.

Primary area:
generative models
Keywords:
watermark, multi-bit watermark, image generation, bit-wise generation, provenance tracing, autoregressive generation

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_zZPTOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces ByteMark, a multi-bit watermarking method for bit-wise image generators that encodes messages through repetition or flipping relationships between adjacent bits. It extends the payload through spatial partitioning and evaluates image quality, robustness to watermark removal attacks, and the inheritance of watermark messages through downstream fine-tuning.

Critical strengths

Encoding messages through relationships between bits rather than directly biasing individual bit values is an intuitive design choice. The experiments cover multiple bit-wise generation models, payload sizes, and attack types. The investigation of multi-bit watermark inheritance through downstream fine-tuning is relevant to tracing the provenance of training data.

Critical weaknesses
  1. The claim of being the first to study multi-bit watermarking is overstated. Although the claim is restricted to bit-wise image generators, prior work on image generation models has already explored both detecting the presence of a watermark and recovering the specific watermark message.
  2. The relatively novel aspect of this work is the shift from directly encoding watermark bits to encoding relationships between bits, but the methodological novelty remains limited. Moreover, the design is restricted to bit-wise image generators, limiting its practical applicability and further constraining its utility.
  3. The method closely resembles BitMark in that both introduce a statistical bias during generation in bit-wise image models, with the distinction largely lying in how this bias is implemented. Could BitMark also be extended into a multi-bit watermarking method by adding a decoding module?
  4. For identity recovery, reporting bit accuracy alone is insufficient to establish successful identification. For example, Table 2 reports a bit accuracy of 95.4% when 50% of the training data is watermarked, but this does not imply that 95.4% of user identifiers are correctly recovered.
  5. The robustness results in Table 1 show low bit recovery accuracy under geometric transformations, raising concerns about the method’s usability for tasks that require complete recovery of the watermark message. These experiments do not adequately support the claimed multi-bit message recovery capability in such scenarios.
  6. For the multi-source training scenario, the paper only evaluates fine-tuning on images carrying the same fixed 31-bit message, which is not a reasonable representation of the intended setting. If the training images carry different watermark messages, oracle detection could become less effective, making blind detection more relevant; however, the current blind detection results still leave substantial room for improvement.
Questions for authors

Please see Weaknesses!

4 comments
AuthorsOpenReview ↗

We limit our claim to bit-wise generators, show why a decoder cannot make BitMark multi-bit, and add attribution among 1M users (≤0.1% misattribution), SyncSeal correction for geometric attacks (bit acc. 0.73→0.95) and multi-user radioactivity.

We thank the Reviewer for the detailed and constructive review. We are glad that the Reviewer finds the transition-based design intuitive and the multi-bit radioactivity relevant for the provenance of training data. We provide detailed answers below.

W1: The claim of being the first to study multi-bit watermarking is overstated.

Our claim is limited to in-generation multi-bit watermarks for bit-wise autoregressive image generators, and it is correct in this scope. We agree that multi-bit watermarks exist for other settings, and we cite and compare against them. Post-hoc schemes such as TrustMark, InvisMark and WAM are our baselines in Table 1. StableSignature is specific to latent diffusion models. We additionally cite other multi-bit work such as UniMark.

W2: The methodological novelty remains limited and the design is restricted to bit-wise image generators, limiting its practical applicability and further constraining its utility.

While our ByteMark is designed for bit-wise image generators, the insights are not limited to this scope. All generation-time watermarks are specific to their model class, such as TreeRing, which is applied to diffusion models and WMAR which is applied to IARs. The model specific design is what drives their robustness and our ByteMark now closes the multi-bit gap for bit-wise generators. The design principles we find also apply beyond bit-wise generators. (1) Multi-bit information can be applied in relations or transitions, rather than in values directly. (2) Image partitions for message embedding should be contiguous and the expected transforms should map each region on itself. (3) The design-effect correction can lower the worst-natural image FPR, as we show in Appendix D.3, which can be applied to other token based schemes.

W3: Could BitMark also be extended into a multi-bit watermarking method by adding a decoding module?

There is no direct way to turn BitMark into a multi-bit method with a decoding module. A decoder can only reconstruct and detect information that was embedded during the generation. Since BitMark applies one fixed green list at every position of every token all watermarks carry the same zero-bit information. To carry a message, the embedding must change, not the decoder. The direct change, a bias of each bit towards the message bit, fails as we explore in Appendix C.2.

W4: Reporting bit accuracy alone is insufficient to establish successful identification.

We agree that bit accuracy is not an identification rate. We therefore added end-to-end user attribution experiments in Appendix F 4. We generate one million random 31-bit user ids. For 10,000 of them, we generate one image with the id and a unique prompt. An image must first pass blind detection (Section 3.2, Equation 3). We then attribute it to the user with the nearest id in Hamming distance, if this distance is at most τ\tau. We select τ\tau as the largest value for which a uniform random message is within distance τ\tau of one of the NN ids with a probability of at most 1%.

AttackNoneBlurJPEGNoiseColorRotCropH-flipV-flipVAECtrlReg+
Attribution ↑, N = 1,000, τ = 399.999.999.996.694.07.923.519.86.0100.04.3
Attribution ↑, N = 10,000, τ = 2100.0100.099.993.691.35.812.912.53.2100.02.3
Attribution ↑, N = 100,000, τ = 1100.0100.099.989.988.84.24.55.01.3100.00.6
Attribution ↑, N = 1,000,000, τ = 0100.0100.099.780.183.42.20.91.30.3100.00.1
Misattribution ↓, N = 1,000, τ = 30.00.00.00.00.00.00.00.00.10.00.0
Misattribution ↓, N = 10,000, τ = 20.00.00.00.00.00.00.10.10.10.00.1
Misattribution ↓, N = 100,000, τ = 10.00.00.00.00.00.00.10.00.00.00.0
Misattribution ↓, N = 1,000,000, τ = 00.00.00.00.00.00.00.00.00.00.00.0

We report the attribution rate and misattribution rate of the end-to-end user attribution for up to one million users. We find that even for one million users we achieve near perfect attribution rate under no and simple attacks. Furthermore, the τ\tau gating results in a misattribution rate of at most 0.1% under all attacks.

We apply the same protocol to the images of the fine-tuned model M2 from Table 2. We report the attribution rate for one image (K = 1) and for 256 pooled images (K = 256) generated by M2M_2.

Watermarked share10%25%50%100%
K = 1, N = 1,0000.012.594.5100.0
K = 1, N = 1,000,0000.00.025.596.0
K = 256, N = 1,000100.0100.0100.0100.0
K = 256, N = 1,000,00093.5100.0100.0100.0

At 50% watermarked we achieve a correct attribution of 94.5% of the single images among 1,000 users. A provider has many images of a suspect model. With 256 pooled images, we identify the user among one million users in 93.5% of the draws, even if the user contributed only 10% of the training data. Misattribution is at most 0.5%.

W5: The robustness results in Table 1 show low bit recovery accuracy under geometric transformations. These experiments do not adequately support the claimed multi-bit message recovery capability in such scenarios.

We provide the attribution rates, also against the geometric attacks in our answer to your W4. Geometric transformations can be corrected using a post-hoc rotation estimation tool, as we show in Table 14. Additionally, ByteMark can be combined with SyncSeal [1], which further enables correction of rotations, crops and flips. We score each image both as received and after correction and keep the larger score, at a slightly stricter threshold that keeps the false-positive rate at 1%. We have added the description of this setup to Appendix F 1.2.

AttackBlind TPR 31 bitBit acc. 31 bitBit acc. 62 bitBit acc. 124 bit
Rotate 20°0.69 → 1.000.82 → 1.000.76 → 1.000.66 → 1.00
Rotate 30°0.25 → 1.000.72 → 1.000.65 → 1.000.58 → 1.00
Rotate 45°0.04 → 1.000.65 → 1.000.58 → 1.000.53 → 1.00
Rotate 90°0.03 → 0.990.61 → 0.960.56 → 0.920.50 → 0.84
Rotate 180°0.12 → 0.990.65 → 0.960.55 → 0.920.48 → 0.83
Rot+crop 10°0.92 → 1.000.90 → 0.970.84 → 0.940.74 → 0.86
Rot+crop 20°0.54 → 0.980.79 → 0.950.72 → 0.900.63 → 0.82
Rot+crop 30°0.21 → 0.970.71 → 0.930.64 → 0.880.56 → 0.79
Rot+crop 45°0.06 → 0.960.62 → 0.940.57 → 0.880.51 → 0.80
Crop 0.70.71 → 0.960.84 → 0.930.75 → 0.860.65 → 0.77
Crop 0.50.13 → 0.190.70 → 0.710.62 → 0.630.55 → 0.56
H-flip0.86 → 1.000.86 → 0.980.79 → 0.950.60 → 0.87
V-flip0.18 → 0.980.73 → 0.970.54 → 0.930.55 → 0.86
Rotate (suite, ±180°)0.16 → 1.000.65 → 1.000.60 → 1.000.55 → 0.99
Rot+crop (suite)0.16 → 0.950.65 → 0.940.59 → 0.890.54 → 0.81

We report the bit accuracy without \to with correction. Over the 15 geometric attacks, the mean bit accuracy increases from 0.73 to 0.95 for 31 bits, from 0.65 to 0.91 for 62 bits and from 0.58 to 0.85 for 124 bits. We have added these results to Table 18.

W6: Radioactivity training only on a single message

We performed two additional experiments: attribution of several users whose images are mixed in the training data.

We use a registry of 32 users using maximum simplex codes. We score each image of M2 with the oracle statistic (Eq. 4) for each registered user and attribute the image to the user with the highest score, if this score passes the calibrated 1% FPR threshold. This differs from the single-user table above, which uses blind detection and nearest-message decoding. We report attribution to a contributor / misattribution to a non-contributor, in % of single images of M2.

Users in training1% each5% each10% each25% each50% each100% each
0 (clean model)– / 1.0–––––
10.0 / 0.01.0 / 0.03.0 / 0.084.0 / 0.0100.0 / 0.0100.0 / 0.0
20.0 / 0.00.0 / 0.04.5 / 0.097.2 / 0.2100.0 / 0.0–
42.5 / 0.81.0 / 0.212.0 / 1.099.6 / 0.0––
81.4 / 0.01.8 / 0.516.4 / 0.1–––
161.4 / 0.42.3 / 0.1––––

Misattribution is at most 1.0% in all models, which is the false-positive level of the threshold. When each user contributes 25% of the training data, 84.0–99.6% of the single images of M2 are attributed to a contributor, with at most 0.2% misattribution. At lower shares, a single image does not carry sufficient signal.


References

[1] SyncSeal: “Geometric Image Synchronization with Deep Watermarking” Pierre Fernandez, Tomáš Souček, Nikola Jovanović, Hady Elsahar, Sylvestre-Alvise Rebuffi, Valeriu Lacatusu, Tuan Tran, Alexandre Mourachko. ICASSP 2026.

Reviewer_zZPTOpenReview ↗

Response to the authors

Ok, I have raised my score to 3. Good luck!

Dear Reviewer zZPT, Thank you for raising the score. If there are any remaining questions, we are happy to address them. Kind regards, The Authors

Reviewer_nU8kOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents ByteMark, the first multi-bit watermarking scheme for bit-wise image generators, extending existing zero-bit watermarking approaches. By encoding information through transitions between consecutive bits, ByteMark preserves the balance between zeros and ones, thereby maintaining high visual fidelity. The authors further extend this mechanism to support watermark payloads exceeding 1,000 bits and demonstrate an interesting property of watermark radioactivity, whereby the embedded watermark can persist in models trained on watermarked images. Extensive experiments also demonstrate strong robustness against a wide range of watermark removal attacks.

Critical strengths

Novelty: This paper addresses an important yet underexplored problem by extending existing zero-bit watermarking approaches to multi-bit in Bit-wise image generators and further demonstrating the property of watermark radioactivity.

Comprehensive experimental evaluation: It provides extensive experimental results across a wide range of settings, thoroughly evaluating the robustness, and radioactivity of ByteMark. The experiments offer strong empirical support for the proposed approach and demonstrate its practical applicability.

Simple yet well-motivated technical design: ByteMark is simple and intuitive, closely aligned with the structural characteristics of bit-wise generation. Its design is technically well-motivated rather than relying on complex heuristics.

Critical weaknesses

1.Limited Generalizability Beyond Bit-wise Image Generators. ByteMark is tightly coupled with the bit-wise generation process, making it inherently architecture-dependent. While Bit-wise Image Generators represent a promising research direction, and developing specialized watermarking methods for this paradigm is certainly valuable, the technical contributions of ByteMark appear to be largely confined to this specific setting. It would be valuable for the authors to further discuss whether the underlying insights of ByteMark could be generalized into broader watermarking design principles, and whether these insights could inspire watermarking approaches for other image generation architectures. Such a discussion would help clarify the broader implications of this work and strengthen its potential contribution to the field of image watermarking.


2.Validity of Security Assumptions and Potential Watermark Forgery Risks. Unlike post-hoc watermarking methods that employ independently trained watermark encoders or decoder, ByteMark relies on the visual encoder of the underlying image generation model for watermark detection and message extraction, and assumes that this encoder remains private. However, this design warrants further scrutiny: ByteMark treats a component primarily designed for visual representation learning, rather than security, as a critical security boundary for watermark detection. Given that many visual encoders and their pretrained weights about Bit-wise Image Generators are publicly available, an attacker could potentially exploit surrogate encoders with similar architectures or encoding mechanisms, or even obtain the exact encoder used by the target model, to attempt watermark forgery attacks. It would be valuable for the authors to further discuss how realistic this security assumption is in practical deployments and provide small-scale experiments evaluating the proposed method against such forgery attacks.


3.Insufficient security evaluation against watermark forgery attacks. The paper lacks a systematic evaluation of ByteMark's robustness against common watermark forgery attacks [1][2][3]. It would be beneficial to evaluate the low-cost forgery attack proposed in [1] and visualize ByteMark's watermark residuals following Figures 8 and 9 of [1] to investigate whether a fixed residual pattern exists.

Questions for authors
  • The paper's main contribution lies in its encoding scheme. Have the authors considered any potential drawbacks of the proposed encoding scheme or observed any limitations during empirical evaluations? For example, could the transition-based encoding amplify bit errors during decoding, among other potential issues?

  • ByteMark encodes messages in transition encoding rather than direct bit encoding to preserve a balanced bit distribution. Have the authors explored alternative approaches to achieving this goal, and how do they perform?

[1] Pei Yang, Hai Ci, Yiren Song, and Mike Zheng Shou. Can simple averaging defeat modern watermarks? NeurIPS, 2024.

[2] Ruowei Wang, Chenguo Lin, Qijun Zhao, and Feiyu Zhu. Watermark faker: Towards forgery of digital image watermarking. ICME, 2021.

[3] Ziping Dong, Chao Shuai, Zhongjie Ba, Peng Cheng, Zhan Qin, Qinglong Wang, and Kui Ren. WMCopier: Forging invisible watermarks on arbitrary images. NeurIPS, 2025.

5 comments
AuthorsOpenReview ↗

ByteMark works on four bit-wise generators and its transition design transfers beyond them. We test seven forgery attacks; content binding (SemBind) cuts white-box forgery 99.0%→2.7%, others to ≤1.3%. Extra bit errors fall mostly on message bits.

We would like to thank the Reviewer for the insightful comments. We appreciate the Reviewer recognizing that we address an important yet underexplored problem with a technically well motivated design. We answer each point in detail in the following.

W1: Limited Generalizability Beyond Bit-wise Image Generators.

All generation-time watermarks are specific to their model class, such as TreeRing, which is applied to diffusion models and WMAR which is applied to IARs. The model specific design is what provides their robustness and our ByteMark now closes the multi-bit gap for bit-wise generators. It applies without changes to four different models with different token layouts, namely Infinity-2B, Infinity-8B, BiGR and Instella (Table 3 and 5).

The design principles we find also apply beyond bit-wise generators. (1) Multi-bit information can be applied in relations or transitions, rather than in values directly. (2) Image partitions for message embedding should be contiguous and the expected transforms should map each region on itself. (3) The design-effect correction can lower the worst-natural image FPR, as we show in Appendix D.3, which can be applied to other token based schemes.

W2: Validity of Security Assumptions and Potential Watermark Forgery Risks

Our security assumptions are shared with prior work on token- and bit-based watermarking, like WMAR and BitMark. A commercial provider would keep their encoder private. We use the openly available models for reproducibility.

We additionally tested how much a surrogate encoder helps an attacker, by running LatentOpt [1], which optimizes an image until its latent under a proxy encoder matches the latent of a watermarked image. We forge the 31-bit message of one victim onto 100 MS-COCO images and report the percentage of images that pass blind detection and are attributed to the victim among 10000 users.

Proxy encoderAccessPasses detectionAttributed, no bindingAttributed, SemBind
Clean covers (no attack)–0.00.00.0
Infinity-2B VAE (exact)white100.099.02.7
Infinity-8B VAEgrey91.062.01.3
Infinity d=64 VAEgrey77.052.00.5
Infinity d=16 VAEgrey44.04.00.0
LlamaGen VQblack7.00.00.0

Without binding, the attack fails with an encoder from a different family (LlamaGen). Its success increases with knowledge of the architecture and is highest with the exact encoder.

To close this gap, we bind the watermark to the image content, following SemBind [2]. The provider masks the message of each user with a keyed semantic hash of the image. A watermark that is copied onto other content is unmasked with the hash of that content and no longer decodes to the victim. With this binding, white-box LatentOpt with the exact encoder falls from 99.0% to 2.7% at a cost of 1.5 points of mean genuine attribution. The binding uses no payload, which means that all 31 bits still identify the user.

W3: Insufficient security evaluation against watermark forgery attacks.

We additionally performed experiments analyzing the potential watermark forgery risks. Therefore we evaluate the proposed attacks of Averaging [3], Watermark Faker [4] and WMCopier [5]. We also include more attacks, namely Frequency injection [1], the high-pass variant of averaging [1] and WMForger [6].

AttackAccessPSNRLPIPSPasses detection, no DEFFPasses detection, DEFFAttributed (exact), no bindingAttributed (exact), SemBind
Clean covers (no attack)–––4.50.00.0 (0.0)0.0 (0.0)
Frequency injection [1]none36.20.0967.50.50.0 (0.0)0.0 (0.0)
Averaging (4,800 refs)images28.90.0332.538.518.5 (5.5)0.0 (0.0)
Averaging (1,800 pairs)pairs34.70.0316.021.58.0 (2.5)0.0 (0.0)
Averaging, high-pass (500 refs, ×4)images33.30.1681.588.069.0 (45.0)0.5 (0.0)
WMForger (1 ref)image33.20.1536.044.521.0 (6.0)0.1 (0.0)
Watermark Faker, pixel expansionpairs12.51.025.51.00.0 (0.0)0.0 (0.0)
Watermark Faker, RGBpairs13.70.6763.00.00.0 (0.0)0.0 (0.0)
WMCopier (5,000 refs)images28.80.404.00.00.0 (0.0)0.0 (0.0)

The fixed pattern that allows Frequency injection to pass the blind test on 67.5% of images, is caught by DEFF alone before binding is applied. Also attacks that learn the watermark fail. Watermark Faker and WMCopier reach at most 1% acceptance, i.e. our FPR threshold, and no image is attributed to the victim. Watermark Faker needs aligned pairs of the same image with and without the mark. These pairs do not exist in the same sense for a generation-time mark, because the mark changes the sampling process.

Attacks copying the watermark carry some success, before content binding. We find that averaging many images that carry the same message yields a consistent residual (correlation 0.89 between two disjoint sets of 1,000 images), and part of it is shared between all messages. Averaging the images of many different users attributes no image to any user, so only averaging hundreds of images known to come from the same user enables message forgery. With such images, plain averaging attributes 18.5% of the forged images to the victim, its high-pass variant 69.0%. WMForger, using only a single reference, achieves 21.0%. These attacks forge post-hoc watermarks as well [3, 6]. Content binding successfully defends all these attacks.

We additionally added the requested figures as Figure 13 and 14 to the paper to visualize the forged watermarks of [3].

Q1 : Have the authors considered any potential drawbacks of the proposed encoding scheme or observed any limitations during empirical evaluations? For example, could the transition-based encoding amplify bit errors during decoding, among other potential issues?

We have carefully evaluated the impact of our watermarking scheme on image quality and performance in Section 4.4 as well as Table 18-20. We performed additional experiments evaluating the bit-error rate during decoding with and without watermarking. Therefore, we store the bits produced during generation of 5000 images, decode them to an image and encode the image back to bits. The limitation that we see is capacity: longer payloads have fewer votes for each bit, and bit accuracy decreases (Table 1).

cleanByteMark
error rate, mean per image0.26810.3174
error rate, bits that follow the message0.26800.3301
error rate, bits that go against it0.26850.2709
share of bits following the message, as generated0.50020.8009
share of bits following the message, after re-encoding0.50030.5857

Our results show that ByteMark slightly increases the error rate, however the bit error focuses on the bits that embed the message, i.e. exactly the bits changed by our message. On the bits that don’t follow the message the error rate hardly changes.

Q2: Have the authors explored alternative approaches to achieving this goal, and how do they perform?

Alternative approaches to achieve a balanced distribution or mitigating the effects of an imbalanced one are in Appendix C.2. We bias the bit values directly and keep the 0/1 balance in other ways: a mirrored message (16 bits and their inverse), an interleaved message, a lower bias, a bias applied by chance, and a bias on later scales only. None restores the quality (Fig. 4). This still fails due to the codeword clustering, which occurs also for balanced messages. In this case the image converges towards a single token value, which further disrupts image quality and is not robust against attacks.


References

[1] Andreas Müller, Denis Lukovnikov, Shingo Kodama, Minh Pham, Anubhav Jain, Jonathan Petit, Niv Cohen, and Asja Fischer. On the robustness of watermarking for autoregressive image generation. NeurIPS, 2026.

[2] Xin Zhang, Zijin Yang, Kejiang Chen, Linfeng Ma, Weiming Zhang, and Nenghai Yu. SemBind: Binding diffusion watermarks to semantics against black-box forgery attacks. ICML, 2026. [3] Pei Yang, Hai Ci, Yiren Song, and Mike Zheng Shou. Can simple averaging defeat modern watermarks? NeurIPS, 2024. [4] Ruowei Wang, Chenguo Lin, Qijun Zhao, and Feiyu Zhu. Watermark faker: Towards forgery of digital image watermarking. ICME, 2021. [5] Ziping Dong, Chao Shuai, Zhongjie Ba, Peng Cheng, Zhan Qin, Qinglong Wang, and Kui Ren. WMCopier: Forging invisible watermarks on arbitrary images. NeurIPS, 2025. [6] Souček, T., Rebuffi, S.-A., Fernandez, P., Jovanović, N., Elsahar, H., Lacatusu, V., Tran, T., & Mourachko, A. (2025). Transferable Black-Box One-Shot Forging of Watermarks via Image Preference Models. NeurIPS, 2025.

I sincerely appreciate the authors' efforts in conducting such comprehensive additional experiments to address my concerns, and I am happy to raise my rating accordingly.

Reviewer_W7fAOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper embeds messages in bit-wise image generators by biasing adjacent bits to repeat or flip. It uses spatial partitions for longer messages and evaluates decoding, image quality, resistance to image transformations, and watermark inheritance after fine-tuning.

Critical strengths

The transition-based construction addresses the need for multi-bit messages in bit-wise generators. Evaluation across several generators provides evidence that the implementation is useful beyond one model, although I do not find this sufficient to outweigh the concerns below.

Critical weaknesses
  1. My main concern is the limited conceptual advance over existing watermarking schemes. Biasing a selected subset of generation outcomes is already central to Kirchenbauer et al. (2023), and message-dependent subset selection and payload allocation are developed in Yoo et al. (2024). More directly, Section 3.1 identifies the repeat/flip construction with the green/red lists used by BitMark. ByteMark adds message-dependent selection of these transitions and spatial payload partitioning. This is a concrete extension, but the paper does not sufficiently explain what new principle or capability makes it more than an incremental multi-bit adaptation. The failure of direct bit-value embedding is useful evidence, but does not by itself settle this comparison.

  2. The strongest application claim, fine-grained source attribution, is not adequately validated. Average bit accuracy and detection against a known message do not establish reliable identification of an unknown user. Several geometric attacks also reduce long-payload decoding close to chance. The radioactivity experiment uses one fixed message, leaving attribution under mixtures of different users' messages unresolved. Appendix D provides blind detection and calibration corrections, but these do not replace an end-to-end identification evaluation.

Overall, I recommend rejection because the conceptual advance is limited and the evidence does not sufficiently establish the claimed attribution capability.

References: Kirchenbauer et al. (2023), A Watermark for Large Language Models. https://proceedings.mlr.press/v202/kirchenbauer23a.html Yoo et al. (2024), Advancing Beyond Identification: Multi-bit Watermark for Large Language Models via Position Allocation. https://arxiv.org/abs/2308.00221 Kerner et al. (2025), BitMark: Watermarking Bitwise Autoregressive Image Generative Models. https://arxiv.org/abs/2506.21209

Questions for authors

What is the essential advance beyond message-dependent list selection applied to BitMark's transition alphabet? Do you have complete identifier-recovery and false-attribution results, including when downstream training mixes different payloads? Please distinguish existing results from suggestions for future experiments.

6 comments
AuthorsOpenReview ↗

We show that MPAC-style list selection is our Naive baseline and fails in bit-wise generators, and add attribution among 1M users (≤0.1% misattribution), SyncSeal correction (bit acc. 0.73→0.95) and radioactivity attribution for mixed users.

We would like to thank the Reviewer for the insightful comments. We appreciate the Reviewer recognizing that we address the need for multi-bit message embedding and the usefulness of our method.

W1 & Q1: What is the essential advance beyond message-dependent list selection applied to BitMark's transition alphabet?

The advance of our method is not the choice of a specific list, but rather to find the channel that survives in bit-wise generators. The direct transfer of Yoo et al. (2024) is exactly our Naive baseline. MPAC selects one list per message chunk from a partition of the vocabulary. In bit-wise generators, the vocabulary at each step is {0, 1}. The message-dependent list would then bias the single bit value (bjb_j to mjm_j). We show in Section 3.1 and Appendix C.2 that this approach fails in two ways, namely quality and robustness. Specifically this direct embedding shifts the balance of 0s and 1s in a generated image. Our method on the other hand keeps the delicate balance intact by biasing in the transitions between bit positions. We added the Yoo et al. (2024) reference to Appendix C.2.

W2a & Q2a: The strongest application claim, fine-grained source attribution, is not adequately validated. Do you have complete identifier-recovery and false-attribution results?

We performed additional experiments to evaluate the ability to trace an image back to its user. Therefore, we generate one million random 31-bit sequences where each of them represents a single user’s id. Of these ids, we randomly sample 10,000, for which we generate images that are marked with their respective id and a unique prompt. To evaluate the attribution, we leverage the blind detection procedure, where we first test if we can detect a watermark within an image (Section 3.2, Equation 3). After passing the detection, the message is extracted and the image attributed to the user with the nearest message. We further add a threshold that requires the decoded message to have a maximal hamming distance of τ\tau to the nearest user id. We select τ\tau to be the largest value for which the probability of a uniform random 31-bit message to land within distance τ\tau of one of the NN user ids is at most 1%.

AttackNoneBlurJPEGNoiseColorRotCropH-flipV-flipVAECtrlReg+
Attribution ↑, N = 1,000, τ = 399.999.999.996.694.07.923.519.86.0100.04.3
Attribution ↑, N = 10,000, τ = 2100.0100.099.993.691.35.812.912.53.2100.02.3
Attribution ↑, N = 100,000, τ = 1100.0100.099.989.988.84.24.55.01.3100.00.6
Attribution ↑, N = 1,000,000, τ = 0100.0100.099.780.183.42.20.91.30.3100.00.1
Misattribution ↓, N = 1,000, τ = 30.00.00.00.00.00.00.00.00.10.00.0
Misattribution ↓, N = 10,000, τ = 20.00.00.00.00.00.00.10.10.10.00.1
Misattribution ↓, N = 100,000, τ = 10.00.00.00.00.00.00.10.00.00.00.0
Misattribution ↓, N = 1,000,000, τ = 00.00.00.00.00.00.00.00.00.00.00.0

We report the attribution rate and misattribution rate of the end-to-end user attribution for up to one million users. We find that even for one million users we achieve near perfect attribution rate under no and simple attacks. Furthermore, the τ\tau gating results in a misattribution rate of at most 0.1% under all attacks. We added these results to Appendix F 4

W2b: Several geometric attacks also reduce long-payload decoding close to chance

Geometric attacks can be undone before decoding, either by estimating the transformation post-hoc (Table 14) or by adding a synchronisation mark such as SyncSeal [1] directly after generation. We performed experiments with SyncSeal with two additional changes. (1) The extractor learns the rotation angle based on the black corner artifacts left behind by rotation. On rotate-and-crop it therefore misjudges angles and misses rotations below 2°. We fine-tuned the extractor to fix this error. (2) SyncSeal always returns a transformation without a presence signal. When colour changes or regeneration damage its mark, its estimate is confidently wrong, and SyncSeal's own unwarping would apply it anyway. We instead train a small classifier that predicts whether an estimate is usable (AUC 0.98 on held-out attacks and image sources). We score each image both as received and after correction and keep the larger score, only when the classifier reports usable presence, at a slightly stricter threshold that keeps the false-positive rate at 1%. We have added the description of this setup to Appendix F 1.2.

W2b (continued)

AttackBlind TPR 31 bitBit acc. 31 bitBit acc. 62 bitBit acc. 124 bit
Rotate 20°0.69→1.000.82→1.000.76→1.000.66→1.00
Rotate 30°0.25→1.000.72→1.000.65→1.000.58→1.00
Rotate 45°0.04→1.000.65→1.000.58→1.000.53→1.00
Rotate 90°0.03→0.990.61→0.960.56→0.920.50→0.84
Rotate 180°0.12→0.990.65→0.960.55→0.920.48→0.83
Rot+crop 10°0.92→1.000.90→0.970.84→0.940.74→0.86
Rot+crop 20°0.54→0.980.79→0.950.72→0.900.63→0.82
Rot+crop 30°0.21→0.970.71→0.930.64→0.880.56→0.79
Rot+crop 45°0.06→0.960.62→0.940.57→0.880.51→0.80
Crop 0.70.71→0.960.84→0.930.75→0.860.65→0.77
Crop 0.50.13→0.190.70→0.710.62→0.630.55→0.56
H-flip0.86→1.000.86→0.980.79→0.950.60→0.87
V-flip0.18→0.980.73→0.970.54→0.930.55→0.86
Rotate (suite, ±180°)0.16→1.000.65→1.000.60→1.000.55→0.99
Rot+crop (suite)0.16→0.950.65→0.940.59→0.890.54→0.81

The table shows the bit accuracy and TPR of ByteMark without (before the arrow) → and with the SyncSeal correction (after the arrow). Over the 15 geometric attacks, the mean bit accuracy increases from 0.73 to 0.95 for 31 bits, from 0.65 to 0.91 for 62 bits and from 0.58 to 0.85 for 124 bits. Post-hoc watermarks gain the same robustness from SyncSeal. We have added these results to Table 18 along with its effect on multi user attribution in Table 28

W2c & Q2b: The radioactivity experiment uses one fixed message, leaving attribution under mixtures of different users' messages unresolved. Do you have complete identifier-recovery and false-attribution results when downstream training mixes different payloads?

We performed two additional experiments: attribution of a single user through radioactivity, and attribution of several users whose images are mixed in the training data.

Single user. We follow the fine-tuning setting of Section 4.2 and the attribution of our answer to your W2a. We have a single user's watermark in the fine-tuning data and report attribution, when the extracted message is sufficiently close to the embedded one. We report the attribution / misattribution rate for K=1 image and for the mean message extracted across K=256 images.

N (users)τM2 images K0%1%5%10%25%50%100%
1,00031-/0.00.0/0.00.0/0.00.0/0.012.5/0.094.5/0.0100.0/0.0
1,0003256-/0.00.0/0.026.0/0.0100.0/0.0100.0/0.0100.0/0.0100.0/0.0
10,00021-/0.00.0/0.00.0/0.00.0/0.05.5/0.581.0/0.0100.0/0.0
10,0002256-/0.00.0/0.012.5/0.0100.0/0.0100.0/0.0100.0/0.0100.0/0.0
100,00011-/0.00.0/0.00.0/0.00.0/0.00.5/0.558.0/0.099.5/0.0
100,0001256-/0.00.0/0.02.5/0.0100.0/0.0100.0/0.0100.0/0.0100.0/0.0
1,000,00001-/0.00.0/0.00.0/0.00.0/0.00.0/0.025.5/0.096.0/0.0
1,000,0000256-/0.00.0/0.00.0/0.093.5/0.0100.0/0.0100.0/0.0100.0/0.0

Our results show that when M2 is trained only on watermarked data (100% column), 96.0% of its single images are attributed to the correct user among one million users. Additionally, in all cases the misattribution rate never exceeds 0.5%.

Multiple users. We use a registry of 32 users using maximum simplex codes. We score each image of M2 with the oracle statistic (Eq. 4) for each registered user and attribute the image to the user with the highest score, if this score passes the calibrated 1% FPR threshold. This differs from the single-user table above, which uses blind detection and nearest-message decoding. We report attribution to a contributor / misattribution to a non-contributor, in % of single images of M2.

Users in training1% each5% each10% each25% each50% each100% each
0 (clean model)–/1.0–––––
10.0/0.01.0/0.03.0/0.084.0/0.0100.0/0.0100.0/0.0
20.0/0.00.0/0.04.5/0.097.2/0.2100.0/0.0–
42.5/0.81.0/0.212.0/1.099.6/0.0––
81.4/0.01.8/0.516.4/0.1–––
161.4/0.42.3/0.1––––

Misattribution is at most 1.0% in all models, which is the false-positive level of the threshold. When each user contributes 25% of the training data, 84.0–99.6% of the single images of M2 are attributed to a contributor, with at most 0.2% misattribution. At lower shares, a single image does not carry sufficient signal.

Q3: Please distinguish existing results from suggestions for future experiments.

We have clarified which experiments already exist in the paper. The submission contains the robustness evaluation (Tables 1 and 13), the radioactivity of a single message (Table 2) and the detection calibration (Appendix D). New in this rebuttal are the end-to-end user attribution for up to one million users (W2a), the geometric correction with SyncSeal (W2b), and the attribution through radioactivity for single and multiple users (W2c & Q2).

References:

[1] SyncSeal: “Geometric Image Synchronization with Deep Watermarking” Pierre Fernandez, Tomáš Souček, Nikola Jovanović, Hady Elsahar, Sylvestre-Alvise Rebuffi, Valeriu Lacatusu, Tuan Tran, Alexandre Mourachko. ICASSP 2026.

AuthorsOpenReview ↗

The Discussion Phase Is Coming to an End

Dear Reviewer,

We would again like to thank you for your review. We give a short summary of our response below.

  • W1 & Q1 Novelty: The direct transfer of Yoo et al. (2024) to bit-wise generators is our Naive baseline. It fails in quality and robustness (Section 3.1, Appendix C.2). ByteMark embeds the message in the transitions between bits, which keeps the balance of 0s and 1s intact.
  • W2a & Q2a User Attribution: We performed end-to-end user attribution with up to one million users. Under no and weak attacks, the attribution rate is at least 99.7%. Misattribution is at most 0.1% under all attacks.
  • W2b Geometric Attacks: We evaluated SyncSeal correction for ByteMark, which increases the mean bit accuracy over 15 geometric attacks from 0.73 to 0.95.
  • W2c & Q2b Multi-User Radioactivity: With 1 to 4 users at a share of 25% each, 84.0 - 99.6% of the single images of M2M_2 are attributed to a contributor, with at most 0.2% misattribution. Misattribution is at most 1.0% in all models, which is the false-positive level of the threshold.
  • Q3: We clarified which results exist in the paper and have added all rebuttal experiments to the manuscript.

As the rebuttal period is coming to an end, we would highly feedback on our response, and we are happy to answer any further questions.

Kind regards,

The Authors of ByteMark

Comments on the paper

2 comments
AuthorsOpenReview ↗

New results: user attribution among 1M users (≤0.1% misattribution), SyncSeal correction for geometric attacks (31-bit acc. 0.73→0.95), attribution through radioactivity for single and multiple users, and six forgery attacks, which content binding (SemBind) reduces to ≤2.7% attribution.

Dear Reviewers & AC,

Thank you for your time and effort in evaluating our work. We addressed all comments, ran new experiments, reported detailed results and clarified the points that were unclear in the initial submission.

To summarize, ByteMark “addresses an important yet underexplored problem” (nU8k) and “the need for multi-bit messages in bit-wise generators” (W7fA). The Reviewer (nU8k) describes the transition-based encoding as a “well-motivated technical design”. All Reviewers rate the presentation as “clear and easy to follow” (W7fA, nU8k, zZPT).

During the rebuttal, we added the following results: End-to-end user attribution (W7fA, zZPT). We attribute images to different users in experiments with one million users. We achieve near perfect attribution rate under no and simple attacks. The misattribution rate is at most 0.1% under all attacks. Improved geometric attack robustness (W7fA, zZPT). We integrate SyncSeal [1] into ByteMark, which increases the mean bit accuracy against geometric attacks from 0.73 to 0.95 for 31 bits, from 0.65 to 0.91 for 62 bits and from 0.58 to 0.85 for 124 bits. The false-positive rate stays at 1%. Attribution through radioactivity (W7fA, zZPT). We maintain the possibility of per-user attribution at a misattribution rate at or below 1.0% even for downstream models with 93,5% correct attribution among 1 million users using 256 images when users contribute a 10% share of data. Forgery attacks and defence (nU8k). We evaluate Averaging [2], Watermark Faker [3], WMCopier [4], LatentOpt [5], Frequency injection [5] and WMForger [6]. Frequency injection, Watermark Faker and WMCopier fail outright. Averaging, LatentOpt and WMForger can forge a user’s message, requiring many images, knowledge of the encoder and, as the current best one-shot attack, only a single image to succeed. Binding the watermark to the image content, following SemBind [7], also defeats all these attacks: at most 2.7% of forged images are still attributed to the victim. The binding uses no payload bits and costs 1.5 points of genuine attribution.

Once again, we thank the Reviewers for their careful reading and constructive feedback, which has helped us improve our work. We have done our best to address the raised concerns through detailed explanations and additional experiments. If our rebuttal addresses the Reviewer's concerns, we would appreciate it very much if they consider updating the score. We are also happy to address any remaining questions.

References:

[1] SyncSeal: “Geometric Image Synchronization with Deep Watermarking” Pierre Fernandez, Tomáš Souček, Nikola Jovanović, Hady Elsahar, Sylvestre-Alvise Rebuffi, Valeriu Lacatusu, Tuan Tran, Alexandre Mourachko. ICASSP 2026.

[2] Pei Yang, Hai Ci, Yiren Song, and Mike Zheng Shou. Can simple averaging defeat modern watermarks? NeurIPS, 2024.

[3] Ruowei Wang, Chenguo Lin, Qijun Zhao, and Feiyu Zhu. Watermark faker: Towards forgery of digital image watermarking. ICME, 2021.

[4] Ziping Dong, Chao Shuai, Zhongjie Ba, Peng Cheng, Zhan Qin, Qinglong Wang, and Kui Ren. WMCopier: Forging invisible watermarks on arbitrary images. NeurIPS, 2025.

[5] Andreas Müller, Denis Lukovnikov, Shingo Kodama, Minh Pham, Anubhav Jain, Jonathan Petit, Niv Cohen, and Asja Fischer. On the robustness of watermarking for autoregressive image generation. NeurIPS, 2026.

[6] Souček, T., Rebuffi, S.-A., Fernandez, P., Jovanović, N., Elsahar, H., Lacatusu, V., Tran, T., & Mourachko, A. (2025). Transferable Black-Box One-Shot Forging of Watermarks via Image Preference Models. NeurIPS, 2025.

[7] Xin Zhang, Zijin Yang, Kejiang Chen, Linfeng Ma, Weiming Zhang, and Nenghai Yu. SemBind: Binding diffusion watermarks to semantics against black-box forgery attacks. ICML, 2026.

Area_Chair_QvuiOpenReview ↗

Participate in the discussion period

Dear reviewers,

The authors have replied to your comments and the discussion is close to the end. We encourage you to participate in the discussion to see whether your questions and concerns have been answered by the authors. Thank you!

Best, Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.