- 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
- 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.
- 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.
- 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?
- 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.
- 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.
- 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
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.
| Attack | Blind TPR 31 bit | Bit acc. 31 bit | Bit acc. 62 bit | Bit acc. 124 bit |
|---|---|---|---|---|
| Rotate 20° | 0.69 → 1.00 | 0.82 → 1.00 | 0.76 → 1.00 | 0.66 → 1.00 |
| Rotate 30° | 0.25 → 1.00 | 0.72 → 1.00 | 0.65 → 1.00 | 0.58 → 1.00 |
| Rotate 45° | 0.04 → 1.00 | 0.65 → 1.00 | 0.58 → 1.00 | 0.53 → 1.00 |
| Rotate 90° | 0.03 → 0.99 | 0.61 → 0.96 | 0.56 → 0.92 | 0.50 → 0.84 |
| Rotate 180° | 0.12 → 0.99 | 0.65 → 0.96 | 0.55 → 0.92 | 0.48 → 0.83 |
| Rot+crop 10° | 0.92 → 1.00 | 0.90 → 0.97 | 0.84 → 0.94 | 0.74 → 0.86 |
| Rot+crop 20° | 0.54 → 0.98 | 0.79 → 0.95 | 0.72 → 0.90 | 0.63 → 0.82 |
| Rot+crop 30° | 0.21 → 0.97 | 0.71 → 0.93 | 0.64 → 0.88 | 0.56 → 0.79 |
| Rot+crop 45° | 0.06 → 0.96 | 0.62 → 0.94 | 0.57 → 0.88 | 0.51 → 0.80 |
| Crop 0.7 | 0.71 → 0.96 | 0.84 → 0.93 | 0.75 → 0.86 | 0.65 → 0.77 |
| Crop 0.5 | 0.13 → 0.19 | 0.70 → 0.71 | 0.62 → 0.63 | 0.55 → 0.56 |
| H-flip | 0.86 → 1.00 | 0.86 → 0.98 | 0.79 → 0.95 | 0.60 → 0.87 |
| V-flip | 0.18 → 0.98 | 0.73 → 0.97 | 0.54 → 0.93 | 0.55 → 0.86 |
| Rotate (suite, ±180°) | 0.16 → 1.00 | 0.65 → 1.00 | 0.60 → 1.00 | 0.55 → 0.99 |
| Rot+crop (suite) | 0.16 → 0.95 | 0.65 → 0.94 | 0.59 → 0.89 | 0.54 → 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 training | 1% each | 5% each | 10% each | 25% each | 50% each | 100% each |
|---|---|---|---|---|---|---|
| 0 (clean model) | – / 1.0 | – | – | – | – | – |
| 1 | 0.0 / 0.0 | 1.0 / 0.0 | 3.0 / 0.0 | 84.0 / 0.0 | 100.0 / 0.0 | 100.0 / 0.0 |
| 2 | 0.0 / 0.0 | 0.0 / 0.0 | 4.5 / 0.0 | 97.2 / 0.2 | 100.0 / 0.0 | – |
| 4 | 2.5 / 0.8 | 1.0 / 0.2 | 12.0 / 1.0 | 99.6 / 0.0 | – | – |
| 8 | 1.4 / 0.0 | 1.8 / 0.5 | 16.4 / 0.1 | – | – | – |
| 16 | 1.4 / 0.4 | 2.3 / 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.
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
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.
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.
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.
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.
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. We select τ as the largest value for which a uniform random message is within distance τ of one of the N ids with a probability of at most 1%.
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 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 M2.
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%.