acceptodds
Withdrawn from ICLR 2027

SelfCheckMark: Confidence-Driven Dual-Channel Watermarking and Self-Correcting Extraction for Video Diffusion Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

As video diffusion models gain wider use and generated videos become more prevalent, reliable copyright attribution and provenance tracing become increasingly important. In-generation watermarking addresses this need by embedding information during synthesis. A prominent route modifies the initial diffusion noise and recovers the embedded watermark information through video inversion. Yet generation, video distortions, and inversion errors yield recovered latent evidence of uneven reliability, so weak observations can dilute strong ones. Moreover, correcting an inverted noise estimate before combining redundant observations can fail when its errors exceed the code's tolerance. To address these challenges, we propose SelfCheckMark, a blind video watermarking framework with confidence-aware extraction and post-fusion self-correction. During embedding, extended Hamming(8,4) coding and a dual-channel design assign each coded bit to a strong-response anchor and continuous redundant elements. During extraction, local values are summed so elements with larger absolute values contribute more, while channel–frame replicas are weighted by inverse standard deviation, reducing the influence of unreliable evidence. Only after replica fusion is the most likely legal Hamming codeword selected, allowing correction to act on the combined evidence. Experiments show that, with a 512-bit watermark, SelfCheckMark achieves 100% mean bit accuracy on clean videos, over 98% under Gaussian noise and blur, and 85.36% under H.264 compression. Ablations confirm both confidence fusion and post-fusion correction.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
diffusion model watermarking, generative watermarking, confidence fusion, error-correction coding, maximum likelihood decoding

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 2 reviews, above 0% of the others.
Reviewer_TJvMOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes SelfCheckMark, an in-generation video watermarking method that embeds information into the initial diffusion noise. The method combines repeated sign embedding, Hamming error-correction coding, and confidence-weighted extraction to improve watermark recovery after video distortions.

Critical strengths

Using continuous latent values instead of hard bit decisions during extraction is a reasonable idea.

Critical weaknesses
  • Several components are introduced, but their actual contributions are unclear. For example, the authors describe a dual-channel design with a strong-response anchor and redundant elements. However, all four elements undergo the same embedding operation and are simply summed during extraction. I do not understand what makes this a dual-channel design or what the anchor actually adds. At 512 bits, the simple repetition baseline outperforms the proposed Hamming-based approach under all evaluated attacks. This makes me question the choice of Hamming coding. The additional improvement from the weighting is also very small.

  • Only three attacks are evaluated, each at a single strength. Important video transformations such as cropping, resizing, frame dropping, frame reordering, and other temporal edits are missing. The evaluation is also limited to one video generation model.

  • Another concern is the baseline performance. VideoMark and SIGMark perform poorly under several conditions, sometimes close to random guessing, without sufficient explanation. I would like to understand whether these results are consistent with their original implementations. The improvements over VideoShield are also relatively small, with no confidence intervals or repeated-run analysis to establish their significance.

  • The authors only report an aggregate VBench-2.0 score. No qualitative examples compare watermarked and unwatermarked videos. I would expect a more detailed analysis of visual quality and temporal consistency, especially since the method modifies the initial noise. The authors also claim that sign scrambling preserves the noise distribution, but this is not experimentally demonstrated.

  • The method repeats the same watermark information across channels and frames and uses a global key with a fixed nonce. I am concerned about the security of this construction when the same key or watermark information is reused across multiple videos. There is no analysis of possible watermark removal, forgery, or information leakage from repeated patterns.

  • More importantly, the authors only report bit accuracy. There are no detection metrics such as TPR/FPR, wrong-key verification, or exact payload recovery. Recovering watermark bits is not the same as verifying the origin of a video, which is one of the main motivations of the paper.

Questions for authors
  • What does the dual-channel design actually add compared with simple repeated sign embedding?

  • Why is Hamming coding preferred when pure repetition performs better at 512 bits?

  • How is the proposed confidence measure related to actual recovery errors?

  • Can the authors provide broader video robustness experiments, detection metrics, and more convincing video-quality results?

  • How secure is the method when the same key and watermark information are reused across videos?

Reviewer_4wdnOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

SelfCheckMark is an initial-noise watermark for image-to-video diffusion (HunyuanVideo-I2V). A 512-bit payload (480 custom + 32 version bits) is encoded with extended Hamming(8,4) into 1,024 bits. Each coded bit is written into the signs of a 2×2 block, with magnitudes kept, in every one of the 256 channel–frame planes of the 16×16×64×64 initial latent. The signs are then scrambled with a ChaCha20 mask derived from a global key and a fixed nonce. The embedding builds on Gaussian-Shading-style sign embedding; the claimed contributions are on the extraction side.

Extraction inverts the query video (50-step FlowMatch Euler, empty prompt) and descrambles the signs. It then fuses the evidence at two levels, which is the paper's first claimed contribution. The four soft values in each block are summed rather than sign-voted (the "dual-channel" anchor-plus-redundancy design), and the 256 replicas are combined with weights inversely proportional to each replica's standard deviation. The second claimed contribution is "post-fusion self-correction": only after fusion is each 8-bit block decoded, by maximum correlation with the 16 legal codewords (maximum-likelihood soft decoding).

Results on 400 HunyuanVideo-I2V videos with a 512-bit payload and three distortions (Gaussian noise, H.264, blur) show bit accuracy on par with VideoShield, the strongest baseline (within about 1 point; 85.36% vs 84.52% under H.264), and no measurable quality change. Ablations also show that soft summation is the main source of gain, while the Hamming code beats plain repetition only at ≤128 bits and is worse at the 512 bits used for the main results.

Critical strengths

Clean ablations of the extraction side.

  • Keeping soft magnitudes instead of sign voting gives consistent gains, largest under H.264 (+4.1–4.5 points, Table 3).
  • The method is training-free, keeps model weights fixed, and extraction is blind (global key, no per-video template).
Critical weaknesses

Limited novelty

  • The embedding is Gaussian-Shading-style sign embedding with spatial, channel and frame repetition plus ChaCha20 scrambling.
  • The "dual-channel" design amounts to summing the four elements of each 2×2 block. The "anchor" element is embedded identically and is not treated differently at extraction (Eq. 1).
  • The layout (Sec. 3.3) differs from Gaussian Shading only by being more local, and that is not an improvement.
  • Soft combining before decoding and ML (correlation) decoding of Hamming(8,4) are textbook techniques.
  • More specifically, soft extraction for Gaussian-Shading-style watermarks is prior art, and none of the following is cited:
    • Gaussian Shading++ replaces GS's hard voting with maximum-likelihood decoding of the repetition code, i.e. summed per-copy evidence. It also introduces a "double-channel design", a name that collides with this paper's "dual-channel".
    • The PRC watermark decodes soft, magnitude-weighted values with BP-OSD.
    • VideoMark supports soft decisions.

The paper's own results undercut its second contribution ("post-fusion self-correction"):

  • At the evaluated 512-bit capacity, pure repetition with the same soft combining beats the Hamming code under every distortion (H.264 89.39% vs 85.36%; noise 99.37% vs 98.90%; blur 98.68% vs 98.04%; Sec. 4.3 and Table 8). The proposed coding therefore lowers the headline numbers.
  • With repetition, the method would beat VideoShield by ~5 points under H.264. That gain would come from soft combining, not from the proposed code.

The evaluation is too narrow to support robustness or state-of-the-art claims

  • One model; three distortions, each at a single strength.
  • Bit accuracy only: no detection TPR at fixed FPR on unwatermarked videos.
  • No temporal attacks (frame drop, insertion or swap; temporal crop; frame-rate change), no spatial crop, resize or rotation, no other codecs or bitrates, and no regeneration or adversarial attacks.
  • The paper cites WAVES as its "unified attack protocol" but uses only three of its attacks.
Questions for authors
  1. Address weaknesses.
  2. Broader attacks, with strength sweeps.
    • Temporal: frame drop, insertion, swap or averaging; frame-rate change.
    • Spatial: crop, resize, rotation.
    • Codecs: MPEG-4/HEVC at several bitrates.
    • Regeneration attacks.
  3. Replica weighting.
    • How does sigma_r separate signal from noise?
    • Please compare with SNR-based or maximal-ratio weights, e.g. decision-directed estimates of ArA_r and srs_r per replica.
  4. Prior soft decoding and geometric attacks.
    • Please compare with Gaussian Shading++'s soft ML decoding of repetition codes, and with PRC's soft decoding.
    • Please test geometric attacks. Inversion-based video watermarks such as VideoShield drop to ≈0.52 and ≈0.51 bit accuracy under a 90% center crop or a 15° rotation in SPDMark's evaluation (SVD-XT).

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

Related papers

Discussion (0)

Sign in to comment.