acceptodds
Under review as a conference paper at ICLR 2027

mAVE: A Watermark for Joint Audio-Visual Generation Models

Abstract

Watermarking joint audio-visual generation supports vendor copyright protection and content provenance. However, independently valid audio and video watermarks do not establish a shared generation session. An adversary can splice watermarked modalities from different sessions, causing the pair to be mistaken for the vendor's original joint output. We introduce mAVE (Manifold Audio-Visual Entanglement), a training-free watermarking framework that strengthens vendor attribution through session binding in native joint audio-visual diffusion transformers. mAVE separates public record retrieval from secret session authentication: a fixed public index locates the server record, while a randomized payload binds audio bits to a session-keyed video grid through a cryptographic digest. One prompt-conditioned joint inversion supports provider-assisted verification of both modalities against a session record, without modifying generator weights or training auxiliary watermark networks. Our analysis establishes implementation-matched distribution preservation and a full-initialization routing/clipping budget, alongside adaptive session-pool security and stable local-perturbation bounds. Experiments on LTX-2 and MOVA show comparable generation quality. mAVE achieves 99.8% true-positive rate and 0% observed false-positive rate in the evaluated swap test, and retains 99.2% true-positive rate under FrameAvg temporal averaging. Same-prompt and similarity-selected swaps further test session authentication beyond perceptual compatibility.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Joint Audio-Visual Generation, GenAI Watermark, Audio-Visual Binding

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_5PzsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents mAVE, a training-free watermarking method for authenticating whether audio and video originate from the same generation session. It links their latent payloads through a hash and uses a public index to retrieve a private session record. Verification combines inversion conditioned on the original prompt with record-based payload and binding checks. The paper provides conditional security analysis and evaluates two joint audio-video generators, reporting 99.8% TPR and zero observed FPR in the principal swap tests, with comparable generation quality.

Critical strengths
  1. The paper distinguishes provider-level watermark detection from authentication of an audio-video pairing. The same-prompt and similarity-selected swaps in Table 5 provide useful tests beyond semantic compatibility.

  2. The method requires no watermark-specific training and is evaluated on two native joint generators. Tables 2 and 4 support its feasibility under common processing operations while maintaining broadly comparable generation quality.

Critical weaknesses
  1. The proposed embedding procedure closely follows Gaussian Shading, making the session-binding mechanism central to the contribution. However, the verifier hashes a server-reconstructed grid rather than the recovered video (Section 4.3). Although Table 5 reports an advantage over a random session-specific reference, the paper does not sufficiently explain why deriving the reference from the video grid produces this advantage.

  2. Section 4.4 assumes conditional uniformity of the recovered binding vector and bounds on perturbation-induced bit changes. The main text does not establish these properties for the actual inversion and decoding pipeline or instantiate a useful perturbation bound. The reported swap tests and correctly stated 0.30% upper confidence limit support performance on the evaluated distributions, but do not establish broader adaptive security. Targeted disruption of the public index is also untested, leaving the robustness of record retrieval unresolved.

Questions for authors
  1. What explains the SR-Rand/mAVE gap in Table 5? Please provide the actual calibrated thresholds, binding-score distributions, and counts passing index retrieval, modality checks, and binding verification.

  2. How does verification perform when neither the original prompt nor generation-request metadata is available for joint inversion? Please provide experimental results, including authentication TPR/FPR and index-recovery rates, and specify how the missing conditioning is handled.

  3. For the binding lengths and inversion steps already evaluated in Section 5.6, how do full authentication TPR/FPR and index-recovery rates vary? Mean BA alone does not establish authentication reliability.

  4. Does the removal threat include targeted corruption of the public index, and what evidence supports robustness to it? When request metadata supplies the original prompt, does it also identify the session record, and what additional role does the embedded index serve?

8 comments
AuthorsOpenReview ↗

Response to Reviewer 5Pzs — Part 1/5

Q1 — Reference reconstruction, calibration and stage statistics

Scope: The additional prompt-free, parameter-sweep and attack experiments below use LTX-2 / T2AV (text-to-audio-video generation). Original main-paper comparisons retain their reported evaluation scope.

Thank you for requesting the intermediate statistics and distinguishing protocol availability from complete authentication.

Q1. Reference reconstruction and the SR-Rand/mAVE gap.

Reference reconstruction addresses recovery noise; the SR-Rand comparison addresses the binding construction. The provider rebuilds the ideal video grid because directly hashing a recovered grid would turn even a small video recovery error into a different digest. The received video must still pass its own comparison against the selected record. This provides a stable binding target without removing the requirement to authenticate the query video.

SR-Rand is a matched control for this reference choice. Its audio target is an independently sampled string stored in the video-selected session record. It retains the payload budget, carrier positions, repetition, recovery pipeline, noise-free reference and dedicated binding gate. It is therefore a session-referenced control, not a comparison against the donor audio's own reference. The observed advantage cannot be attributed merely to comparing mAVE with a noisier reference or a weaker record-selection rule.

Calibrated operating point. We use strict thresholds greater than 0.8 for video BA, audio BA and the effective binding score. On each method's 2,000 clean validation cases, we scan a common scalar threshold over {0.70,0.75,0.80,0.85,0.90}\{0.70,0.75,0.80,0.85,0.90\}, applying it to all three scores simultaneously, and select the largest candidate achieving at least 99.7% full authentic acceptance. All three methods select 0.8, with validation recall of 99.8% for mAVE and 99.7% for both controls. The thresholds are then frozen for all test groups. SR-ID's effective score is the minimum of its two reference-match scores. Supplementary Table R4 gives the complete scan.

The table below gives the principal counts.

MethodAuthentic acceptedPre-binding passes (cross / same / selected)Final accepts (cross / same / selected)
mAVE998/10001 / 24 / 380 / 0 / 0
SR-Rand997/100098 / 94 / 10269 / 66 / 71
SR-ID997/1000116 / 130 / 11682 / 92 / 81

Each swap group contains 1,000 attempts. R1 gives the complete routing and gate counts, including correct-index recovery, successful lookup, wrong-record and no-record outcomes. In the similarity-selected group, all methods have 999 successful lookups and 990 video-gate passes. Their joint modality-pass counts then differ, with 38 for mAVE, 102 for SR-Rand and 116 for SR-ID. Binding verification further reduces those counts to 0, 71 and 81. The gap therefore occurs both in the samples reaching the binding gate and in the fraction accepted by that gate. Lookup failures alone do not account for it.

Recovered-score structure. R2-R3 report distributions before modality filtering and among its survivors, including means, SDs, quantiles and histograms. Over all scorable swaps, mAVE has 0/1,000, 0/999 and 0/999 scores above 0.8. The corresponding SR-Rand counts are 105/1,000, 101/999 and 112/999; SR-ID has 124/1,000, 138/999 and 126/999. Thus the high-score cases in the controls are already present before modality selection.

For similarity-selected SR-Rand swaps, the median is 0.50781, but 112 scores exceed 0.8, including 85 above 0.9. A mean of 0.54839 consequently hides a substantial high-score subgroup alongside the central random-like scores. Of those 112 high-score cases, 71 pass the modality gates, a rate of 63.4%; among the other 887 scorable cases, only 31 pass, a rate of 3.5%. The high-score subgroup therefore constitutes 71/102=69.6%71/102=69.6\% of the modality survivors. These are direct proportions from the reported counts, not an assumed mixture model.

AuthorsOpenReview ↗

Response to Reviewer 5Pzs — Part 2/5

Q1 — Gate interaction, mechanism and statistical scope

Continuing Q1, Part 1 gives the high-score subgroup and stage counts. Here we explain the gate interaction and delimit the statistical claim.

Why the gates favor this subgroup. The binding field is included in the full audio payload, so the two scores share information. With N=128N=128 and L=512L=512, let SrestS_{\mathrm{rest}} denote accuracy on the other 384 bits. Then

BAa=128512Sbind+384512Srest.\mathrm{BA}_{a}=\frac{128}{512}S_{\mathrm{bind}}+\frac{384}{512}S_{\mathrm{rest}}.

At BAa>0.8\mathrm{BA}_{a} > 0.8, a sample with Sbind=0.5S_{\mathrm{bind}}=0.5 needs Srest>0.9S_{\mathrm{rest}} > 0.9, while one with Sbind=0.92S_{\mathrm{bind}}=0.92 needs only Srest>0.76S_{\mathrm{rest}} > 0.76. A high binding score therefore also helps a sample pass the aggregate audio check. Once such cases exist in the recovered-score distribution, the two gates do not act as independent opportunities to reject them. The exact empirical decomposition for the similarity-selected group is

FPRSR−Rand=1021000×71102=7.1\mathrm{FPR}_{\mathrm{SR-Rand}}=\frac{102}{1000}\times\frac{71}{102}=7.1\\%.

FPRmAVE=381000×038=0\mathrm{FPR}_{\mathrm{mAVE}}=\frac{38}{1000}\times\frac{0}{38}=0\\%\quad\text{(observed)}.

This decomposition requires no independence assumption. It explains how the score distribution and the gates combine into the reported FPR. It also identifies the limit of the selection explanation. Filtering enriches the existing high-score subgroup but does not explain its original appearance.

How this relates to random tags and cryptographic dependency. Independently sampling stored tags does not, by itself, establish a uniform conditional law for the vector recovered after joint inversion and decoding of a mismatched pair. The verifier scores that recovered vector against the selected record; it does not merely compare two ideal tags drawn independently at generation time. Under a uniform recovered-vector model, a random reference and a fixed digest reference would indeed have the same Hamming-ball matching probability. The measured control gap therefore concerns end-to-end recovery and verification. It does not imply an intrinsically large collision probability for 128-bit random strings. The aggregates characterize the differing recovered-score distributions and the gate interaction, but do not identify a hash-only cause for that distributional difference.

This interpretation is consistent with the same-sample ablation. On mAVE media, adding the binding gate rejects 1, 24 and 38 otherwise accepted swaps without changing generation, decoding or authentic acceptance. Aggregate audio accuracy can hide concentrated errors in the binding field. For example, 64 matching binding bits and 346 matching remaining bits give BAa=410/512>0.8\mathrm{BA}_{a}=410/512 > 0.8 while Sbind=0.5S_{\mathrm{bind}}=0.5. The explicit field check therefore adds a requirement that the modality average alone does not enforce. The matched-reference controls then show that its benefit depends on the recovered scores supplied by the complete construction, not just on the presence of an extra gate.

Statistical scope. Correct-index recovery is an evaluation diagnostic, not an extra ground-truth gate. We distinguish exceedances among all scorable cases, conditional acceptance among modality survivors, and final acceptance among all attempts. For mAVE the respective zero-exceedance denominators are 1,000/999/999; the survivor denominators are 1/24/38; and the complete-verifier denominator is 1,000 in each group. A one-sided independent-binomial 95% upper limit for 0/1,000 is approximately 0.30%; the nominal conditional limits for 0/1, 0/24 and 0/38 are 95.0%, 11.73% and 7.58%. These statements have different populations and require the relevant sampling assumptions; they are not interchangeable or bounds for correlated adaptive queries.

The mAVE swap-score means near 0.5 and SDs near 0.044 are compatible with the shape of Binomial⁡(128,1/2)/128\operatorname{Binomial}(128,1/2)/128, whose SD is approximately 0.04419. They do not prove the full joint conditional law or certify a deployment error rate of 9.67×10−139.67\times10^{-13}. The five-step false acceptance discussed below provides a concrete reason to keep that ideal model separate from the numerical verifier.

Taken together, the evidence supports an operational contribution. mAVE couples a session-specific binding construction with a field-level verification rule that separates the evaluated genuine and swapped pairs at comparable genuine acceptance. The same-sample ablation isolates the added check; the matched-reference controls demonstrate the complete construction's advantage at the reported operating point. We will make this distinction explicit in the paper, with reference reconstruction justified by stability and the session-authentication benefit supported by the measured score distributions and full-verifier outcomes.

AuthorsOpenReview ↗

Response to Reviewer 5Pzs — Part 3/5

Q2 — One-pass verification without the original prompt

Q2. Verification without the original prompt or request metadata.

We clarify a one-pass, caller-prompt-free route. The caller supplies the AV pair; the verifier performs one joint inversion with empty text conditioning and extracts the public index from the fixed, unmasked video positions. This index requires no session secret. The provider then retrieves the selected record's original prompt and secret, derives KsessK_{\mathrm{sess}} using that stored prompt as in Eq. (2), and derandomizes and verifies the same recovered latents. No second inversion is required.

The empty inversion condition and the original prompt used in key derivation have distinct roles. This route does not use a caller-supplied prompt or request handle, but still depends on the provider's session database. Empty-text inversion has methodological precedent in Gaussian Shading and VideoShield; the following results address mAVE's LTX-2 / T2AV verification performance.

On LTX-2 / T2AV, the empty-text results have video/audio BA 0.935/0.919, corresponding to changes of -0.1/+0.4 percentage points relative to the LTX-2 / T2AV reference values 0.936/0.915 in Table 1. T2AV has no input-image condition. The verifier does not require the original input image, an extracted conditioning frame, the original prompt or a request handle. The generator and its configuration are known to the provider; the session prompt and secret are retrieved only after index extraction. The table below reports the complete outcomes.

Pair groupAttemptsCorrect indexWrong recordNo recordFinal acceptsRate (%)
Authentic100099730997TPR 99.7
Cross-prompt1000999100FPR 0.0
Same-prompt1000998200FPR 0.0
Similarity-selected1000998110FPR 0.0

The positive exact-index recovery and authentication TPR are both 99.7%. There are three wrong-record positive lookups and no missing positive lookups; the overall final accepted count is 997. Observed swap FPR is 0% in each of the three 1,000-attempt groups, with all failed lookups retained in the denominator. These results establish the measured performance of this LTX-2 text-free route. A change relative to original-prompt conditioning requires a matched LTX-2 cohort; the original aggregate 99.8% TPR and the manuscript's MOVA BA values are not used as that paired baseline.

We will revise Section 4.3 and the corresponding workflow descriptions in Appendices E.2, E.5 and I to distinguish cinvc_{\mathrm{inv}}, the inversion condition, from PrecordP_{\mathrm{record}}, the stored prompt used in key derivation. We will present the original-prompt route and this one-pass, empty-text route separately. Both use the record prompt for key derivation; only the former uses it to condition inversion.

Reviewer_P8uDOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces mAVE, a training-free watermarking framework for joint audio-visual diffusion transformers. Rather than independently watermarking each modality, mAVE binds the audio and video latents so that cross-session splicing can be detected. The method achieves 99.8% TPR and 0% observed FPR on the evaluated swap test, but the performance drops a bit under other attacks.

Critical strengths

The problem is well-motivated and practically relevant. The idea of linking audio and video latents is interesting and reasonable. The experiments are reasonably comprehensive, covering various attack types and baselins comparisons.

Critical weaknesses

The paper's presentation is a significant concern. The design choices and symbols are introduced without sufficient motivation or intuitionm, making it hard to follow the details in each step. The authors should consider to substantially revise the paper before resubmission.

The robustness to various attacks is another concern. The performance under some attacks drops noticably compared with other benchmark systems, espeically for the audio attacks.

Questions for authors

I just use Sections 3 and 4.1 as examples to illustrate which aspects need further clarification, but the same issue occurs throughout the paper.

  • The threat model assumes the adversary has generator weights. Which practical use cases may motivate this assumption? Typically the adversary only has access to the outputs.
  • The paper mentions perceptually plausible swaps but does not explain how they are constructed in practice.
  • The terms "bound grids" and "bit-grid construction" are used without prior definition. It would be better to provide a brief description and clarify why it is designed as grid and what the benefits compared with other options.
  • Regarding the "initial latent spaces" Z_v and Z_a, are they the initial noise tensors at the start of the denoising process?
  • For Equations (1) and (2), it would be helpful to add a brief intuitive motivation before the formal definitions, including the rationale for each component.
  • The sets I_bind and I_time in Equation (4) are not clearly defined.
  • B_v and B_a are described as one-dimensional grids but appear as matrices in Figure 2. Pleaes clarify this inconsistency.
  • It would be better to visually distinguish the three disjoint regions of B_v in Figure 2 and briefly explain the motivation for this partition.
  • Similarly, for the audio grid B_a construction, it would help to clarify the intuition behind this design choice.
4 comments
AuthorsOpenReview ↗

Response to Reviewer P8uD — Part 1/2

Method motivation, threat model and logical representation

Scope: The additional prompt-free, parameter-sweep and attack experiments below use LTX-2 / T2AV (text-to-audio-video generation). Original main-paper comparisons retain their reported evaluation scope.

Thank you for identifying both the presentation problem and the robustness trade-off. We agree that the original exposition introduced several implementation details before explaining what the verifier is trying to establish. We have prepared a replacement explanation organized around the session record, the three logical fields in each modality, and the verification decision, together with a revised Figure 2. The proposed replacement passages accompanying this response make these changes concrete.

Overview before the equations. Every generation request creates a private session record. The video carries a public address for that record. The audio carries a binding field derived from the record's video grid, in addition to its other payload bits. At verification, the public address selects one record; the provider then checks the recovered video, the recovered audio, and the designated binding field against that same record. A valid watermark in each track does not by itself mean that the two tracks belong together. This operational explanation will precede the notation and cryptographic derivation.

Why assume access to the generator weights? A provider can operate an openly released generator, such as LTX-2, while keeping its per-request session secrets private. An attacker may therefore know the generator architecture, weights and inversion procedure even when it cannot obtain those secrets. Weight access models this deployment and a stronger attacker than output access alone; it is not an assertion that all real attackers possess the weights. The verification service holds the cryptographic secret, which is separate from the generator parameters.

How are compatible swaps constructed? Appendix J.4 uses three operational constructions involving audio from a different prompt, audio from a separate session with the same prompt, and audio selected for highest ImageBind similarity from a fixed pool of 20 other sessions. In each case we splice the actual media and jointly invert the resulting pair; we do not assemble separately extracted bitstrings. Same-prompt selection and high embedding similarity control semantic compatibility but do not guarantee lip synchronization or human-perceived plausibility. We will use the more precise description "same-prompt and similarity-selected swaps" when no human plausibility judgment is available.

What is a grid, and why use it? BvB_v and BaB_a are length-LL one-dimensional indexed bit arrays, not inherently two-dimensional matrices. A grid specifies which logical bit serves which purpose. Its representation makes field allocation, repetition and mapping to different latent tensor shapes explicit; it does not provide a separate geometric security property. The latent tensor may have channel, time and spatial dimensions even though the logical payload is one-dimensional. A matrix depiction represents reshaping, not a different definition of BvB_v or BaB_a.

Why partition the video grid? Its disjoint fields serve three different jobs. The synchronization template reserves alignment information, the public index allows record retrieval before the verifier knows the session key, and the base payload provides session-specific bits for the modality check. Public index copies remain at fixed, unmasked video coordinates JIJ_I; other coordinates lie in JRJ_R and undergo keyed permutation and per-copy masking. Combining these jobs into an undifferentiated field would obscure why routing must be possible before secret-dependent decoding. The revised Figure 2 uses distinct field colors and separately shows their mapping into latent coordinates.

What are the initial latent spaces and Eq. (1)? zvz_v and zaz_a are the video and audio noise tensors at the start of denoising, and ZvZ_v and ZaZ_a are their respective spaces. The implemented hash acts on the discrete logical video grid, not on a floating-point tensor. We will introduce the generation-time binding condition directly as Ba[i]=SHA ⁣− ⁣256(Bv)[ϕ(i)]B_a[i]=\mathrm{SHA\!-\!256}(B_v)[\phi(i)] for i∈Ibindi\in I_{\mathrm{bind}}, then explain how repetition, randomization and inverse-CDF sampling realize it in initial noise. "Authentic manifold" names this constrained family of initializations; the construction does not require a smooth geometric manifold. This also separates an exact generation-time relation from the error-tolerant verification rule.

AuthorsOpenReview ↗

Response to Reviewer P8uD — Part 2/2

Key derivation, audio binding, verification and robustness

Why the components of Eq. (2)? A fresh server secret mm distinguishes generation requests, including requests with identical prompts. The normalized prompt identifies the text component used in key derivation but contributes no assumed secret entropy. Hashing each component produces fixed-length inputs before their concatenation; the outer hash produces KsessK_{\mathrm{sess}}. The secret-hash prefix has fixed length independent of NN, so changing the number of binding comparisons does not change the key-security parameter. Lowercasing and trimming apply to key derivation, while original-prompt generation uses the original text. Purpose-separated subkeys assign different roles to payload generation, masking and permutation. These are implementation choices for session-specific and purpose-separated derivation, not a claim that this exact hash nesting is the only valid design.

What are IbindI_{\mathrm{bind}} and ItimeI_{\mathrm{time}}, and why construct the audio this way? ItimeI_{\mathrm{time}} contains logical synchronization-template slots. IbindI_{\mathrm{bind}} is a disjoint set of NN audio slots reserved for distinct selected digest bits; ϕ\phi maps those slots to digest positions. Basea\mathrm{Base}_a fills the remaining positions. We will define these sets before the piecewise equation and label them in Figure 2. Audio has no public-index field. Both modalities are checked using the one record selected by the video index. A separate binding threshold prevents a high average over the rest of the audio payload from hiding errors concentrated in the binding field. For example, 64 correct binding bits and 346 correct remaining bits yield audio BA 410/512>0.8410/512>0.8 even though the binding score is only 0.5.

Why reconstruct the reference on the server? Directly hashing a recovered video grid would turn even one recovery error into a largely different digest. We instead rebuild the ideal grid from the selected record and use its digest as a stable target. The received video still has to pass its own comparison with that same record. Reference stability explains this verification design; it does not by itself establish superiority to a stored random session tag. We address the empirical SR-Rand comparison separately. This explanation does not support a claim of universal superiority.

Verification flow made explicit. Received AV pair →\rightarrow one joint inversion →\rightarrow public video-index decoding →\rightarrow session-record lookup →\rightarrow reconstruction of the grids, masks and reference →\rightarrow video/audio BA checks plus the binding check →\rightarrow Accept/Reject. For the additional LTX-2 / T2AV empty-text route, no image condition is required. Its record prompt is retrieved after inversion for key derivation; it is distinct from the empty text used to condition inversion. This sequencing will be stated alongside the generation figure so that the public and secret-dependent stages are easy to follow.

Robustness and its limits. We agree that mAVE does not uniformly match the extraction robustness of specialized modality watermarks. Table 4 reports video BA 0.927 for mAVE versus 0.948 for VideoShield under FrameAvg, and 0.914 versus 0.943 under the brightness attack. The clean audio BA comparisons with WavMark and Timbre also do not support a blanket claim of higher extraction accuracy. AudioSeal scores and BA are different quantities, so those particular columns are not a common-scale detection comparison; this distinction does not remove the limitations visible in the comparable extraction results.

The relevant additional capability is session-pair authentication. In the reported audio tests, complete mAVE authentication retains TPR/FPR of 98.5%/0.1% under 64-kbit/s MP3 compression and 97.4%/0.2% under 20 dB audio noise. These are within-system results, not proof of superior sample-level detection to AudioSeal. A relative robustness claim would require each detector's genuine detection rate at a common false-alarm operating point and matched distortion settings. We will keep those objectives and metrics separate.

Temporal robustness also needs qualification. Appendix G reports mAVE video BA of 0.710 under local FrameSwap, substantially below its 0.927 under FrameAvg. BA alone does not give the complete authentication TPR for FrameSwap, so we will not infer one. The additional LTX-2 / T2AV results likewise expose concrete trade-offs. One-step inversion gives only 65.7% genuine acceptance, and the strongest listed targeted-index perturbation reduces genuine acceptance from 99.8% to 88.1%. We will describe robustness for the evaluated transformations and budgets, acknowledge these extraction and availability limitations, and retain session binding as the claimed additional capability. We do not assert uniform superiority under every distortion.

Thanks for the repsonse and clarification. They address some of my concerns, and I'v raised my score to 2.

Thank you for your detailed feedback on the presentation and robustness of our work. We have uploaded a revised manuscript, with substantive revisions marked in blue. Below we indicate where the manuscript addresses your concerns.

1. Method presentation, notation, and Figure 2.

We have revised Figure 2 (p. 4) to show the video and audio payloads as one-dimensional logical bit sequences, with distinct routing, synchronization-template, and base/binding fields. The figure separately illustrates their mapping into model-shaped latents and the verification workflow.

The corresponding definitions appear in Section 4.1 (pp. 3–5): the latent-level notation summarizes a construction whose hashes operate on discrete bit grids; the video payload has three distinct fields; and the audio construction defines I_time, I_bind, and the digest-bit selector before the piecewise equation. Appendices E.1–E.3 (p. 28) further explain the public index, server-held secret, prompt normalization, and purpose-separated keys. Appendix K (pp. 41–44) collects the notation.

2. Threat model and construction of compatible swaps.

Section 3 (p. 3) distinguishes access to generator weights from access to the provider’s private session secrets. The motivating deployment is a provider using publicly available generator weights while retaining per-request secrets privately; weight access is a stronger threat model, not an assumption that every attacker possesses the weights.

Appendix J.4 (p. 39) specifies the three swap constructions: cross-prompt, same-prompt/different-session, and similarity-selected audio from a fixed pool of 20 other sessions. These tests splice the actual media and jointly invert the resulting pair. Same-prompt and similarity-based selection control semantic compatibility; they do not establish human-perceived plausibility or guaranteed lip synchronization.

3. Verification sequence and reference reconstruction.

Section 4.3 (pp. 5–6) and Appendix E.5 (pp. 29–30) clarify the sequence from joint inversion and public-index decoding to record lookup, reference reconstruction, and the three acceptance checks. They distinguish the inversion condition from the stored prompt used for key derivation. The additional caller-prompt-free results appear in Table 6(a) (p. 9).

4. Robustness and its limitations.

We agree that session authentication does not imply uniformly superior extraction robustness. Table 1 (p. 7) and Table 4 (p. 8) distinguish extraction bit accuracy, AudioSeal scores, and complete-pair authentication rates. For example, mAVE’s full-verifier TPR/FPR is 98.5%/0.1% under MP3 compression and 97.4%/0.2% under 20 dB audio noise; these are within-system results, not evidence of uniformly better detection than AudioSeal.

Section 5.5 (p. 9) and Appendix G (pp. 32–33) document processing and temporal limitations. Table 6(b) and Table 7 (p. 9) additionally expose targeted-index availability loss and finite-step authentication failures. Appendix B.3, Table B.2 (p. 26) reports complete authentication alongside extraction accuracy, including the 97.1% TPR and one false acceptance at five steps.

Thank you again for identifying where the original presentation needed clarification. We hope these revisions make both the construction and its empirical limitations substantially easier to assess.

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

This paper studies cross-modal watermarking and session-level provenance authentication for joint audio-visual generative models. It proposes mAVE (Manifold Audio-Visual Entanglement), a joint audio-visual watermarking framework that requires neither modifying the parameters of the generative model nor training additional watermarking networks. In the initial latent variables of the joint generative model, it establishes a cryptographic dependency between the video and audio watermark payloads, and during verification, this relationship is recovered through joint inversion. It achieves promising results in experiments on generation quality, cross-session swap attacks, and robustness.

Critical strengths

The authors not only focus on whether the watermark can be extracted, but also consider scenarios where the watermarks are legitimate but the audio-video pairing relationship is invalid.

The paper's approach establishes a cryptographic relationship between the audio watermark and the video watermark, checking not only whether each modality carries a legitimate watermark, but also whether their binding relationship corresponds to the same generation session.

Critical weaknesses

I am confused about whether the paper's core advantage comes from cross-modal cryptographic dependency or from watermark encoding and reference verification schemes that are easier to recover and more reliable.

The theoretical design considers some degree of adaptivity, but the experiments have not yet sufficiently demonstrated that the system remains equally reliable under feedback-driven attacks. For example, an attacker may continuously modify the audio and video based on authentication results and repeatedly submit verification attempts for the same target session.

Questions for authors

Q1:Could the authors clarify whether the paper's advantage stems from the cross-modal dependency itself or from differences in encoding and decoding reliability?

Q2:How does the system address adaptive attackers who iteratively modify swapped audio–video pairs based on repeated accept/reject feedback?

5 comments
AuthorsOpenReview ↗

Response to Reviewer YzEd — Part 1/3

Q1 — Binding contribution and reference controls

Scope: The additional prompt-free, parameter-sweep and attack experiments below use LTX-2 / T2AV (text-to-audio-video generation). Original main-paper comparisons retain their reported evaluation scope.

Thank you for distinguishing individual watermark validity from authentication of an audio-video pairing.

Q1. Does the improvement come from binding or from encoding/decoding reliability?

The evidence supports two complementary conclusions. The dedicated binding check contributes beyond the recovered modality scores, and the complete mAVE construction gives better session separation than the matched server-reference controls at the evaluated operating point. We distinguish these conclusions because an additional verification rule and the construction that supplies its recovered bits are different interventions.

First, the same-sample ablation isolates the binding check. It reuses exactly the same generated media, inversion outputs and recovered scores, changing only the final acceptance rule. Removing the binding gate produces 1/1,000, 24/1,000 and 38/1,000 false acceptances for cross-prompt, same-prompt and similarity-selected swaps. Restoring it rejects all of those cases while authentic acceptance remains 998/1,000. This improvement cannot be attributed to using a different encoder, a more accurate inversion or a different recovered bitstream, since those quantities are held fixed.

The table below shows that binding scores also separate the evaluated sessions before selection by the modality gates.

mAVE groupScorable nMean binding scoreSDScore >0.8
Authentic10000.920970.024021000
Cross-prompt10000.501230.044120
Same-prompt9990.499400.044130
Similarity-selected9990.500490.043730

Thus the observed separation is not confined to the small subset that already passes the modality checks. Failed record lookups have no binding score and remain in the denominator of complete-verifier performance.

Second, SR-Rand and SR-ID address whether the gain is simply due to access to a reliable server reference. SR-Rand retains the embedding budget, recovery pipeline, stored noise-free reference and dedicated gate, while changing the binding target's construction. SR-ID checks both recovered tags against the video-selected record. Both therefore have session-referenced verification; SR-Rand is not an unbound baseline and is not checked against the donor's own record. Their measured swap FPRs are 6.6 to 7.1% and 8.1 to 9.2%, respectively, versus 0% observed for mAVE, at authentic TPRs of 99.7%, 99.7% and 99.8%. A difference between noisy and noise-free references cannot account for this comparison.

The stage statistics explain how the measured construction gap reaches the final decision. For similarity-selected SR-Rand swaps, 112/999 scorable samples exceed the binding threshold before modality filtering; 71 of those survive the modality gates. They make up 71/102, or 69.6%, of the survivors. The binding field also contributes to aggregate audio BA, so those two checks share information. In contrast, mAVE has no binding exceedance over the 999 scorable swaps, and its binding check rejects all 38 modality survivors. Our response to 5Pzs gives the full distributions and the exact gate decomposition.

The resulting claim is specific. The same-sample experiment establishes the incremental value of explicit pairing verification, while the matched controls establish an empirical advantage of the complete session-bound construction at comparable genuine acceptance. Independent random tags and a fixed digest reference have the same ideal matching tail under a uniform recovered-vector model; the experiments do not establish a universal cryptographic ordering between them or identify a hash-only cause for the recovery-distribution difference. The contribution demonstrated here is reliable session discrimination through the implemented generation, recovery and verification pipeline, beyond two individually acceptable modality scores.

AuthorsOpenReview ↗

Response to Reviewer YzEd — Part 2/3

Q2 — Attack objective, constraints and search methods

Q2. Repeated accept/reject feedback.

We evaluate repeated verification against a fixed target session, with 1,000 targets per strategy and at most 1,000 verification calls per target. Every decision is made by the complete verifier, and the attack sees only Accept/Reject. The reported target-level attack success rates at the largest budget are 0/1,000 for random proposals and 3/1,000 for each of the other two strategies. Thus the observed risk is small in these tests but is not zero.

Attack objective and admissibility. For an original swap x−=(vA,aB), A≠Bx^-=(v_A,a_B),\ A\ne B, a successful candidate must retain the mismatched donor content, remain within the perturbation budget relative to x−x^-, and pass full authentication. We use the constraints ∥v′−vA∥∞≤ϵv\lVert v'-v_A\rVert_\infty\le\epsilon_v and 10log⁡10∥aB∥22∥a′−aB∥22≥Smin⁡10\log_{10}\frac{\lVert a_B\rVert_2^2}{\lVert a'-a_B\rVert_2^2}\ge S_{\min}. Their reference remains the original swap throughout the search; a sequence of small updates cannot reset the total budget. Equivalently, the audio ℓ2\ell_2 radius is ρa=∥aB∥2 10−Smin⁡/20\rho_a=\lVert a_B\rVert_2\,10^{-S_{\min}/20}. Acceptance of an authentic initialization sample does not constitute forgery.

Media constraints and content retention. Video is normalized to [0,1][0,1] with ϵv=4/255\epsilon_v=4/255; audio is normalized to [−1,1][-1,1] with Smin⁡=30 dBS_{\min}=30\,\mathrm{dB}. Frame count, audio sample count, duration and alignment are preserved. We require cosine similarity of at least 0.98 between the candidate and original donor CLAP audio embeddings. For speech, we additionally require WER no greater than 5% relative to the donor transcript obtained with the same fixed ASR system and normalization. These criteria operationalize donor-content retention; an accepted authentic initialization does not satisfy the cross-session forgery objective merely because it passes authentication.

Random proposals independently sample the available attack parameters. Rejection-history search uses the same candidate representation but excludes parameter configurations that have already been rejected at the search's parameter resolution. This tests non-repeating search under binary feedback; it does not use a secret score or gradient. Because all responses before the first successful attack are rejections, such a sequence can also be generated without replacement in advance. We do not present this strategy alone as a strong test of decision-boundary optimization or attribute its difference from random proposals solely to information gained from feedback.

Candidate representation. The proposal pool is indexed by video amplitude in {0,1,2,3,4}/255\{0,1,2,3,4\}/255, audio SNR in {30,35,40,45,+∞}\{30,35,40,45,+\infty\} dB, and pseudorandom direction seed in {0,…,1023}\{0,\ldots,1023\}; +∞+\infty denotes zero audio perturbation. Both proposal strategies use the same perturbation parameterization. Random proposals sample parameter configurations with replacement, whereas rejection-history search excludes previously tested configurations. This exclusion is defined over the parameter configuration, not by access to any private verification score.

The known-positive-seed strategy performs decision-boundary search from an accepted authentic pair. The search seeks an accepted candidate inside the permitted neighborhood of the original swap; remaining close to the authentic seed is not the attack objective. Its five initialization/diagnostic calls are included in the total budget. A positive diagnostic response is excluded from the attack-success count, and internal record-identity diagnostics are not returned to the attack.

Positive-seed update rule. We initialize with x+=(vA,aA)x^+=(v_A,a_A) from the target session and minimize the normalized distance d(x,x−)=max⁡{∥v−vA∥∞/ϵv, ∥a−aB∥2/ρa}d(x,x^-)=\max\{\lVert v-v_A\rVert_\infty/\epsilon_v,\ \lVert a-a_B\rVert_2/\rho_a\} over accepted candidates. The search alternates movement toward x−x^- with random tangential proposals around the current accepted point, retaining an update only when the verifier accepts and dd decreases. The initial inward and tangential proposal scales are 0.01 in normalized search coordinates. After every 10 calls, the scales are multiplied by 1.2 when at least half of the proposals were accepted, or by 0.5 otherwise, and clipped to [10−4,0.1][10^{-4},0.1]. An accepted initialization may lie outside the admissible neighborhood and cannot then count as a forgery. Success requires d≤1d\le1, the donor-content checks and full authentication. Initialization, boundary probes and final validation all count toward qq, including the five initialization/diagnostic calls.

AuthorsOpenReview ↗

Response to Reviewer YzEd — Part 3/3

Q2 — Query outcomes and scope of the security analysis

The following results use the 1,000-target binary-feedback evaluation described in Part 2.

The table below reports successful targets at each budget.

Strategyq=1q=1q=10q=10q=100q=100q=1,000q=1{,}000Target ASR at q=1,000q=1{,}000 (%)
Random proposals00000.0
Reject-history search00030.3
Known-positive-seed searchN/A0130.3

N/A indicates that a one-call budget cannot complete the five-call initialization. It does not represent a completed attack with zero successes. Every other entry counts targets with at least one admissible false acceptance by that budget; a target is counted only at its first success. The three successful targets for each feedback strategy satisfy the stated media and donor-content constraints and retain the intended target record. The two strategies may succeed on overlapping targets. Record identity is checked by the evaluator and is not additional feedback available to the attack.

Interpretation and relation to theory. Target-level ASR is the probability of at least one successful candidate within the budget, not a per-query FPR. The small number of observed successes does not establish a statistical ranking between strategies or security against the optimal attack. Our fixed-pool result does not require independent queries, while the first-success bound min⁡{1,qϵsingle}\min\{1,q\epsilon_{\mathrm{single}}\} requires ϵsingle\epsilon_{\mathrm{single}} to upper-bound the best allowed no-verification-feedback constructor, including offline candidate selection. A random-swap FPR cannot be substituted for that quantity. The positive-seed experiment also includes genuine-positive diagnostic queries, outside the attack-only first-success formulation as stated.

More fundamentally, the ideal recovered-bit law is an assumption about the complete decoder, not a consequence of naming a cryptographic primitive. The five-step experiment in our response to 5Pzs contains one false acceptance at N=128N=128; it therefore cannot be assigned the ideal 9.67×10−139.67\times10^{-13} single-test probability as an established deployment guarantee. Similarly, the successful query attacks are not explained by a known certified perturbation bound, since we have not instantiated the required bit-stability and routing certificate for these media budgets. We report the empirical attacks directly, retain the conditional theorem with its premises, and do not infer adaptive security from zero observed random-swap errors.

Comments on the paper

6 comments

Dear Reviewers,

Thank you for your time and effort in reviewing our paper. We have provided detailed responses and additional experimental results to address the concerns raised in the reviews.

We would greatly appreciate any further feedback or follow-up questions you may have. We are happy to provide additional clarifications and hope to continue a constructive discussion.

Thank you again for your consideration!

Area_Chair_rv5gOpenReview ↗

ICLR Review Discussion – Follow-up on Author Rebuttal

Dear Reviewers,

Thank you for your time and effort in reviewing this submission.

The authors have submitted their rebuttal. Could you please take a moment to review their responses and assess whether your concerns have been adequately addressed? If you have any remaining concerns or new questions, please feel free to raise them in the discussion.

If you have already responded to the authors, please kindly disregard this reminder.

Thank you again for your valuable contributions to the review process.

Best,

AC

We thank all reviewers and the Area Chair for their time, careful reading, and constructive feedback throughout the discussion.

We have now uploaded a revised manuscript incorporating the clarifications, additional experimental results, and revisions discussed in our responses. Substantive additions and revisions are marked in blue to facilitate review.

The revision clarifies the method and verification protocol, includes fuller empirical evaluations and diagnostics, and refines the theoretical framing and causal attribution to reflect the evidence and its limitations. Our individual replies provide pointers to the relevant sections, tables, and pages.

Thank you again for helping us improve the clarity, completeness, and scientific precision of the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.