acceptodds
Under review as a conference paper at ICLR 2027

KeyBound: Keyed and Host-Bound Learned Audio Watermarking for Speech Provenance

Abstract

Audio watermarking is a proactive route to attributing synthetic speech to its source. Learned audio watermarks are typically judged by payload recovery after a fixed catalog of signal distortions such as noise, compression, filtering, and resampling. That test is necessary but not sufficient for provenance. A mark offered as evidence of origin should not be readable by an unauthorized party, should not be transferable to unrelated audio, and should not vanish when the recording is re-synthesized by a modern generative model. We present KeyBound, a learned audio watermark that restores the two ingredients classical watermarking supplied and learned schemes set aside, a secret key and a host-aware carrier. KeyBound masks the payload with a secret key and embeds the masked bits through a carrier modulated by a host representation derived from a fixed log-mel transform and a learned projection, so the key governs payload access while the host-conditioned carrier resists direct transplantation. A key-independent presence head lets any party detect a mark, whereas only a key holder reads its attribution, and under the single-sample uniformity assumption a wrong-key decode clears our verification rule with probability at most . On LibriSpeech against WavMark, AudioSeal, and Timbre, KeyBound holds 1.00 detection accuracy, both at a threshold calibrated on clean audio and at an attack-matched 1% false-positive rate, and 0.98 bit accuracy under a spectral denoiser that costs every baseline its detection, decodes at chance without the key, and rejects transplanted carriers. Detection further transfers to held-out DAC and BigVGAN re-synthesis, though exact payload recovery degrades. Speech provenance is thus better posed as a keyed, host-bound attribution problem than as payload recovery under a fixed catalog of distortions.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
audio watermarking, speech provenance, neural re-synthesis

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_4EgXOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a post hoc speech watermarking algorithm, KeyBound that the watermark is both host-dependent and keyed. To achieve this, KeyBound masks the watermark payload with a private key and embeds the watermark audio, generated conditioned on the masked watermark payload. The framework's watermark decoder decodes the watermarked audio and outputs, if the audio is watermarked, the watermark payload. The watermark payload can only be decoded with the private key. This is how the watermark is keyed. For the host-dependent feature, it projects the masked watermark payload into latent space through an MLP and a FiLM module together with the projected host latent features. During the training process, it introduces a forgery loss that encourages the model to differentiate a watermark that is not from the matched host.

Critical strengths
  1. The paper identifies the research gap where the keyed watermark and host-dependent watermark are not unified, and the current methods only focus on one or another.

  2. The introduction of the mask key and the dual verification heads can further prevent watermark forgery by the user. With the secret key, it mathematically proves that the probability of an attacker successfully forging a watermark is very low.

  3. To make the watermark generation process host-dependent, the paper uses a FiLM module to condition the watermark generation process on the masked watermark payload.

Critical weaknesses
  1. It is not clear why the host anchor module ( C()+FiLM ) provides the host-dependent property that is unique in this paper. C converts host audio into the spectrogram domain, followed by a CNN layer. Then the FiLM fuses this feature with the masked watermark payload feature. However, a similar architecture was used in both Timbre and AudioSeal. Timbre converts the host audio into the spectrogram domain and concatenates the feature with the watermark payload features. AudioSeal also does a similar feature-merging operation. This type of concatenation can also be a type of feature merging, thus creating a host-dependent watermark generation. If the paper argues that the FiLM module provides more fine-grained feature merging, the experiment is needed. Without it, I don't see the novel design of the anchor module.

  2. For the keyed watermark mechanism, the dual-head module is from AudioSeal. The XAttnMark has a similar design, where it uses an embedding table to encode the watermark payload. This embedding table can also be considered to be a secret key.

  3. Try to use simpler words and short sentences, which can help the reader follow the paper. There are some math notations not explained and the Table 1 is not described in the paper.

Questions for authors

See the weaknesses and the following questions

  1. What is y in Section 3?

  2. What is host analysis C in Sec. 3.1

  3. In the Fig. 1 caption, not all notations appear in the figure. Can you elaborate on it?

  4. What is m~\tilde{m} and z in Sec. 4?

  5. Where is the explanation of transplant-negative training in Sec. 4?

  6. What is RR in Sec. 4.2

  7. Where is the description of Table 1?

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

This paper proposes KeyBound, a new audio watermarking scheme for speech provenance that uses secret-key protection and host binding. First, they use a secret key to mask the watermark payload, preventing unauthorized users from recovering it. Then, they use audio features to condition the watermark signal, making it harder to transfer the watermark to unrelated audio. They also train the model with different signal distortions and re-synthesis attacks to improve robustness. Experiments show that KeyBound achieves better robustness than existing audio watermarking methods under the evaluated attacks.

Critical strengths
  1. Interesting design. The paper combines secret-key payload protection with host-conditioned watermark embedding to improve watermark security. The separation between watermark detection and keyed payload verification is a useful design.

  2. Strong experimental results. The proposed method achieves better robustness than existing audio watermarking methods under various signal distortions and re-synthesis attacks, while keeping good audio quality.

Critical weaknesses
  1. Unclear motivation. The introduction does not clearly explain why payload protection and host binding are necessary. Specifically, why would knowing the payload make an attacker to estimate or remove the watermark? How can an attacker forge a watermark without host binding? These questions are not answered, and make further watermark design not that necessary.

  2. Missing related work. The paper misses several relevant studies about audio watermark security, for example, watermark robustness [1] and watermark removal [2]. A more comprehensive discussion is needed to justify the position of this paper.

[1] Wen et al., SoK: How Robust is Audio Watermarking in Generative AI Models?, arXiv 2025.

[2] Yao et al., Audio Pirates: Black-box Audio Watermark Removal via Diffusion Priors, arXiv 2026.

Questions for authors
  1. Could the authors clarify the motivation for payload protection and host binding?

  2. Why does knowing the payload help attackers estimate or remove the watermark?

  3. Why is host binding necessary to prevent watermark forgery? WavMark already resists transplantation, what additional security benefits from host-binding mechanism?

  4. The paper should clearly explain what security problems these designs aim to solve.

Reviewer_HY2nOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces KeyBound, an audio watermarking scheme that requires a key to unlock the payload. The authors claim that the proposed design is host-bound, meaning that it is conditioned on the host signal itself, and key-protected. They also claim that it shows significant advantages in robustness against perturbations compared to baselines under the same conditions.

Critical strengths

I applaud the paper's focus on audio watermarking that survives neural codec attacks. This is a very timely and challenging subject.

The experimental results are encouraging, showing strong robustness compared to baselines.

Critical weaknesses

However, there are fundamental issues with the authors' claims and weaknesses in the experimental setup.

  1. There is a closely related prior work, MaskMark [1], that shares many of the same features as the proposed method. First, MaskMark is host-bound, with bits embedded in a similar fashion to this paper. Second, the watermark is also keyed, as MaskMark converts a 6-bit key into a 512-bit code before embedding it into the audio. To detect the watermark, the key is needed; otherwise, the bit match rate will be around 50%. There is a slight difference in that MaskMark uses the key to detect the presence of a watermark rather than to unlock the payload. Third, MaskMark is shown to be robust against resynthesis and denoisers. Hence, this method should be cited and compared against.

  2. Because the key mapping is essentially a mapping from a 16-bit code to another 16-bit code, it is entirely possible to add such a mapping to any existing audio watermarking method that has sufficient payload capacity. This also means that methods in audio steganography can be trivially turned into keyed "audio watermarking" methods by the authors' definition. For example, Pavlovic et al. [2] also embed watermarks using a trained DNN to conditionally hide the input code in the signal, making the method host-bound.

  3. Prior watermarking methods, such as AudioSeal and the additional references [1] and [2], are all general-purpose methods that support watermark localization in long audio signals. However, the proposed method is trained and evaluated in a significantly more specialized setting: (A) It is trained and tested on the same type of data—clean audio samples from LibriSpeech—which may bias the model toward this particular data distribution. (B) It is trained on 1-second clips. Because the audio length is fixed, the embedder can easily hide information associated with sample positions, making the watermark easier to detect in the experimental setting but harder to localize in real-world scenarios. (C) The experiments (Table 1) do not include any tests involving phase shifts or sample shifts, effectively avoiding exposure of the issue described in (B).

  4. While the method is robust under the tested conditions, one metric that stands out is the PESQ score of 3.03, which suggests that the embedder introduces substantial distortion that may be perceptible. To put this number into perspective, both AudioSeal and [2] achieve PESQ scores above 4.4, indicating significantly better perceptual quality.

Suggested references:

[1] P. O'Reilly, et al., "MaskMark: Robust Neural Watermarking for Real and Synthetic Speech," ICASSP 2024. [2] K. Pavlovic, et al. "Robust Speech Watermarking by a Jointly Trained Embedder and Detector Using a DNN." Digital Signal Processing, 122, 103381. https://doi.org/10.1016/j.dsp.2021.103381.

Questions for authors

I would be interested to hear whether the authors agree or disagree with my criticisms, especially point (2). In particular, I believe that an existing host-bound audio watermarking method with sufficient payload capacity could share the key features of the proposed method simply by adding a key-mapping step. I would appreciate hearing how the authors would argue against this interpretation and clarify the novelty of their approach. I am open for discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.