acceptodds
Under review as a conference paper at ICLR 2027

Simplifying Robust Audio Watermarking

Abstract

Multi-bit audio watermarking models are generally found vulnerable under deep watermark removal attacks, such as neural codec resynthesis and voice cloning. Existing methods jointly apply strategies like attack simulation, latent watermarking, and perceptual losses to improve robustness while preserving imperceptibility. However, we discover that most of these strategies are not necessary to achieve state-of-the-art robustness. To figure out key designs that enable robust audio watermarking, we propose SimpleMark, a simple yet effective audio watermarking framework. SimpleMark narrows down robust audio watermarking to two design choices: First, we embed the watermark through **bounded magnitude scaling** to control imperceptibility. Second, we detect watermark from **low-level audio representations** using a lightweight readout head to achieve robustness. Following these two designs, we build a series of lightweight audio watermarking models that surpass existing complex methods in robustness while maintaining competitive imperceptibility.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
audio watermarking, audio provenance, low-level audio representations, speech security

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_ioA5OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces SimpleMark, a minimalist audio watermarking framework that breaks away from paradigms relying on complex designs such as latent space embedding, attack simulation, and perceptual loss. By employing only bounded multiplicative scaling in the magnitude spectrum for embedding and lightweight linear decoding based on low-level audio representations, the model achieves state-of-the-art robustness against potent attacks, while maintaining high imperceptibility.

Critical strengths
  1. This work challenges the conventional belief that strong robustness requires complex latent-space operations or training-time attack simulations, demonstrating that a simple linear scaling predictor and a linear detection head can outperform existing sophisticated baselines.

  2. The authors evaluate the watermarking model under extensive realistic attack scenarios, including 18 traditional editing attacks, 30 neural codecs, 7 zero-shot voice cloning, and 4 AI resampling attacks.

Critical weaknesses
  1. SimpleMark exhibits substantial performance degradation on certain specific codecs, such as Language-Codec (approximately 50%), WavTokenizer (approximately 64%–71%), and SQ-Codec (approximately 52%).

  2. Section 4.4 reveals that the watermark learned by SimpleMark may primarily manifest as a fixed frequency pattern shared across time and independent of audio content. The authors do not sufficiently demonstrate that the watermark can withstand adaptive removal attacks specifically designed to exploit its structural characteristics.

  3. The methodological novelty is questionable, as the proposed approach appears to be an effective combination of existing designs rather than an entirely new watermarking mechanism.

Questions for authors
  1. The description of the training pipeline is oversimplified and lacks crucial details. The paper presents only a single detection loss in Eq. (2) without explaining how the end-to-end gradient flows from the Detector back through the frozen audio representation backbone, the iSTFT/STFT operators, and into the Embedder's Scale Predictor. It remains unclear whether the Embedder and Detector are trained jointly or in alternative stages, and how numerical stability/gradients are maintained through complex magnitude scaling and phase recombination during iSTFT.

  2. Section 4.4 reveals that the watermark learned by SimpleMark may primarily manifest as a fixed frequency pattern shared across time and independent of audio content. The authors do not sufficiently demonstrate that the watermark can withstand adaptive removal attacks specifically designed to exploit its structural characteristics. For example, can an attacker remove the watermark using spectral averaging followed by notch filtering or spectral subtraction while preserving audio quality? Furthermore, could averaging multiple watermarked audio clips enable an attacker to extract the common frequency template and forge ownership claims?

  3. In Appendix C.1, the authors adopt a 75% bit-matching threshold for existence classification without explanation. What is the justification for this choice? For K=16K=16, a 75% threshold yields a high False Positive Rate (FPR≈3.84%\text{FPR} \approx 3.84\%) under random guessing.

3 comments

We sincerely appreciate your constructive feedback and insightful comments. Now we are going to respond to all your questions and concerns.

Question 1: What is the detail of the training pipeline?

Thank you for pointing out these ambiguities. We jointly train the watermark embedder and detector in a single stage using the detection loss in Eq. (2) as the only supervision. In each training step, we assign a randomly-initialized ground-truth watermark message m\mathbf m to each audio sample, train the embedder and detector in an end-to-end manner with the target of decoding m\mathbf m correctly from the watermarked audio. Although the audio representation extractor is frozen, gradients with respect to its input waveform (aka the watermarked audio xm\mathbf x_m) are retained. From xm\mathbf{x}_m, gradients backpropagate through the differentiable iSTFT to the watermarked magnitude spectrogram M~\widetilde{\mathbf{M}}, then through the magnitude scaling and tanh⁡\tanh operations to update the scale predictor S\mathcal{S}. This gradient flow requires neither differentiating phase P\mathbf P nor dividing by low-energy amplitude bins, so numerical stability naturally holds. Meanwhile, the tanh⁡\tanh function helps prevent gradient explosion by limiting the gain applied to each bin. After all, the training pipeline of SimpleMark is deliberately designed to be simple and requires no additional stabilization techniques.

Question 2: Is SimpleMark robust against watermark removal or forgery attacks based on spectral averaging?

Thanks for this insightful question. Indeed, we find that existing anti-cloning audio watermarking models, including SimpleMark, are generally vulnerable under adaptive watermark removal and/or forgery attacks. To the best of our knowledge, this kind of attack has not been reported in other literature. Specifically, we consider two types of black-box attacks:

  1. Adaptive Watermark Removal Attack: In this kind of attack, the attacker wants to remove the watermark from a watermarked audio sample$A$. Let MA′M'_A and PA′P'_A denote the STFT amplitude spectrogram of AA, and μt\mu_t the mean across time frames. The attacked STFT amplitude is produced by dividing the original spectrogram with per-bin average, or formally, M~A=MA′/μt(MA′)\widetilde M_A=M'_A/\mu_t(M'_A). Then the attacked waveform is produced by A~=iSTFT⁡(M~A,PA′)\widetilde A=\operatorname{iSTFT}(\widetilde M_A, P'_A).

  2. Adaptive Watermark Forgery Attack: In this kind of attack, the attacker wants to transfer the watermark from a watermarked audio sample$A$ to an unwatermarked sample BB. Let MA′M'_A and MBM_B denote the STFT amplitude spectrogram of AA and BB, respectively. The attacked STFT amplitude of B is then produced by M~B′=MB⋅μt(MA′)μt(MB)\widetilde M_B'=M_B\cdot\frac{\mu_t(M'_A)}{\mu_t(M_B)}.

We introduce three metrics to evaluate whether the attack is effective:

  1. Bitwise Accuracy (%): Bitwise accuracy after attack. Lower indicates more effective removal attacks, whereas higher indicates more effective forgery attacks.

  2. Attack Successful Rate (ASR, %): For removal attacks, ASR is the proportion of attacked samples predicted as unwatermarked after the attack. For forgery, it is the proportion of samples predicted as watermarked after the attack.

  3. PESQ: Attacked audio quality. An attack should not be considered successful if it significantly damages audio quality.

Below we report attack results on both baselines and our proposed system:

MethodRemoval Bit Acc. (%) ↓Removal ASR (%) ↑Removal PESQ ↑Forgery Bit Acc. (%) ↑Forgery ASR (%) ↑Forgery PESQ ↑
AudioSeal99.790.03.4749.851.23.22
WavMark99.680.03.4450.141.03.22
Timbre55.9389.93.2792.6395.83.12
AWARE99.900.03.4949.780.23.16
VoiceMark54.4590.33.6180.3659.23.16
AlignMark94.611.33.4854.323.13.23
SimpleMark55.3290.53.1895.0297.93.10

The results show that all three systems designed for anti-cloning audio watermarking, namely Timbre, VoiceMark and SimpleMark, are susceptible to both adaptive removal and adaptive forgerty attacks. This indicates that watermarks produced by these systems may share the same nature of time-invariant, content agnostic frequency pattern. However, this kind of attack significantly degrades watermarked audio quality. It is therefore questionable whether this kind of attack should be considered practical in reality.

(Continued in the next comment...)

Question 3: Why is 75% chosen as the threshold used for watermark existence prediction?

As shown in the table below, we choose 75% as the bit-matching threshold since it maximizes mean existence accuracy across four conditions: unattacked audio, codec resynthesis, voice cloning, and AI attacks. We think that this choice balances FPR and FNR under diverse attacks. However, one can also adopt a stricter threshold like 87.50% when FPR matters more. At that threshold, our method can achieve 0.16% FPR with only a 2.06pp decrease in mean existence accuracy.

Threshold (%)Theoretical FPR (%)Mean Empirical FPR (%)No Atk (%)Codecs (%)Cloning (%)AI (%)mean (%)
50.0059.8259.9370.7368.3069.5067.7269.06
56.2540.1840.2680.0676.7279.4075.6377.95
62.5022.7222.5288.7483.6987.7181.5685.42
68.7510.5110.3394.9287.1394.0985.4490.40
75.003.843.7598.0587.2396.8186.9992.27
81.251.061.0199.5084.8996.2587.2891.98
87.500.210.1799.9481.0892.8786.9590.21
93.750.0260.02399.9475.7484.9285.5186.53
100.000.00150.000999.2767.2671.9579.4879.49

Question 4: Why does SimpleMark perform poorly on Language-Codec, WavTokenizer, and SQ-Codec?

To begin with, it is worth noticing that some codecs are challenging across all evaluated watermarking methods: As shown in the following table, even the strongest baseline suffers from substantial performance degradation on SQ-Codec and WavTokenizer. Therefore, performance degradation on these codecs does not imply that SimpleMark fall behind baselines or is not robust to general neural codecs.

As for the reason why SimpleMark performs poorly on certain codecs, we observed in the following table that recostructed audio quality of these codecs is also comparably low. One plausible explanation is that these codecs heavily alter acoustic details that carry watermark information, making the embedded bits harder to recover.

Nevertheless, we would like to emphasize that SimpleMark achieves substantial gains on several other codecs. For example, compared with the strongest baseline, it improves bit accuracy by 9.52pp on SemantiCodec_0.35 kbps (E3), 6.06pp on SoCodec (M), and 4.66 pp on SemantiCodec_0.70 kbps (E2). On average, SimpleMark achieves the highest mean bit accuracy under neural codec resynthesis attacks, indicating its superior robustness.

CodecBit Acc. (%)Best Baseline (%)PESQ
DAC (B)99.7599.943.809
Spectral Codecs (V)99.0099.232.331
Vocos_encodec_12 (L2)98.3998.723.001
EnCodec-6b24k (D1)96.3998.892.331
…………
Language-Codec (J)50.0091.261.066
WavTokenizer (I1)64.0074.261.918
WavTokenizer (I2)71.4175.991.993
SQ-Codec (W)52.7951.791.222

Question 5: What is the methodological novelty beyond an effective combination of existing watermarking designs?

SimpleMark is not a combination of existing designs. Instead, it is a framework building upon two novel mechanisms: Bounded magnitude scaling for imperceiptibiliy control and low-level audio representation-based watermark detection for robustness. These two mechanisms enable superior robustness and competive imperceptility without mainstream techniques like simulated attacks, latent watermarking and perceptual losses, which builds up the core novelty of this paper.

Our default system comprises a linear watermark embedder and a QWen3-ECAPA-based watermark detector. To the best of our knowledge, this kind of architecture has not been reported in existing literature. Indeed, we show in Tab. 2 that our framework is compatible with some network architectures and representations used in existing audio watermarking models. However, choices these alternative modules hardly affect robustness or imperceiptibility. Finally, we also regard simplicity of our framework as an innovation: We demonstrate that state-of-the-art robust audio watermarking can be achieved with few model parameters (12.6M), training data (10k utterances) and simple model architecture.

Thank you for the reply and the additional experiments. Being simple and effective is the selling point of the design, and it is supported. The rating is increased to 3.

Reviewer_dG1COpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper propose SimpleMark, a simple yet effective audio watermarking framework. It first embeds the watermark through bounded magnitude scaling to control imperceptibility, and then implements a lightweight model to robustly detect watermark. It performs well on strong attacks like voice cloning and neural codec processing. It outperforms more complex existing methods in robustness while keeping audio quality competitive.

Critical strengths

SimpleMark achieves state-of-the-art robustness against the hard attacks, such as neural codecs and voice cloning, without needing attack simulation, perceptual losses, or complex architectures. The model architecture is simpler, more interpretable, and easier to train than prior work.

Critical weaknesses

Overall, it is still a robustness-imperceptibility tradeoff. The high accuracy on nueral codecs and voice cloning attacks seem to come at the cost of audio quality. The Audio quality metrics (PESQ ~3.9, MOS ~3.78) fall below 4, meaning watermarked audio has noticeable distortion.

Questions for authors
  • It would be useful to explain why the loss in Equation (2) was designed this way, and how sensitive the results are to the choice of the margin value gamma.

  • The scaling residual for each frame seems to be predicted using only that frame's information. Would it help to also use nearby frames as context when making this prediction?

  • In Table 1, both PESQ and MOS scores are below 4 for SimpleMark, which suggests noticeable artifacts. Would it be possible to compare the robustness given the same level of Imperceptibility (which may be achievable by adjusting tau)?

  • Table 3 compares various audio representations for watermark detection. However, the detector representation only affects how the watermark is retrieved given the watermarked audio. That is, changing the detector should not change audio quality (imperceptibility) since the watermarked audio has been generated. It is unclear why the imperceptibility changes along with the detector representatoin in Table 3 (also the Table 2).

4 comments

Question 1: Why is the loss in Eq. (2) designed this way? How sensitive are the results to the classification margin γ\gamma?

Thanks for pointing out this issue. The loss in Eq. 2 is a smoothed variant of the margin-based hinge loss max⁡(0,γ−mℓ)\max(0,\gamma-m\ell). It encourages the predicted watermark logit ℓ\ell not to just cross but to lie at least γ\gamma away from the decision boundary. At γ=0\gamma=0, it gegrades to the standard binary cross-entropy loss with logits.

The table below shows the ablation results on margin γ\gamma. Larger γ\gammaenhances robustness at the cost of lower audio quality. To balance robustness and imperceiptibility, we set γ=2\gamma=2 where the robustness gain saturates. Nevertheless, the performance of SimpleMark is not sensitive to the classification margin γ\gamma: Performance varies within 1% on bitwise accuracy and 0.02 on PESQ when γ\gamma sweeps from 0 to 3.

γ\gammaHiFiCodec Bit Acc (%)PESQ↑STOI↑NISQA↑VISQOL↑
081.593.95870.95933.77924.0380
181.693.94210.95883.77204.0300
1.581.943.94300.95843.77594.0229
282.293.94720.95833.78034.0130
2.582.353.94610.95843.78104.0022
382.323.94790.95843.78073.9923

Question 2: Would incorporating neighboring frames as context improve the framewise scaling-residual predictor?

Not necessarily. In Tab. 2, we compare three kinds of scale predictors: Linear predictors that produce per-frame scaling factor independently, gated convolution networks that capture local context with 3×33\times3 convolutions, and conformers that capture global context through self-attention (see Fig. 6 in Appendix B.2.1 for detailed network architecture). The results show that all three scale predictors perform equally well, with gated convolution incorporating neighboring frames performs slightly worse. Therefore, we conclude that larger perceptual field does not necessarily improve the performance of scale predictors.

Question 3: Would it be possible to compare the robustness given the same level of Imperceptibility (which may be achievable by adjusting tau)?

Thanks for your constructive suggestion! In the following table, we show the full pareto frontier of SimpleMark, obtained by adjusting τ\tau (or equivalently, SDR‾\overline{SDR}). The results show that SimpleMark retains its robustness advantage at the same level of imperceptibility: Indeed, SimpleMark with SDR‾=9dB\overline{SDR}=9dB is the only system that can achieve mean bitwise accuracy > 85% and MOS > 4. Even within the tightest quality budget, our system still surpasses AlignMark, the strongest high-quality baseline, by 16.04pp on bitwise accuracy under voice cloning attacks. The pareto frontier conparison between SimpleMark and AWARE in Fig. 3 also demonstrates that our system achieves significantly better robustness–imperceptibility trade-off, rather than re-balancing on the old pareto frontier of existing methods.

MethodSDR‾\overline{SDR} (dB)No Atk (%)Edits (%)Codecs (%)Cloning (%)AI (%)Mean (%)PESQSTOIViSQOLNISQAMOS
AudioSeal100.0091.4650.7450.2058.7962.804.4220.9984.3054.0064.27 ± 0.75
WavMark100.0092.8354.5449.8757.8263.774.1870.9973.6864.0543.66 ± 1.06
Timbre†100.0099.7260.3969.7986.0678.993.7880.9924.5913.7633.67 ± 1.01
AWARE699.9998.5062.8251.5987.5075.104.2390.9854.8413.6934.38 ± 0.75
VoiceMark98.8287.0982.9391.6381.9585.902.1110.9102.9084.0622.37 ± 0.75
AlignMark97.6888.0184.9668.1388.3482.362.9750.9474.4114.0663.97 ± 0.99
SimpleMark699.9098.0984.9992.9587.8390.973.9340.9604.0213.7763.54 ± 0.83
SimpleMark799.7896.4783.9291.3787.1189.724.0590.9684.1373.8503.84 ± 0.81
SimpleMark999.2493.9381.2488.0585.6187.214.2440.9794.3103.9394.24 ± 0.70
SimpleMark1198.0090.5278.2284.1783.7684.174.3570.9874.4343.9834.43 ± 0.64

Question 4: Why does changing the detector representation affect imperceptibility in Tables 2 and 3 if the detector only retrieves the watermark from generated audio?

Thank you for raising this point. We would like to clarify that although weights of the representation extractors are frozen, gradients from the watermark detection loss still pass through the representation extractors and update the scale predictor during training. Therefore, changing the detector representation changes the training signal received by the watermark embedder, leading it to produce a different scaling pattern. Though these scaling factors are bounded within the same range, their actual strength still slightly differs. Therefore, quality of watermarked audio produced by these different embedders also slightly varies.

Thank you for the response. Could you please clarify further: In the table for Q3, is the SDR the average computed over the entire test set? A few rows are missing the SDR. Could you please show them?

AuthorsOpenReview ↗

Feel free to tell us if you have further questions

Thanks for your timely response, which pointed out a crucial ambiguity in notation. We actually meant to use SDR‾\underline{SDR} rather than SDR‾\overline{SDR}, which denotes the lower bound of per-bin SDR. It is actually τ\tau under decibel scale (see eq. (6)), a tunable hyperparameter consistently applied to all test utterances. We have revised our manuscript to replace SDR‾\overline{SDR} with τdB\tau_{dB} to avoid misunderstanding. Furthermore, we only report SDR‾\overline{SDR} / τdB\tau_{dB} of AWARE and our proposed system since they are the only two systems that support per-bin SDR control.

Thank you again for pointing out this issue! Feel free to tell us if you have further questions.

Thank you for the response and the additional clarification. My concerns have been adequately addressed. I have also read the authors’ responses to the other reviewers, and I am happy to raise my score to 3.

Reviewer_BT7FOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper argues that most mechanisms used in robust multi-bit audio watermarking (attack simulation, latent-space embedding, perceptual losses) are unnecessary, and proposes SimpleMark, which embeds the message by bounded multiplicative scaling of STFT magnitudes (imperceptibility controlled by a hard SDR bound) and decodes it with a linear head over a frozen low-level audio representation after attentive statistics pooling, trained with only a detection loss and no simulated attacks. Across RAW-Bench edits, 30 neural codecs, seven zero-shot voice-cloning systems and DeepMark AI attacks, SimpleMark reports state-of-the-art bit accuracy under codec resynthesis and voice cloning while remaining competitive elsewhere, with nine predictor/representation combinations behaving consistently. Ablations show that the representation level matters, that attack simulation adds nothing and perceptual losses hurt robustness, and that the learned watermark is a content-agnostic frequency pattern shared across frames.

Critical strengths
  1. The position of this paper is easy to follow.
  2. The authors identify and verify the actual mechanism (a fixed, content-agnostic spectral pattern) rather than leaving it implicit.
  3. The subtractive study is a useful clarification for the field: it shows that codec- and cloning-robust watermarking does not require attack simulation, latent embedding, or perceptual losses.
Critical weaknesses
  1. The claim of “competitive imperceptibility” is contradicted by the paper’s own numbers: SimpleMark’s MOS (3.78) is second-lowest among the seven systems and its PESQ trails AudioSeal, WavMark and AWARE, yet Table 1 compares robustness at unmatched quality. The only Pareto comparison (Fig. 3) covers one baseline and one attack, and the MOS study has 10 listeners. Since the robustness–imperceptibility trade-off is the central question, the evidence for a better trade-off is insufficient.
  2. The watermark is a fixed, content-agnostic frequency pattern, and the implications are not addressed. Security-wise, a non-adaptive attacker can estimate the pattern by averaging several watermarked clips and remove or forge it, a known weakness of spread-spectrum-style marks that the threat model sidesteps. Mechanistically, the pattern rides on the spectral envelope/timbre, which both the speaker-encoder detector and voice-cloning systems preserve—so “cloning robustness” may amount to the clone copying an altered timbre, and false-positive rates on unwatermarked (especially same-speaker) audio are not reported in the main text. The conceptual distinction from Timbre, which shows similar tonal artifacts in Fig. 4, is also unclear.
Questions for authors
  1. What is the false-positive / existence-detection rate on unwatermarked audio, including unwatermarked audio from the same speakers?
  2. If an attacker averages the spectra of several watermarked clips and applies a notch/equalization filter, does the watermark survive?
  3. Is cloning robustness due to the clone reproducing the altered timbre? How does SimpleMark behave with cloning systems that transfer prosody/content but not timbre?
4 comments

We sincerely appreciate your constructive feedback and insightful comments. Below we provide our answers to your questions.

Question 1: What is the false-positive / existence-detection rate on unwatermarked audio, including unwatermarked audio from the same speakers?

Thanks for pointing out this issue. First of all, we used paired watermarked / unwatermarked audio for calculating binary accuracy: whenever there is a watermarked audio, there will be paired unwatermarked source audio with exactly the same content and speaker. That said, we provide FPR (%) of our system under the 75% threshold and compare it with the baselines.

MethodNo AtkCodecsCloningAI
Timbre3.363.556.183.24
AWARE3.633.904.013.84
VoiceMark3.213.474.043.82
AlignMark4.083.923.523.78
SimpleMark3.893.803.773.57

The results show that compared with the baselines, SimpleMark has relatively low FPR. Additionally, the following table lists FPR (%) of the SimpleMark instances in Tab. 2. These results indicate that the FPRs of all SimpleMark instances lay around 3.84%, the theoretical FPR under random guessing.

Scale PredictorRepresentationNo AtkCodecsCloningAI
LinearLog-Mel4.313.893.493.93
LinearQWen3-ECAPA3.893.803.773.57
LinearSpeechTokenizer3.473.893.613.94
Gated Conv.Log-Mel3.783.903.063.66
Gated Conv.QWen3-ECAPA3.243.703.863.70
Gated Conv.SpeechTokenizer4.473.903.214.36
ConformerLog-Mel4.164.093.814.26
ConformerQWen3-ECAPA3.363.803.553.87
ConformerSpeechTokenizer4.583.904.264.04

Question 2: If an attacker averages the spectra of several watermarked clips and applies a notch/equalization filter, does the watermark survive?

Thank you for this insightful question. We find that existing anti-cloning audio watermarking models, including SimpleMark, are generally vulnerable under equalization-based watermark removal attacks. To the best of our knowledge, this kind of attack has not been reported in other literature.

In this kind of attack, the attacker wants to remove the watermark from a watermarked audio sample$A$. Let MA′M'_A and PA′P'_A denote the STFT amplitude spectrogram of AA, and μt\mu_t the mean across time frames. The attacked STFT amplitude is produced by dividing the original spectrogram with per-bin average, or formally, M~A=MA′/μt(MA′)\widetilde M_A=M'_A/\mu_t(M'_A). Then the attacked waveform is produced by A~=iSTFT⁡(M~A,PA′)\widetilde A=\operatorname{iSTFT}(\widetilde M_A, P'_A). We introduce three metrics to evaluate whether the attack is effective:

  1. Bitwise Accuracy (%): Bitwise accuracy after attack. Lower indicates more effective attacks.

  2. Attack Successful Rate (ASR, %): Fraction of samples predicted as unwatermarked after attack.

  3. PESQ: Attacked audio quality. An attack should not be considered successful if it significantly damages audio quality.

Below we report attack results on both baselines and our proposed system:

MethodBit Acc. (%)↓ASR (%)↑PESQ↑
AudioSeal99.79%0.0%3.47
WavMark99.68%0.0%3.44
Timbre55.93%89.9%3.27
AWARE99.9%0.0%3.49
VoiceMark54.45%90.3%3.61
AlignMark94.61%1.3%3.48
SimpleMark55.32%90.5%3.18

The results show that all three systems designed for anti-cloning audio watermarking, namely Timbre, VoiceMark and SimpleMark, are susceptible to this kind of equalization-based watermark removal attack. This indicates that watermarks produced by these systems may share the same nature of time-invariant, content agnostic frequency pattern. However, this kind of attack significantly degrades watermarked audio quality. It is therefore questionable whether this kind of attack should be considered practical in reality.

(Continued in the next comment...)

Question 3: Is SimpleMark fragile against adaptive watermark forgery attacks?

Similar to the attack in Question 2, we can derive a similar equalization-based watermark forgery attack. In this kind of attack, the attacker wants to transfer the watermark from a watermarked audio sample$A$ to an unwatermarked sample BB. Let MA′M'_A and MBM_B denote the STFT amplitude spectrogram of AA and BB, respectively. The attacked STFT amplitude of B is then produced by M~B′=MB⋅μt(MA′)μt(MB)\widetilde M_B'=M_B\cdot\frac{\mu_t(M'_A)}{\mu_t(M_B)}. Metrics are similar to those introduced in Question 2, except that higher bitwise accuracy indicates more effective attacks here. The table below lists the attack results on both baselines and our proposed system:

MethodBit Acc. (%)↑ASR (%)↑PESQ↑
AudioSeal49.85%1.2%3.22
WavMark50.14%1.0%3.22
Timbre92.63%95.8%3.12
AWARE49.78%0.2%3.16
VoiceMark80.36%59.2%3.16
AlignMark54.32%3.1%3.23
SimpleMark95.02%97.9%3.10

The results are similar as in Question 2: anti-cloning audio watermarking models are generally vulnerable under equalization-based watermark forgery attacks. A minor difference is that the forgery attack is less effective on VoiceMark, indicating more variation of watermark patterns. However, the forgery attack also damages audio quality. Additionally, the attacker can only transfer the watermark from watermarked to unwatermarked audio, but has no access to the actual payload or control over payload message.

Question 4: Is cloning robustness due to the clone reproducing the altered timbre? How does SimpleMark behave with cloning systems that transfer prosody/content but not timbre?

First, we do not think that our watermark alters timbre. Instead, we regard it as a global style that voice cloning models unintentionally transfer along with timbre to synthesized speech. As for the second question, we cannot find a publicly-available voice cloning model that transfers only prosody or content but not timbre from the speech prompt to synthesized speech: Content generally comes from text (TTS) or source speech (VC), whereas prosody (F0) is usually generated to fit specific target content. Therefore, we only consider voice cloning attacks where the voice cloning system transfers timbre.

Question 5: Does SimpleMark achieve competitive imperceptibility or better robustness–imperceptibility trade-off?

Yes. An advantage of SimpleMark is that one can tune the threshold τ (or equivalently, SDR‾\overline{SDR}) to explicitly balance between robustness and imperceiptibility. In the main experiment, we choose SDR‾=6dB\overline{SDR}=6dB to show that SimpleMark can achieve state-of-the-art robustness against neural codec resynthesis and voice cloning attacks. The following table shows that with SDR‾=11dB\overline{SDR}=11dB, our system can also achieve state-of-the-art imperceiptibility while maintaining **versatile robustness **against neural codecs, voice cloning and AI attacks: Even within the tightest quality budget, our system still surpasses AlignMark, the strongest high-fidelity baseline, by 16.04pp on bitwise accuracy under voice cloning attacks.

As for robustness–imperceptibility trade-off, we would like to clarify that AWARE is the only baseline that supports explicit robustness–imperceptibility control. Therefore, we can only compare our proposed system with AWARE in Fig. 3, where our system achieves significantly better robustness–imperceptibility trade-off. The full pareto frontier of SimpleMark, as shown in the table below, confirms its superior robustness–imperceptibility trade-off: Indeed, SimpleMark with SDR‾=9dB\overline{SDR}=9dB is the only system that can achieve mean bitwise accuracy > 85% and MOS > 4.

MethodSDR‾\overline{SDR} (dB)No AtkEditsCodecsCloningAIMeanPESQSTOIViSQOLNISQAMOS
AudioSeal100.0091.4650.7450.2058.7962.804.4220.9984.3054.0064.27 ± 0.75
WavMark100.0092.8354.5449.8757.8263.774.1870.9973.6864.0543.66 ± 1.06
Timbre†100.0099.7260.3969.7986.0678.993.7880.9924.5913.7633.67 ± 1.01
AWARE699.9998.5062.8251.5987.5075.104.2390.9854.8413.6934.38 ± 0.75
VoiceMark98.8287.0982.9391.6381.9585.902.1110.9102.9084.0622.37 ± 0.75
AlignMark97.6888.0184.9668.1388.3482.362.9750.9474.4114.0663.97 ± 0.99
SimpleMark699.9098.0984.9992.9587.8390.973.9340.9604.0213.7763.54 ± 0.83
SimpleMark799.7896.4783.9291.3787.1189.724.0590.9684.1373.8503.84 ± 0.81
SimpleMark999.2493.9381.2488.0585.6187.214.2440.9794.3103.9394.24 ± 0.70
SimpleMark1198.0090.5278.2284.1783.7684.174.3570.9874.4343.9834.43 ± 0.64

(Continued in the next comment...)

Question 6: What is the conceptual distinction between SimpleMark and Timbre, given their similar tonal artifacts in Fig. 4?

We would like to respectfully clarify that although we found some shared watermark pattern between SimpleMark and Timbre, this by no means indicate that they share the same model architecture: Timbre detects watermark in STFT amplitude spectrogram, controls imperceptibility with perceptual loss, and enhances robustness with simulated attack (Griffin-Lim). In contrast, SimpleMark controls imperceptibility with bounded magnitude scaling and enhances robustness by detecting watermark from low-level audio representations. Even though a gated-convolutional network adapted from Timbre is included as one choice of scale predictor, our experiment shows that choice of the scale predictor hardly affects robustness or imperceiptibility. That said, the only thing that SimpleMark and Timbre share in common is that they embed watermark in STFT amplitude spectrogram.

Thanks for the rebuttal. I am raising the score to 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.