acceptodds
Under review as a conference paper at ICLR 2027

FreqWak: Lightweight Inversion-Free Frequency-Domain Watermarking For Time Series Data

Abstract

Synthetic time series produced by diffusion models make it possible to share privacy-sensitive datasets, such as patients' functional MRI records. Important criteria for synthetic data include high data utility and traceability for verifying the data source. Existing state-of-the-art methods (e.g., TimeWak) rely on costly diffusion inversion, incurring approximation errors and yielding low bit accuracy (55%–96%), and suffering from low detection efficiency. We propose FreqWak, to the best of our knowledge, the first lightweight frequency-domain watermarking framework that eliminates diffusion inversion and achieves over 99.9% bit accuracy under our evaluation settings. FreqWak embeds watermarks by modulating the phase of the residual frequency band in the post-sampling time series. Detection requires only a single rFFT, enabling highly accurate and inversion-free verification without modifying the diffusion training process. We extensively evaluate FreqWak across five datasets and multiple baselines, in terms of its impact on synthetic data quality, watermark detectability, and robustness under various post-editing attacks. Our results show that FreqWak achieves nearly 100% bit accuracy across all tested lengths, with detection 19–1,256× faster than TimeWak on 10,000 samples. Against the strongest SOTA baseline, FreqWak improves context-FID by 51.26% and correlational scores by 92.67%, while keeping the synthetic data quality statistically indistinguishable from the unwatermarked baseline. Moreover, owing to the phase invariance of non-DC components, FreqWak is provably immune to offset attacks under linear transformations, with a formal proof. Overall, FreqWak achieves competitive performance across all evaluated benchmarks and attack configurations. Our code is available as supplementary material.

Primary area:
generative models
Keywords:
Watermarking, Time Series, Diffusion Models, Frequency-Domain Analysis

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_E9DvOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes FreqWak, a post-sampling watermarking method for multivariate time series generated by diffusion models. After standard DDIM sampling, it shifts the phase of the high frequency components toward a key dependent template. Detection requires only a single rFFT and a phase alignment check, without using diffusion inversion. They report nearly 100% bit accuracy and fast detection against TimeWak while maintaining high synthetic data quality.

Critical strengths

Detection is cheap and does not require diffusion inversion.

Critical weaknesses
  1. The first two sentences in the abstract use the same content and structure as TimeWak's abstract. The evaluation sentence in the abstract is also nearly identical. In section 2.1, the descriptions of Tree Ring, Stable Signature and Tabwak are copied from TimeWak section 2. In section 4.4, the attack definitions are copied from TimeWak as well.
  2. The layout format does not follow the ICLR template. The authors reduced the margins of the paper.
Questions for authors

None.

Reviewer_M8UmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces FreqWak, a post-generation watermarking approach for synthetic time-series data. Rather than modifying the diffusion process or recovering its initial noise through inversion, FreqWak embeds a keyed signature by adjusting the phases of selected high-frequency components. Watermark detection then relies on Fourier-domain phase alignment. The authors provide theoretical results concerning detection reliability, signal distortion, and invariance to positive affine transformations. Experiments on five datasets compare FreqWak with existing watermarking methods in terms of detection accuracy, computational efficiency, data quality, and resistance to post-processing attacks. The reported results show substantial detection speedups and nearly perfect phase-alignment accuracy under unmodified conditions.

Critical strengths

S1: FreqWak is computationally attractive. Moving watermark embedding outside the diffusion process removes the need for expensive inversion during verification. The algorithm is relatively simple, does not require retraining the generator, and could be incorporated into existing time-series generation pipelines with limited modifications.

S2: The experimental coverage is extensive, spanning five datasets, different sequence lengths, and several post-editing attacks. The analysis of frequency-band selection and injection strength is useful for understanding the method's behavior. The invariance to constant offsets is also a desirable property, particularly for time-series data that may undergo normalization or rescaling.

Critical weaknesses

W1: I have difficulty reconciling the reported data-quality improvements with several results elsewhere in the manuscript. Some comparisons rely on unwatermarked reference values from the original TimeWak study rather than those obtained through the authors' reproduction. More importantly, the discriminative scores indicate that watermarked samples become substantially easier to distinguish from real data. These findings weaken the claim that FreqWak preserves the quality of the generated time series.

W2: The detection results need further justification. The default decision threshold does not establish a controlled false-positive rate, and the theoretical reliability guarantee does not fully account for the constraints imposed by real-valued Fourier reconstruction. The robustness evaluation also overlooks transformations that directly affect the fixed-frequency phase signature. Consequently, the results do not yet establish the claimed reliability in a practical verification setting.

Questions for authors
  1. The quality results appear to use inconsistent reference values. For Stocks, Table 1 reports an unwatermarked Context-FID of 0.258, whereas Appendix Table 6 gives 0.106 from the authors' reproduction. FreqWak achieves 0.135, which is actually worse than the reproduced unwatermarked result. There are further discrepancies across Tables 11, 13, and 15. How were these results obtained, and why were values from different evaluation pipelines combined? A consistent comparison using the same generator, evaluation procedure, and random seeds is necessary to substantiate the reported quality gains.
  2. Table 1 discriminative scores are concerning. FreqWak obtains values around 0.49–0.50 across all datasets, compared with substantially lower scores for unwatermarked samples. Under the stated metric, this means the watermarked data are much easier to distinguish from real data. The explanation that the discriminator detects the injected phase pattern does not resolve the issue, since the purpose of this metric is to assess synthetic-data fidelity. The predictive scores also deteriorate on MuJoCo and Energy. How do the authors reconcile these results with the claim of negligible quality degradation?
  3. Algorithm 2 sets the detection threshold to 0.5, which is also approximately the expected phase-alignment accuracy of an unwatermarked signal against an independent random template. Such a threshold would ordinarily produce a substantial false-positive rate. However, Fig. 2 reports detection at 0.1% FPR without explaining the corresponding threshold selection. Appendix Tables 8 and 13 also label values near 0.5 as false positives, although they appear to represent phase-alignment accuracy. Please specify how the threshold was calibrated and report the actual sample-level false-positive rates.
  4. Theorem 1 guarantees perfect alignment for α ≥ 0.5, yet Table 11 reports an accuracy of 0.929 at α = 0.5. There is also a potential implementation issue at the Nyquist frequency: the selected injection band includes k = L/2 for even sequence lengths, where the Fourier coefficient must be real, while Eq. (1) may assign it a complex phase. The irFFT operation cannot preserve that arbitrary imaginary component. This could explain why the theoretical guarantee does not always hold after reconstruction. The relationship between the theorem, the implemented transform, and the reported ablation results needs to be established.
  5. I am concerned that the robustness tests miss a natural weakness of the method. Because detection compares phases at fixed frequency indices, even a small temporal shift changes the phase alignment without necessarily altering the underlying signal information. A one-sample circular shift, mild smoothing, or temporal resampling would be particularly informative tests. The existing results already show substantial vulnerability to aggressive cropping. It would be useful to see whether FreqWak remains detectable under these simpler transformations, using TPR at a controlled false-positive rate rather than Z-scores alone.
  6. The technical distinction from existing post-processing watermarking methods is not sufficiently developed. Embedding keyed patterns through Fourier-phase modifications is an established signal-watermarking strategy, and the paper already includes HTW, another post-generation method that does not require diffusion inversion. FreqWak may still offer a useful adaptation to multivariate time series, but its specific methodological novelty remains unclear. A direct comparison with a conventional frequency-domain watermark, using identical data and distortion settings, would help show what is gained by the proposed phase-blending design.
  7. The detection statistics and assumptions behind the theoretical analysis need closer examination. For ETTh with window size 64, Appendix Table 8 reports phase-alignment accuracy of 0.504 after 30% cropping, while Table 9 gives a Z-score of −99.74 for what appears to be the same setting. Several Z-score signs also differ between Table 1 and the attack tables. Please explain the Z-score definition and calibration procedure. In addition, the imperceptibility analysis assumes that the selected residual band contains less than 5% of the spectral energy, but the manuscript does not establish this proportion for each dataset and sequence length. Reporting these measurements would make the theoretical bound more meaningful.
4 comments
AuthorsOpenReview ↗

Our reply to reviewer M8Um ( from Question 1 to Question 5)

We very much thank the reviewer for the serious reading and professional suggestions.

Since current research mostly uses diffusion models to generate time series and watermarks, our study focuses on diffusion models and aims to address the inability of existing time-series watermarking methods to simultaneously achieve high detectability (100% bit accuracy), imperceptibility, strong robustness, and lightweight, low-cost, efficient detection. In particular, for watermark detection accuracy (Bit Accuracy), almost all diffusion-generated time-series watermarks have poor Bit Accuracy (BA). For example, as shown in Table 2 of our paper, the current best method, TimeWak, generally achieves only 60%–80% BA, whereas our FreqWak can achieve 100% Bit Accuracy and has very high detection efficiency.This is of practical significance for watermark detection in large-scale time series, especially those with longer sequence lengths. We address the reviewer's concerns , from Question 1 (Q1) to Question 7 (Q7) in turn.

To Q1 (reference values).

To make a fair comparison with TimeWak, we used all the baselines which are same with Timewak. And the revision will state each row's provenance: TR/GS/HTW/TabWak/TabWak$^\top$/TimeWak rows come from the TimeWak paper; W/O and FreqWak rows are our own reproduction (Stocks W/O 0.106), because Timewak is an excellent paper and Timewak' data can be verified. So do we. The Freqwak's abstract claim will be restated watermarked-vs-watermarked (Stocks: FreqWak 0.135 vs TimeWak 0.277) plus per-dataset Δ\DeltaFID vs our reproduced W/O. The revision will also state each row's provenance: TR/GS/HTW/TabWak/TabWak$^\top$/TimeWak rows come from the TimeWak paper; W/O and FreqWak rows are our own reproduction (Stocks W/O 0.106). The abstract claim is restated watermarked-vs-watermarked (Stocks: FreqWak 0.135 vs TimeWak 0.277) plus per-dataset Δ\DeltaFID vs our reproduced W/O.

To Q2 (discriminative score).

We isolated the cause. Neutralization control: embedding a random, non-keyed phase pattern into unwatermarked samples at matched detection strength (per-sample BA≈1BA\approx1) reproduces the full discriminative gap — control 0.497/0.500/0.497 vs watermarked 0.495/0.497/0.500 on MuJoCo/Energy/fMRI (3 runs) — while unwatermarked samples sit at 0.01–0.14. The gap is thus entirely the deliberate high-frequency carrier modification; the generator itself is unaffected. Corrections from archived samples: Stocks/ETTh disc near W/O (0.07/0.23 vs 0.07/0.09); predictive (MuJoCo 0.007→0.010, Energy 0.253→0.302); Energy corr 0.145 not reproducible (fresh 0.94/1.34/0.89 vs W/O 1.299).

To Q3 (threshold & FPR).

τ\tau=0.5 in Alg. 2 is illustrative, not the decision rule. Detection uses the calibrated statistic: per-window BA vs the no-watermark reference (mean1, std1), Z=mean−mean1std1/1000\frac{\text{mean}-\text{mean1}}{\text{std1}/\sqrt{1000}}, threshold at the 99.9th percentile → 0.1% target FPR. Revision: restate in Sec. 4.1, relabel the Tables 8/13 rows (≈0.5\approx0.5 values are calibration BAs, not FPR), and report measured sample-level FPR: 0.000–0.001 across all 30 configurations.

To Q4 (Thm 1 vs α=0.5\alpha=0.5; Nyquist).

Your diagnosis is right. The implemented band includes the Nyquist bin k=L/2, constrained to be real under rFFT — irFFT discards the injected imaginary part, so the template phase cannot be enforced there and Thm 1 does not cover it. Per feature f, that bin aligns with probability 1 if ∣Re⁡(Wf)∣Wf∣∣>1−αα,else1/2;\left|\frac{\operatorname{Re}(W_f)}{|W_f|}\right| > \frac{1-\alpha}{\alpha}, else 1/2; at exactly α\alpha=0.5 it is 1/2 regardless, giving expected BA=(n−1+0.5)/n=6.5/7=0.929 — precisely Table 11. Revision: Nyquist excluded from injection and detection — re-verified: clean BA remains 0.998–1.000, [email protected]%FPR = 1.0, all 30 configurations; Thm 1 restated for α>12\alpha>\frac{1}{2}, where detection is deterministic.

To Q5 (shift / smoothing / resampling).

You are right that fixed-index phase detection is inherently shift-sensitive: a 1-sample circular shift drops BA to 0.09–0.14 (TPR 0 at 0.1% FPR) on all datasets. New appendix (5 datasets × 2 window sizes × 3 runs): (i) smoothing and downsampling attacks (moving-average, brick-wall low-pass, 5%/30%) — mild smoothing preserves detection; 30% low-pass at ws=24 degrades BA to 0.27–0.65; (ii) a shift-search detector (s∈{−2,…,2}s \in \{-2,\dots,2\}, constant-factor cost) restores BA 0.998–1.000, TPR 1.0. Temporal resampling by non-integer factors is not yet covered; it rescales the spectrum and shifts phases across bins, so we will add it. Cropping is information-theoretic — removed samples carry no watermark, so any output-domain scheme degrades (TimeWak likewise). All tables as [email protected]% FPR.

(to be continued)

AuthorsOpenReview ↗

Our reply to reviewer M8Um ( from Question 6 to Question 7)

To Q6 (post-processing distinction).

To the best of our knowledge, for large-scale multivariate time series, conventional frequency-domain watermarking methods could generally suffer from two main problems when attempting to maintain very high watermark detection accuracy (Bit Accuracy) while simultaneously achieving watermark robustness, imperceptibility, high detection efficiency, and lightweight deployment. First, they cannot effectively handle large-scale multivariate time series. Second, they cannot achieve very high watermark detection accuracy (Bit Accuracy), nor can they strike an ideal balance among robustness, detection efficiency, and lightweight deployment.

Current research mostly uses diffusion models to generate multivariate time series and watermarks. So Freqwak targets diffusion models to address their inability to achieve high detectability (100% bit accuracy), simultaneously with strong robustness, and lightweight, low-cost, efficient detection. And our contribution is not frequency-domain phase modulation per se but a deployable watermarking protocol for diffusion-generated multivariate time series.

HTW is the Heads-Tails Watermark ( reference [27] in TimeWak paper): values are pseudo-randomly assigned heads/tails via a keyed hash, then nudged by the smallest representable increment until the heads ratio matches a target — that is the value-adjustment mechanism; we will define and cite it. The HTW is a post-watermarking technique designed for univariate time series,not for multivariate time series. To extend its applicability, TimeWak have adapted it for multivariate time series by iterating through each variate in the generated synthetic sample. In short, we treat each variate as a single series when using HTW.

By contrast, although HTW achieves better quality, it performs poorly under attacks, indicating that it struggles to balance the trade-off between quality and robustness.

Direct comparison on identical samples, band- and FPR-matched: Arnold-style PM (A Phase Modulation Audio Watermarking Technique (Arnold et al., IH 2009), BPSK PN phase template, informed detection) and Ngo-style dynamic phase coding (Robust and Reliable Audio Watermarking Based on Dynamic Phase Coding and Error Control Coding (Ngo et al., EUSIPCO 2015), blind quantization to {π\pi/4, 3$\pi$/4}) both reach BA 1.000; what the phase-blending design gains: (i) 42–46% lower distortion than full quantization at equal reliability; (ii) higher [email protected]% FPR in 20/24 attack settings (9 strictly); (iii) provable normalization immunity (Thm 3) and a distortion bound (Thm 2).

To Q7 (Z-score; sign discrepancy; 5%).

Z is defined as in Q3. Z=−99.74Z=-99.74 is real (per-run calibration; recomputed 3-run mean −99.80±1.38-99.80 \pm 1.38); large negative Z under crop reflects systematic anti-alignment (truncation-induced linear phase ramp coherently flips signs), while random destruction gives Z≈0Z\approx0. Table 8's 0.504/0.534 are TimeWak's crop/insert BAs mis-entered in the FreqWak column (archived runs: FreqWak 0.358/0.594) and will be corrected. Sign differences occur only in near-chance cells (BA≈0.5BA\approx0.5): there the Z sign is sampling noise under the symmetric null; systematic settings have consistent signs — the revision reports [email protected]% FPR throughout. For the <5%<5\% assumption we measured the injection-band energy share per dataset and window size (10,000 unwatermarked samples each): Stocks/ETTh/MuJoCo/Energy 0.11–1.65%, fMRI 3.3–4.9% — all <5%<5\% (fMRI's wider residual band: 26–34%, unfitted diffusion noise); Thm 2 scales with the measured share.

We hope these clarifications and fixes address the concerns, and we are honored to receive your more encouragement.

Thank you for the rebuttal. The rebuttal clarifies the Z-score definition and the threshold calibration, and it reports the measured energy share of the injection band. It also confirms the Nyquist-frequency issue and restates Theorem 1 for α > 0.5. However, the main concerns remain.

• The rebuttal confirms that the baseline rows come from the TimeWak paper, while the unwatermarked and FreqWak rows come from the authors’ reproduction. Against the reproduced unwatermarked result on Stocks, FreqWak is worse (0.135 versus 0.106). The comparison with TimeWak still combines values from different evaluation pipelines. • The neutralization control shows that a random, non-keyed phase pattern reproduces the full gap. This confirms that the high-frequency modification makes watermarked samples easy to distinguish from real data. It does not support the claim of negligible quality degradation. The corrected Energy predictive score is also worse (0.302), and one reported correlation value could not be reproduced. • The response states that τ = 0.5 in Algorithm 2 is illustrative and that detection uses a calibrated statistic. The manuscript therefore does not describe the decision rule that was evaluated. The statistic is normalized over 1,000 windows, so the meaning of the reported sample-level false-positive rate is unclear. • The response confirms that a one-sample circular shift reduces TPR to 0 at 0.1% FPR on all datasets. The proposed shift-search detector covers only shifts of up to two samples, and its false-positive rate after the search is not reported. Temporal resampling is not covered. • The new comparison shows that conventional phase-modulation and phase-coding watermarks also reach a bit accuracy of 1.000. The remaining advantage is lower distortion and higher TPR in some attack settings. This does not establish a substantial methodological advance. • The response states that the Table 8 cropping results were TimeWak values entered in the FreqWak column. Together with the other corrections, this means several tables and Theorem 1 require revision, and no revised manuscript is available to verify them. • The statistical equivalence claim, the distinction between presence detection and multi-bit recovery, the overlapping train–test windows, and the timing comparison with a simple post-generation watermark are not addressed in the rebuttal.

Overall, the rebuttal is candid and provides useful clarifications. However, it confirms the central concerns about data quality, detection reliability, and robustness, and the required changes are substantial. The recommendation remains 2: Weak rejection.

AuthorsOpenReview ↗

Our second reply to reviewer M8Um

We very much thank the reviewer for the serious reading and professional suggestions.

To the best of our knowledge, for large-scale multivariate time series, conventional frequency-domain watermarking methods could generally suffer from two main problems when attempting to maintain very high watermark detection accuracy (Bit Accuracy) while simultaneously achieving watermark robustness, imperceptibility, high detection efficiency, and lightweight deployment. First, they cannot effectively handle large-scale multivariate time series. Second, they cannot achieve very high watermark detection accuracy (Bit Accuracy), nor can they strike an ideal balance among robustness, detection efficiency, and lightweight deployment.

Current research mostly uses diffusion models to generate multivariate time series and watermarks. So Freqwak targets diffusion models to address their inability to achieve high detectability (100% bit accuracy), simultaneously with strong robustness, and lightweight, low-cost, efficient detection. And our contribution is not frequency-domain phase modulation per se but a deployable watermarking protocol for diffusion-generated multivariate time series.

In short, to the best of our knowledge, our study is the first to propose a two-stage watermarking architecture for multivariate time series generated by diffusion models, namely “diffusion-generated sequences + frequency-domain watermark embedding/temporal watermark detection.” This architecture eliminates the inherent shortcomings of existing diffusion-model-based watermark generation and detection methods and differs from traditional phase-based watermarking for synthetic time series.

Our baselines and experimental methods are fully consistent with the current state-of-the-art TimeWak method to ensure experimental fairness. Moreover, all experimental data are verifiable, and the validity of the experimental results is very clear.

We therefore sincerely hope that the reviewers can appreciate our innovative work and experimental results and provide more encouraging review comments. Thank you.

Reviewer_kojNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FreqWak, a post-sampling watermarking method for synthetic time series. After ordinary diffusion sampling, the method blends selected high-frequency Fourier coefficients toward a keyed pseudo-random phase template. Detection uses a single rFFT followed by a phase-alignment test, without accessing the diffusion model or performing inversion. The paper provides analyses of clean-signal detection, perturbation energy, and invariance to positive affine transformations. Experiments across five datasets report high phase-alignment accuracy and substantial detection-speed improvements over TimeWak.

Critical strengths
  1. The paper is easy to follow.

  2. The paper has practical motivation and low computational overhead.

  3. The embedding and detection operations are explicitly specified and have transparent signal-processing interpretations. Besides, the paper evaluates five datasets and three sequence lengths, includes injection-strength and frequency-band ablations, and examines compatibility with an alternative MLP denoiser.

Critical weaknesses
  1. The methodological novelty beyond existing phase-based watermarking is not enough. Phase modulation and phase-spectrum encoding have been studied extensively in watermarking, including A Phase Modulation Audio Watermarking Technique (Arnold et al., IH 2009) and Robust and Reliable Audio Watermarking Based on Dynamic Phase Coding and Error Control Coding (Ngo et al., EUSIPCO 2015). These works concern audio rather than the exact setting of this paper, but they are directly relevant to its central mechanism and should be discussed. FreqWak’s inversion-free detection and backbone independence follow directly from applying watermarking to the final output rather than to the diffusion trajectory. Algorithms 1–2 do not use model-specific information after sampling. Consequently, the paper needs to explain what new embedding or detection principle is introduced, and which time-series-specific technical difficulty it resolves beyond adapting phase-based watermarking to generated sequences. The concern is not the simplicity of the method, but the insufficiently demonstrated technical advance.

  2. The large speedup over TimeWak is practically relevant, but primarily demonstrates the computational difference between direct signal-domain verification and diffusion inversion. It does not establish that FreqWak offers an advantage over other inexpensive post-processing watermarking methods. The inclusion of HTW is useful, but its value-adjustment mechanism is different from the phase-based approaches most closely related to FreqWak. A comparison with appropriately adapted phase-coding or spectral watermarking methods, using the same generated inputs and matched distortion and false-positive-rate constraints, is needed to determine whether the proposed design provides a better quality–detectability–robustness trade-off. Without such comparisons, the empirical evidence does not sufficiently distinguish the contribution of this particular method from the general benefit of moving watermarking to the output domain.

  3. Theorem 1 establishes clean-signal phase alignment because the embedding rule directly forces the relevant inner products to be positive when the injection strength exceeds one half; it does not establish reliability under general perturbations or control false positives. Theorem 3 concerns positive, feature-wise affine transformations, rather than arbitrary linear transformations. These are useful characterizations, but largely direct consequences of the construction and Fourier-transform properties.

Questions for authors
  1. What is the principal methodological advance over established phase-modulation and phase-coding watermarking? Please identify the specific difficulty introduced by synthetic multivariate time series and explain how the proposed mechanism resolves it. How should inversion-free detection and backbone independence be distinguished from properties inherited from post-processing the final output?

  2. Could you compare FreqWak with appropriately adapted conventional phase-based or spectral watermarking methods? Please use a shared set of generated samples and evaluate detection and robustness under comparable distortion budgets and fixed false-positive rates, so that the benefit of the proposed embedding rule can be isolated.

  3. Could you provide spectral-energy statistics supporting the less-than-5% assumption across datasets, features, and sequence lengths? Please also justify the Context-FID inequality in Appendix B.2 and distinguish the guarantees for clean detection and positive affine transformations from robustness to other perturbations. How should amplitude preservation be stated for the default configuration?

3 comments
AuthorsOpenReview ↗

our reply to reviewer kojN

Thank lots for the professional suggestions.
About Arnold et al. and Ngo et al. , we are honored that our ideas align with theirs which will be added as references.

Current research mostly uses diffusion models to generate time series and watermarks; So Freqwak targets diffusion models to address existing time-series watermarking methods’ inability to achieve high detectability (100% bit accuracy), simultaneously with strong robustness, and lightweight, low-cost, efficient detection. This is practically significant for watermark detection in large-scale time series, especially longer sequences. Finally, our contribution is not phase modulation per se but a deployable watermarking protocol for diffusion-generated multivariate time series, where standard audio-watermarking assumptions (perceptual masking, synchronized detection) do not hold; we address the reviewer’s concerns in turn.

To Q1(&W1). (1) Statistical, not perceptual, distortion budget. Audio schemes hide in psychoacoustic masking; synthetic series are consumed statistically (Context-FID, discriminative/predictive scores). FreqWak (a) embeds in the diffusion residual band [L/4+1,L/2][L/4{+}1, L/2] — fitted last, phases near-uniform w.r.t. the key (unwatermarked BA $\approx$0.50, Tables 8/13) — and (b) proves the budget: ∥xwm−x0∥22∥x0∥22\frac{\lVert \mathbf{x}_{\mathrm{wm}}-\mathbf{x}_0\rVert_2^2}{\lVert \mathbf{x}_0\rVert_2^2}

$\le 4\alpha^2 \cdot \frac{\sum_{k\in\mathcal{I}}\lvert \mathbf{X}[k]\rvert^2}{\sum_k \lvert \mathbf{X}[k]\rvert^2} $ (Thm 2 via Parseval; the ratio is the band’s measured energy share, <5%, yielding ≈0.098 at α=0.7), with Welch t-tests confirming statistical indistinguishability at ws≥64. (2) Calibration-capable detection: classical coding relies on synchronization codes without sample-level FPR control, whereas FreqWak’s statistic is a sign test; under H0 each sign flips with probability 1/2, so thresholds calibrate exactly for any target FPR ([email protected]% FPR, Fig. 2), as required for shared-data traceability. (3) Offset immunity and deterministic reliability: for α>1/2, cos(θ_X−θ_W) > −α/(1−α) < −1 holds for all phase offsets, making clean detection deterministic—stronger than our original measure-zero phrasing, which we will tighten. With Thm 3 (phase invariance under positive affine maps), detection survives recipient-side normalization. Inversion-free detection and backbone independence follow from output-domain processing; the above protocol-level properties do not, and are, to our knowledge, first established for diffusion-generated time series.

To Q2(&W2). We have implemented both adapted baselines — Arnold-style phase modulation (BPSK PN template phase over I\mathcal{I}, informed detection) and Ngo-style dynamic phase coding (blind quantization to {π/4,3π/4}\{\pi/4, 3\pi/4\}) — on the same samples, band, and 0.1% FPR calibration, distortion matched to FreqWak (relative L2L_2 distortion $\approx$0.058 vs 0.037; preliminary: stocks & MuJoCo, ws=64, single run, Nyquist excluded). Both attain BA 1.000 at FPR≤0.1FPR\leq0.1% — consistent with our position that the mechanism transfers; the differences are protocol-level: (i) classical full quantization needs 42–46% higher distortion for the same reliability; (ii) FreqWak's randomized complex template yields a tighter null (Z 268/415 vs 55/162) and [email protected]% FPR at least as high as both baselines in 20/24 attack settings (strictly higher in 9; exceptions: moving-average on stocks); (iii) neither provides normalization immunity (Thm 3) or distortion bounds (Thm 2). Multi-run, quality-metric, and remaining-dataset results will be included in the revision.

To Q3(&W3).

  • We agree Thms 1/3 follow from the construction; they tie design parameters (α\alpha, band) to deployment requirements, making the design auditable. Thm 3 will be restated for feature-wise positive scalar affine maps.
  • 5% energy: measured injection-band shares (10,000 unwatermarked samples each): Stocks 0.11–0.13%, ETTh 0.92–1.12%, MuJoCo 0.28–1.65%, Energy 0.30–0.46%, fMRI 3.3–4.9% — all $<$5% (fMRI's wider residual band reaches 26.5–33.9%, unfitted diffusion noise); Thm 2 scales with the measured share rather than assuming 5% (full per-sample distributions as appendix).
  • Appendix B.2: we correct expectation form: W22≤E∥Φ(xwm)−Φ(x0)∥2≤LΦ2 4α2⋅∑k∈I∣X[k]∣2∑k∣X[k]∣2W_2^2 \le \mathbb{E}\lVert\Phi(x_{\mathrm{wm}})-\Phi(x_0)\rVert^2 \le L_\Phi^2\ 4\alpha^2 \cdot \frac{\sum_{k\in\mathcal{I}}|\mathbf{X}[k]|^2}{\sum_k|\mathbf{X}[k]|^2} under the coupling xwm=x0+δx_{\text{wm}}=x_0+\delta; we will replace B.2 with this bound.
  • Amplitude: not preserved at α\alpha=0.7 (per-bin factor in [0.4,1]); the paper is already revised from magnitude preservation to bounded spectral distortion (total relative energy distortion $\leq$0.2$\alpha^{2}$).
  • Guarantees: theorems cover clean detection and positive affine maps only; the new evasion-attack appendix and the revised Limitations make robustness explicitly empirical.

we are honored to receive your encouragement and to discuss further.

Reviewer_kojNOpenReview ↗

Comments for the rebuttal

Thank you for the detailed rebuttal. The clarification of the theoretical guarantees and the new spectral-energy statistics address some of my concerns. However, two main issues remain.

  1. Contribution. The methodological advance beyond adapting existing phase-based watermarking to synthetic time series remains insufficiently clear.
  2. Baseline comparisons. The additional results are encouraging but still preliminary, with limited experimental coverage and unclear distortion matching. They do not yet convincingly establish FreqWak’s advantage over these methods. I will keep score.
AuthorsOpenReview ↗

Our second reply to reviewer kojN

We very much thank the reviewer for the serious reading and professional suggestions.

To Q1.

To the best of our knowledge, for large-scale multivariate time series, conventional frequency-domain watermarking methods could generally suffer from two main problems when attempting to maintain very high watermark detection accuracy (Bit Accuracy) while simultaneously achieving watermark robustness, imperceptibility, high detection efficiency, and lightweight deployment. First, they cannot effectively handle large-scale multivariate time series. Second, they cannot achieve very high watermark detection accuracy (Bit Accuracy), nor can they strike an ideal balance among robustness, detection efficiency, and lightweight deployment.

Current research mostly uses diffusion models to generate multivariate time series and watermarks. So Freqwak targets diffusion models to address their inability to achieve high detectability (100% bit accuracy), simultaneously with strong robustness, and lightweight, low-cost, efficient detection. And our contribution is not frequency-domain phase modulation per se but a deployable watermarking protocol for diffusion-generated multivariate time series.

In short, to the best of our knowledge, our study is the first to propose a two-stage watermarking architecture for multivariate time series generated by diffusion models, namely “diffusion-generated sequences + frequency-domain watermark embedding/temporal watermark detection.” This architecture eliminates the inherent shortcomings of existing diffusion-model-based watermark generation and detection methods and differs from traditional phase-based watermarking for synthetic time series.

To Q2.

Our baselines and experimental methods are fully consistent with the current state-of-the-art TimeWak method to ensure experimental fairness. Moreover, all experimental data are verifiable, and the validity of the experimental results is very clear.

We therefore sincerely hope that the reviewers can appreciate our innovative work and experimental results and provide more encouraging review comments. Thank you.

Comments on the paper

1 comment
Area_Chair_1dKaOpenReview ↗

Participate in the discussion period

Dear reviewers,

The authors have replied to your comments and the discussion is close to the end. We encourage you to participate in the discussion to see whether your questions and concerns have been answered by the authors. Thank you!

Best, Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.