acceptodds
Under review as a conference paper at ICLR 2027

PhaseLock: Phase-Structured Tokenization for Watermarking Periodic Time Series

Abstract

Generation-time watermarks for time series bias a discrete token sequence toward a secret set of codes and detect the watermark by re-encoding the observed signal. Existing methods tokenize windows starting at arbitrary positions and treat the codebook as an unstructured set, so patches at different phases of the period are represented and marked in the same way. The bias can then favor codes used at other phases, and the detector does not know the phase of the observed window. We take the phase of a periodic series as the organizing principle for the codebook, the watermark, and the detector. PhaseLock partitions the codebook into phase groups and quantizes each patch using the group corresponding to its phase. The tokenizer and generator are trained only on windows aligned to the patch grid of the period. PhaseLock outputs each generated window with a random temporal shift, allowing samples to start at arbitrary times as in real data, and it injects the watermark with a phase-gated bias, so that each phase is marked with its own codes. At detection, a single re-encoding network recovers the grid alignment, phase, and codes directly from the observed signal. On ten periodic benchmarks, PhaseLock remains detectable under post-editing attacks at a lower Context-FID than the same generator with an unstructured codebook of equal size. It also yields fewer false positives under wrong keys and preserves the periodicity of the data more closely than existing watermarks.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
time series generation, watermarking, generation-time watermark, structured vector quantization, periodicity

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_Eyz9OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents PhaseLock, a generation-time watermarking framework for periodic time series. The method is motivated by the observation that vector-quantized (VQ) codebooks exhibit phase-dependent token usage patterns, even without explicit phase supervision. Existing watermarking approaches generally treat the codebook as an unstructured set, potentially introducing watermark biases that conflict with the periodic characteristics of the data.

PhaseLock partitions the codebook into phase-specific groups and restricts token quantization to the group corresponding to each patch's phase. During generation, a phase-gated logit bias is applied to green tokens within the current phase group, and the generated window is randomly shifted to accommodate arbitrary patch-grid alignments. During detection, a phase-aware re-encoding network estimates the patch-grid alignment, phase, and token identities before computing the watermark detection statistic.

The authors evaluate PhaseLock on ten periodic time-series benchmarks, comparing it with LVQMark, TimeWak, Tree-Ring, and Gaussian Shading. The experiments consider offset, crop, and insert attacks, as well as sample quality, wrong-key false positives, and preservation of periodicity. The reported results show improved sample quality at the evaluated detection operating points, fewer wrong-key false positives, and better preservation of periodic structure in most datasets.

Critical strengths

S1. A well-motivated and coherent phase-aware watermarking framework.

The paper identifies a meaningful connection between the periodic structure of time-series data and the organization of VQ codebooks. The observation that different codes are preferentially used at different phases provides a reasonable motivation for phase-partitioned tokenization and phase-gated watermark embedding. The proposed framework consistently incorporates phase information into tokenization, watermark insertion, and detection, making the overall design conceptually coherent.

S2. Comprehensive comparisons within the considered setting, supported by targeted ablations.

The experiments cover ten periodic datasets and compare PhaseLock with several existing watermarking methods. The evaluation considers watermark detectability, sample quality, false positives under incorrect keys, and periodicity preservation. Importantly, the authors compare PhaseLock with LVQMark at both matched and larger codebook sizes, and provide an ablation of phase partitioning. The additional analyses of watermark strength, training-seed robustness, and phase recovery further strengthen the empirical study.

Critical weaknesses

W1. The detection evaluation relies on large-scale aggregation and does not establish the effectiveness of watermark detection with limited observations.

A major concern is the scale at which watermark detection is evaluated. According to Appendix B.5, the detection protocol samples 1,000 windows from a pool of 10,000 generated windows, computes an aggregate green-ratio statistic, and repeats this procedure 100 times to estimate the true positive rate (TPR). Consequently, the reported TPR measures detection over large collections of generated windows rather than individual time-series samples.

This distinction is important because each output window contains only 15–16 or 31–32 scored tokens, depending on the dataset and alignment. Aggregating 1,000 windows provides substantially more statistical evidence than would be available when verifying a single generated series or a small collection of samples.

Although the experimental protocol is explicitly documented, the paper does not investigate how detection performance varies with the number of available windows. Therefore, the reported TPR of 1.00 under the evaluated conditions does not establish that the watermark remains reliably detectable when only a limited number of generated samples are available.

This limitation affects the practical interpretation of the detection results and makes it difficult to assess the applicability of PhaseLock to provenance verification scenarios with smaller observation budgets.

W2. The robustness evaluation is strongly tied to known attack families and attack-matched statistical calibration, limiting the evidence for generalization.

The robustness evaluation considers three post-editing attacks: offset, crop, and insert. However, the corresponding operations primarily modify data values while preserving the length of the time-series window. In particular, the offset attack applies an additive value shift, the crop attack masks part of the observed window, and the insert attack replaces existing time steps with random values rather than inserting new time steps.

Therefore, although PhaseLock explicitly incorporates patch-grid alignment and phase recovery, the evaluated attacks do not directly establish robustness against more general temporal desynchronization, such as deletion or insertion of time steps, changes in sampling rate, or temporal stretching. The authors acknowledge that length-changing edits are not evaluated, but this remains an important limitation given the emphasis on alignment recovery.

Furthermore, Appendix A states that the detection statistic is standardized using non-watermarked data under the same attack and detector. The re-encoding network is also trained using the same attack families considered during evaluation. Consequently, the reported robustness is demonstrated under an attack-aware training and statistical calibration protocol.

It remains unclear how the method would perform when the attack type is unknown, when clean-data calibration statistics are used, or when the observed transformations differ from those encountered during training.

These limitations do not invalidate the reported results within the evaluated setting. However, they substantially restrict the strength of the evidence supporting the broader robustness claims.

Questions for authors

Q1. Detection performance under different observation budgets

Appendix B.5 evaluates watermark detection by aggregating 1,000 windows. Could the authors clarify whether PhaseLock is intended primarily for collection-level provenance verification?

How does the detection performance change when only 10, or 100 windows are available? Reporting TPR at a fixed false positive rate across different observation budgets would help establish the practical detection requirements of the proposed method.

Q2. Robustness to temporal desynchronization

The proposed detector explicitly recovers patch-grid alignment and phase, but the evaluated offset, crop, and insert attacks preserve the window length.

Could the authors clarify the expected behavior under actual temporal editing operations, such as deleting or inserting time steps, temporal stretching, or changes in sampling rate? In particular, does the phase-aware re-encoding mechanism remain effective when the temporal correspondence between the generated and observed series is altered beyond the random sub-patch shift considered in the current experiments?

Q3. Attack-matched calibration and unseen attacks

Appendix A states that the detection statistic is standardized using non-watermarked data under the same attack and detector. Does the detection procedure therefore require knowledge of the attack type or an attack-matched non-watermarked reference set?

How would detection performance change if the null-distribution statistics were calibrated only on clean data and then applied to attacked samples? Additionally, are the samples used for statistical calibration independent of those used to evaluate the false positive rate?

These clarifications would help determine whether the method can be deployed without prior knowledge of the transformations applied to the observed data.

Q4. Selection of watermark strength

Section 4.1 and Appendix B.5 define the operating watermark strength δ⋆\delta^\star as the smallest strength achieving TPR = 1.00 under the clean condition and all three evaluated attacks.

Were these operating points selected using a separate validation set, or were they determined directly from the evaluation pools reported in the main experiments?

5 comments

Thank you for the detailed rebuttal and additional experiments. I appreciate the authors' efforts to address my concerns and provide further experimental evidence.

The new results on smaller observation budgets, temporal editing attacks, and held-out generation seeds have substantially addressed my previous concerns. However, one point regarding statistical calibration remains unclear. Although the authors state that the detector does not require knowledge of the attack type, the paper describes using non-watermarked data under the same attack for standardization. Could the authors clarify whether the reported results rely on attack-matched calibration statistics and whether the calibration and false-positive evaluation samples are independent? Overall, I appreciate the additional experiments and clarifications provided by the authors.

AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer for the kind follow-up, and we're glad the new experiments addressed the earlier concerns. Both remaining questions were covered in [W2, Q2 & Q3 - 1] of our earlier response, but the answer there sat between two large tables, and we apologize for not stating it more clearly. We restate it below in a more direct form, with the same two tables for reference.


[1] Do the reported results rely on attack-matched calibration statistics?

No. As the reviewer points out, the sentence in Appendix A, "The score zz is standardized by the mean and standard deviation of g^\hat{g} on non-watermarked windows under the same attack and the same detector," is not worded in line with how the detector is run. In every reported result, μ0\mu_0 and σ0\sigma_0 are estimated once per dataset from the clean non-watermarked pool, and the same two numbers standardize every row, clean and attacked. The detector never receives the attack type or an attack-matched reference set.

We will replace that sentence in Appendix A with "The score zz is standardized by the mean and standard deviation of g^\hat{g} on clean non-watermarked windows scored by the same detector, and the same two values are used for every condition, clean and attacked."

This can be checked in the supplementary code. get_zscore in Utils/watermark_metrics.py (line 482) computes μ0,σ0\mu_0, \sigma_0 from its first argument, the null pool. In cli/evaluate.py (--mode zscore, lines 701 to 736) that argument is always the pool read from the clean detect-wo directory, for the clean row (line 709), the attacked watermarked rows (line 733), and the attacked non-watermarked rows (line 736). No attacked pool is ever used to estimate the null.


[2] Are the calibration and false-positive evaluation samples independent?

In the paper, the calibration windows and the windows used for the false positive rate come from the same generation run. The null statistics, however, hardly depend on the run. The table gives μ0\mu_0 / σ0\sigma_0 for PhaseLock at δ⋆\delta^\star, one row per generation seed of Table 1, and they differ across seeds by at most 0.004 on every dataset.

SeedETThETTmAirQ-1AirQ-2AirQ-3AirQ-4AirQ-5AirQ-6EnergyWeather
123450.503 / 0.1910.497 / 0.1140.498 / 0.1620.501 / 0.1510.498 / 0.1420.501 / 0.1410.501 / 0.1470.501 / 0.1410.501 / 0.0980.497 / 0.095
123460.503 / 0.1920.497 / 0.1170.498 / 0.1620.501 / 0.1490.497 / 0.1440.501 / 0.1400.499 / 0.1480.499 / 0.1410.501 / 0.1000.499 / 0.094
123470.504 / 0.1920.500 / 0.1130.499 / 0.1610.500 / 0.1480.500 / 0.1430.502 / 0.1400.501 / 0.1480.497 / 0.1420.499 / 0.0990.499 / 0.096

To make the two fully independent, we took μ0,σ0\mu_0, \sigma_0 from the clean non-watermarked pool of one generation seed and applied them to the watermarked and non-watermarked pools of the other two seeds. Cells give the minimum TPR / maximum non-watermarked FPR over clean, offset, crop, and insert, averaged over the two evaluated seeds.

Calibration → evaluationETThETTmAirQ-1AirQ-2AirQ-3AirQ-4AirQ-5AirQ-6EnergyWeather
12345 → 12346, 123471.0 / 0.011.0 / 0.021.0 / 0.011.0 / 0.001.0 / 0.021.0 / 0.001.0 / 0.021.0 / 0.001.0 / 0.011.0 / 0.02
12346 → 12345, 123471.0 / 0.001.0 / 0.011.0 / 0.011.0 / 0.001.0 / 0.011.0 / 0.011.0 / 0.031.0 / 0.021.0 / 0.011.0 / 0.01
12347 → 12345, 123461.0 / 0.001.0 / 0.001.0 / 0.011.0 / 0.011.0 / 0.001.0 / 0.011.0 / 0.001.0 / 0.051.0 / 0.011.0 / 0.01

With an independent calibration, TPR is 1.00 in every cell and the false positive rate stays at or below 0.05.


We hope this resolves the remaining point, and we thank the reviewer again for the careful reading, which led us to clarify the description of the calibration.

Thank you for the further clarification of the calibration procedure and the additional experimental results.

The authors have clarified that the same clean, non-watermarked calibration statistics are used across all conditions. The additional cross-seed experiments also offer useful evidence regarding detection performance with independent calibration and evaluation samples. These explanations help clarify the evaluation protocol and provide further context for interpreting the reported results.

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

Based on the discovery that different phases of a cycle often correspond to distinct local patterns within periodic time series, the paper proposes PhaseLock, which utilizes phase to redesign tokenizer, watermark injection and detection based on the previous work LVQMark. A considerable number of experiments show that PhaseLock reduces distribution distortion in generated data while still maintains well capacity of watermark detection.

Critical strengths
  1. Clear motivation. As shown in fig 1, a standard VQ tokenizer already shows phase-dependent token usage without explicit phase supervision. This suggests that the learned discrete representation implicitly captures the periodic structure of the data. It is a pretty reasonable motivation for later design.

  2. The method is fairly simple and intuitive (fig 2). As mentioned above, restricting the watermark bias to the current phase group seems like a natural way to reduce interference with the periodic structure of the generated data. It does not require changing the generator as well.

  3. Solid evaluation(tab 1, tab 2, tab 4 etc.). The paper provides a fairly thorough experiment across multiple datasets, including comparisons with existing methods, robustness tests under different attacks, and wrong-key experiments.

Critical weaknesses
  1. Limited additional gains from phase partitioning. tab 3 shows that much of the Context-FID improvement already comes from grid-aligned training and random shifts, even without phase partitioning. Adding the partition further improves Context-FID, but the additional gain is relatively modest compared to the improvement already achieved without it.

  2. The method relies quite heavily on having a known and stable period(appendix D.1, fig 6, fig 7). This seems reasonable for the datasets considered, but I'm less sure how it would handle time series with multiple periodicities or weaker periodic patterns. The Exchange results also suggest that the method may not generalize well to non-periodic data.

Questions for authors

The contribution of the phase-aware design to generation quality could be discussed more clearly. As shown in tab 3, the version without phase partitioning already achieves a substantial improvement in Context-FID over LVQMark, suggesting that grid-aligned training and random shifts also play an important role. It would be better to clarify this distinction when discussing improvements.

3 comments

Thank you for the detailed rebuttal. I appreciate the authors' responses. The additional analysis of watermark retention and code recovery helps clarify the contribution of phase partitioning to detection performance.

However, I have a new concern regarding the motivation and the main source of improvement. From tab 1 the most noticeable advantage of PhaseLock over LVQMark appears to be its generation quality. However, the rebuttal clarifies that phase partitioning mainly contributes to detection, while most of the generation quality improvement comes from grid-aligned training and random shifts. In fact, for periodic time series, simply fixing the patch boundaries may already allow recurring local patterns to be consistently tokenized, making them easier for the tokenizer to learn, even without explicit phase partitioning. Tab 3 shows that grid alignment & random shifts already produces substantial quality improvements without phase partitioning. This raises a question about the paper's central motivation: if the main generation quality gains do not come from explicitly exploiting phase-dependent token usage, but instead from a more general patch alignment strategy, to what extent do the results actually validate the proposed phase-based motivation, rather than showing that phase partitioning mainly compensates for the detection limitations introduced by the other design choices? Given this, I prefer to maintain my original score.

AuthorsOpenReview ↗

Rebuttal by Authors

We sincerely thank the reviewer for reading our response closely and for the follow-up. We are glad that the retention and code-recovery analysis clarified what the phase partition contributes to detection. We also apologize that our one-line summary separated the two components more sharply than the method does.


Do the results validate the phase-based motivation?

We thank the reviewer for stating this concern so precisely, and we agree with the observation behind it. Most of the Context-FID gain in Table 3 is already there with grid-aligned training, and the reviewer's explanation is the right one: for periodic series, fixed patch boundaries allow "recurring local patterns to be consistently tokenized."

Where we see it differently is in what this gain says about the motivation.

1. The patterns recur only because the patch boundaries lie on the patch grid of the period.

Take ETTh, hourly data with a daily cycle, where a patch covers 6 hours (τ=24\tau = 24, s=6s = 6).

  • With windows that start at every time step, a patch can cover hours 0 to 6, 1 to 7, 2 to 8, and so on: 24 different slices of the day. Figure 1 shows that a plain VQ tokenizer uses different codes for them.
  • With windows aligned to the patch grid of the period, a patch covers only hours 0 to 6, 6 to 12, 12 to 18, or 18 to 24: P=τ/s=4P = \tau/s = 4 slices (Section 3.1, Eq. 2).

The same codebook then describes four slices of the day instead of twenty-four. This is the consistency the reviewer describes, so the quality gain of grid-aligned training is direct evidence for the phase-based motivation.

2. Grid-aligned training and the phase partition are the two places where one method uses the period.

LVQMarkPhaseLock
Period of the datanot usedused at every stage
Tokenizer trainingwindows starting at every time stepwindows aligned to the patch grid of the period (grid-aligned training)
Codebook and keyunstructured codebookphase groups with a phase-gated bias (phase partition)
Detectorre-encodes the codes directlyrecovers the alignment and the phase, then reads the codes (phase partition)

Table 3 therefore measures how the roles are divided inside one phase-based design. It does not compare a phase-based idea with a phase-free one.

  • Grid-aligned training brings the phase structure into the tokenizer. This shows mainly in sample quality.
  • The phase partition brings it into the key and the detector, where a structure that is only implicit in the tokenizer cannot be used. This shows mainly in detection.

3. With the phase partition, more of the mark survives detection, so a weaker mark is enough.

Our previous table showed the first half on Energy: PhaseLock retains 70% of the embedded mark, against about 20% without the partition and under 40% for LVQMark, which uses neither component.

This returns to sample quality. At the point where each configuration reaches TPR 1.00 in every condition (δ⋆\delta^\star), PhaseLock carries a weaker embedded mark (gg, the green ratio of the emitted tokens) and has the lower Context-FID.

DatasetConfigurationgg at δ⋆\delta^\starC-FID at δ⋆\delta^\star
EnergyPhaseLock w/o partition0.6340.099
EnergyPhaseLock0.5720.070
AirQ-1PhaseLock w/o partition0.6010.032
AirQ-1PhaseLock0.5600.025

To show what lies behind these numbers, we ran the retention analysis on AirQ-1 across watermark strengths (δ\delta from 6 to 12). The full table and its reading are in the next comment.

AuthorsOpenReview ↗

Rebuttal by Authors (Cont')

3 (continued). Strength sweep on AirQ-1 (K=512K = 512)

Methodδ\deltaggg^\hat{g}RetentionTPR clean / offset / crop / insertC-FID
LVQMark60.5680.5210.300.98 / 0.99 / 1.00 / 0.960.090
LVQMark80.6280.5370.291.00 / 1.00 / 1.00 / 1.000.109
LVQMark100.7080.5560.271.00 / 1.00 / 1.00 / 1.000.136
LVQMark120.7830.5720.251.00 / 1.00 / 1.00 / 1.000.171
PhaseLock w/o partition60.5070.5030.500.01 / 0.03 / 0.04 / 0.000.020
PhaseLock w/o partition80.5150.5080.530.15 / 0.09 / 0.24 / 0.070.021
PhaseLock w/o partition100.5400.5220.541.00 / 0.97 / 0.98 / 0.880.027
PhaseLock w/o partition120.6010.5540.541.00 / 1.00 / 1.00 / 1.000.032
PhaseLock60.5100.5080.770.09 / 0.09 / 0.24 / 0.090.021
PhaseLock80.5260.5180.720.76 / 0.73 / 0.94 / 0.630.022
PhaseLock100.5600.5410.681.00 / 1.00 / 1.00 / 1.000.025
PhaseLock120.6310.5860.651.00 / 1.00 / 1.00 / 1.000.038

gg, g^\hat{g}, and retention are defined as in our previous table (green ratio of the emitted tokens, green ratio the detector reads back, and the share of the mark that survives). Bold δ\delta is δ⋆\delta^\star, the smallest strength with TPR 1.00 in all four conditions. Attacks are at strength 0.30, and all values are means over three generation seeds.

The three configurations are different models, and the same δ\delta does not embed the same mark in each (the gg column). We therefore read each configuration along its own sweep and compare them at their own δ⋆\delta^\star.

  • Retention. Over its sweep, PhaseLock retains 0.65 to 0.77 of the mark it embeds. The configuration without the partition retains 0.50 to 0.54, and LVQMark 0.25 to 0.30.
  • Mark needed for detection. At δ⋆\delta^\star the three detectors read back a similar green ratio (g^\hat{g} from 0.537 to 0.554). PhaseLock gets there from the weakest embedded mark (g=0.560g = 0.560, against 0.601 without the partition and 0.628 for LVQMark).
  • Quality at the point of detection. PhaseLock has the lowest Context-FID at δ⋆\delta^\star (0.025, against 0.032 and 0.109).

A watermark needs both at once: samples of high quality, and a mark that survives to detection. In PhaseLock both come from the same source, the phase of the series.


We hope this makes clear how the two components relate, and we thank the reviewer again for a question that helped us explain the method better.

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

This paper introduces PhaseLock, a VQ-based watermarking method that associates phase with its own set of codes. Instead of using static green tokens, the detector utilizes this pattern and recovers the phase of an observed window from the signal to locate green tokens. Experiments compare the proposed method with LVQMark and show advantages of the proposed method in signal distortion while maintained the robustness under 30% attack conditions.

Critical strengths

The proposed method makes sense and the idea is motivated by the observation in Figure 1 which is quite interesting. The experiment results are convincing (though I have some concerns on datasets which is explained in the next section). The ablation study shows clear benefit for the design.

Critical weaknesses

The scope is limited: (1) Periodic signal only: while the paper appears to support the claim well, the subject is limited to non-audio time series data and improvement upon LVQMark is marginal. On the contrary, the baseline LVQMark has broader scope as periodicity is not a requirement for the method to function. The datasets used in the experiments are also chosen to be periodic - a reader may be interested to see if a dataset has weak or no periodicity, can the proposed method still achieve a reasonable strength in robustness while achieving low signal distortion. (2) Non-universal method: the proposed method requires training one model per dataset. It is also a limitation for LVQMark but because LVQMark is not limited to periods, one can potential combine datasets to make a more general codebook. On the contrary, audio watermarking methods are typically universal; you can train a model on speech music sounds and mixtures, and use it for any type of audio in applicaiton.

Coming from an audio watermarking background, I do see an interesting knowledge gap between audio and non-audio time series research. (1) To make the paper more readable for a wide range of readers, it would be great to have a few sentences explain the why watermarking timeseries is important. I can see ownership tracing being an application, but I also worry that the distortion (e.g. 0.04 corr introduced) would harm the usability of the data. It would be great to motivate the application in the intro. (2) audio watermark can deal with defend against tempering even at local level, say one second of samples among minutes. And by math 0.04 corr translates to about 16dB SNR while most modern audio watermark can achieve a 20dB and higher SNR while being robust to a variety of attacks. I can see a world where audio watermarking being used for timeseries since audio is already a type of time series. The paper could have a broader contribution if it can associate audio watermarking in the same context - just a suggestion, not a criticism.

Questions for authors

I am curious if you can argue for broader scope and applications. Please share your thoughts and views.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.