- 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
- 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.
- The layout format does not follow the ICLR template. The authors reduced the margins of the paper.
Questions for authors
None.
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 ΔFID 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 ΔFID 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≈1) 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=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=std1/1000mean−mean1, threshold at the 99.9th percentile → 0.1% target FPR. Revision: restate in Sec. 4.1, relabel the Tables 8/13 rows (≈0.5 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; 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 ∣Wf∣Re(Wf)>α1−α,else1/2; at exactly α=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>21, 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}, 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)