acceptodds
Under review as a conference paper at ICLR 2027

PETS: Inference-Time Differentially Private Synthetic Time Series Generation

Abstract

Existing methods for differentially private (DP) synthetic time series generation inject privacy during model training via DP-SGD, requiring private data in the training phase, expensive hyperparameter tuning, and costly retraining for new domains. We propose Private Evolution for Time Series (PETS), the first inference-time framework for DP synthetic time series generation via Private Evolution (PE). In this setting, private data is not used to train generative models, but only to guide the selection of synthetic outputs at inference time, improving distributional fidelity while satisfying a privacy-budget constraint. PETS is modular by design, enabling modular adaptation without training on private data. Building on top of PE, we instantiate three concrete schemes, ChatTS+VAE, Diffusion-TS, and TimeCraft, each combining a generation module, a variation module, and a shared contrastive embedding for similarity-driven selection. Empirically, PETS achieves a mean lower C-FID than state-of-the-art (SOTA) DP methods across four public–private dataset pairs over privacy budgets ( to , plus ), with gains up to at . For downstream tasks, PETS achieves a mean the activity classification accuracy of SOTA DP methods over these privacy budgets, showing PETS produces high-quality DP synthetic time series with strong utility-privacy trade-offs.

Primary area:
learning on time series and dynamical systems
Keywords:
synthetic data, time series, differential privacy, machine learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ehbyOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposed a method that makes fake time-series data that looks like a private dataset without leaking private information. It doesn't train a new model on the private data. It reuses a public model in a loop: the public model makes fake samples, each private sample "votes" for the fake sample it looks most like, noise is added to the votes to protect privacy, and the winners are kept and slightly changed. After a few rounds, the fake data looks more like the real data. On five types of data (traffic, energy, activity, heart signals and crypto), the authors report better results than other privacy-safe methods, about 3.4× higher classification accuracy than DP-SGD at ε = 1, and sometimes better quality than models trained directly on the real data.

Critical strengths

S1. A practical idea. Making private fake data using only a public model, with no training on private data, is attractive for hospitals and other regulated settings. It is also cheap to run.

S2. A broad evaluation. Five domains, public models that match the private data to different degrees, and four kinds of downstream tasks are more than most papers in this area test.

Critical weaknesses

W1. The privacy proof's own assumption is broken. Theorem 2 assumes the "synthetic population sizes" are "fixed independently of the private dataset." But Appendix F.3 sets the number of fake samples to sf × |S_priv,train|, meaning it depends on the size of the private dataset. So the theorem doesn't cover the method as actually run.

W2. Private data is used outside the privacy budget.

  • Appendix F.2 says the model settings for traffic, activity and crypto were chosen using "a 6-iteration PE audit on the activity training split." The activity private set is UCI-HAR, and the same sentence explicitly calls the ECG audit "public," so the activity audit appears to use private data.
  • Appendix F.3 says the ECG settings were retuned "by the C-FID of PETS runs at ε = 1," which means picking settings based on quality scores computed on the private data.
  • Data scaling uses a "shared public+private percentile clip and min–max scaling" (Section 4), which also reads private data.

None of these uses is counted in the privacy budget, so the real ε is higher than reported.

W3. The classification headline looks broken. The paper says the classification test uses each method's generated labels. Yet every baseline that generates samples without labels (dp-GAN, GS-WGAN, Pub2Priv, TimeDiff, TimeGAN) scores exactly 20.6% at every privacy level, including the non-private ones. That looks like these methods have no meaningful labels, so the classifier always predicts one class. If so, the "3.4× better than DP-SGD" claim is a comparison against a broken setup.

W4. Results are cherry-picked. The paper reports the best round for each metric and the best of three variants for each cell. Using the final output (what a user would actually get), MMD wins drop from 12/12 to 7/12, and wins in the mismatched-domain settings drop from 6/7 to 3/7. Choosing the best round using private-data scores is also another privacy leak (see W2).

W5. The method is graded with the tool it optimizes. PETS picks samples by similarity in the TS2Vec embedding space, and its main quality score (C-FID) is computed in that same space, so it naturally looks good there. This likely explains why PETS beats non-private models trained on the real data by up to 6.3× in C-FID.

Questions for authors
  1. How is the number of fake samples chosen without knowing the private dataset size? Please restate Theorem 2 for the method as actually run, or rerun with a fixed number.
  2. Was the "activity training split" in F.2 private UCI-HAR data? How do you account for tuning on private data (F.2, F.3), the shared public+private scaling, and picking the best round per metric?
  3. How do label-free methods get labels for the classification test? Why do they all score exactly 20.6%, even without privacy?
  4. Can you report all results on the final output only, with one fixed variant chosen in advance, and C-FID in a different embedding than the one used for voting?
  5. Why is GS-WGAN's C-FID the same at every privacy level?
Reviewer_4HcAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper adapts Private Evolution (PE), the inference-time DP synthetic-data framework introduced for images and text (DPSDA), to time series. Private windows are never used to train any model; they enter only through a DP nearest-neighbor voting histogram (Gaussian noise, μ-GDP composition over T iterations, Appendix B) that selects which synthetic candidates to keep and perturb. The contribution is the set of time-series generation/variation/similarity modules that make PE work: (i) ChatTS rule-based generation with a public-data Conv1D VAE as variation operator, (ii) Diffusion-TS with SDEdit-style partial-noising variation and optional classifier guidance, (iii) TimeCraft (TimeDP + TarDiff) with SDEdit variation and domain/class conditioning; TS2Vec embeddings define the voting distance. Experiments use four public/private pairs (PeMS-BAY→METR-LA, BTC→ETH, WISDM→UCI-HAR, MIT-BIH→INCART), 13 privacy budgets (ε = 0.125…32, ∞), six DP baselines (dp-GAN, GS-WGAN, GlucoSynth, Pub2Priv, DPSDA-image, DPSDA-text) and two non-DP baselines, five fidelity metrics (C-FID in the TS2Vec voting space plus MOMENT-FID, DAE-FID, MMD² on raw windows, ACD), four TSTR downstream tasks, a seven-cohort public–private domain-mismatch study, hyperparameter ablations, and a comparison between best-iteration and released-population (S_T) results. Headline claims: mean 6.0× lower C-FID than the strongest DP baseline, 1.4× the activity classification accuracy of the strongest DP baseline, and 11–216× faster per-iteration runtime than DPSDA adaptations. Code is provided.

Critical strengths
  1. Breadth and honesty of the empirical evaluation. Four domain pairs × 13 budgets × 8 baselines × 4 downstream tasks, plus three evaluation spaces that are not the voting space (MOMENT-FID, DAE-FID, MMD²; Table 5, Appendix H) with reported noise floors, a seven-cohort public–private mismatch study (Table 6), and—importantly—an explicit comparison between the per-metric best iteration used in the main tables and the actually released population S_T (Appendix J, Tables 20–24), where the C-FID/FID_M/FID_DAE wins over DP baselines persist. This is consequential because the main risk with PE-style methods is that gains are an artifact of the voting metric or of non-private iteration selection; the paper goes further than most in ruling that out.

  2. Practical value within the DP-time-series niche. Training-free (w.r.t. private data) DP synthesis with modular public-data components, per-iteration GPU cost 11–216× below DPSDA adaptations (Fig. 12), near-zero utility cost from ε = ∞ to ε = 1 on forecasting/imputation (Fig. 6c, Fig. 15), and released code with pinned third-party versions. If the DP-boundary issues below are fixed, this is a usable baseline for the community.

Critical weaknesses
  1. The formal (ε, δ)-DP claim for the reported numbers does not hold as stated, because private data enters through paths outside the accounted histogram mechanism (Sec. 4 "Datasets", App. F.1, F.3, J; code pe/runner/pe.py lines 84–92). (a) Preprocessing applies "a shared public+private percentile clip and min–max scaling" (Sec. 4; F.1), so the clipping range is an un-noised function of the private set. (b) In the conditional setting the per-class synthetic population sizes are derived from the private label counts without noise: main.py does not pass fraction_per_label_id, so pe/runner/pe.py falls back to priv_data...value_counts() (the code even logs "Assuming the fraction of label ids in private data is public information"); Algorithm 1 instead states an equal Nsyn/|C| split, so paper and code disagree. (c) |S_syn| = sf × |S_priv,train| (F.3) ties the output size to the private set size under add/remove adjacency. (d) Every PETS/DPSDA number in Tables 3–6 is the per-metric best evaluated iteration t ≥ 1 selected using the private reference (App. J), and the ECG checkpoints and PE hyperparameters were selected by private C-FID at ε = 1 (F.3). The authors are transparent about (d) and show S_T results in Appendix J, which largely preserve the conclusions—but the main text should report S_T, (a) should use public-only clipping, and (b) should use a DP label histogram or a fixed public split. Each item is individually small; together they are consequential for a paper whose selling point is a clean inference-time privacy boundary, and the Ethics Statement's one-sentence caveat does not substitute for fixing them. This drove Soundness = 2 and kept the rating at weak rather than clear accept.

  2. Baseline credibility and single-seed reporting limit how much of the 6.0× / 3.4× headline can be trusted (Sec. 4 "Baselines", Table 4, App. F.5–F.6). dp-GAN, Pub2Priv and GlucoSynth are author re-implementations (no public code); GS-WGAN reshapes windows into square images; DPSDA-image operates on 64×64 line-plot renderings with an ImageNet diffusion model. In Table 4 (Activity), dp-GAN, GS-WGAN, Pub2Priv, TimeDiff and TimeGAN all report exactly 20.6% accuracy at every ε including ε = ∞—five methods, four budgets, one number—which indicates a degenerate TSTR setup for unconditional generators (a single-class collapse), not a measurement of those methods. The abstract's "3.4× the activity classification accuracy of DP-SGD methods" is therefore computed against a collapsed baseline and should be removed or re-measured; the "1.4×" against conditional DPSDA is the defensible comparison. Relatedly, Pub2Priv beats PETS at ε = ∞ on traffic and crypto C-FID (Table 11: 0.4×, 0.7×) yet loses by 5–10× at finite ε, suggesting the DP-SGD baseline may be under-tuned under noise. All results use seed 0 with no run-to-run variance (F.3, F.8); PE is stochastic (Gaussian noise + resampling), so the ablation's "within ∼6%" robustness claim (F.4) and the smaller margins (e.g., ECG 1.5–1.6×) are not yet distinguishable from noise.

Questions for authors

Clarifications that could change my assessment:

  1. Please confirm or correct each item in Critical Weakness 1: is the percentile clip computed from public+private windows? Are per-class synthetic sizes taken from private label counts (as the code does) or split equally (as Algorithm 1 states)? Would you commit to reporting S_T numbers in the main text and recomputing with public-only clipping?
  2. What produces the constant 20.6% activity accuracy for dp-GAN/GS-WGAN/Pub2Priv/TimeDiff/TimeGAN at all ε? If these generators are unconditional and labels were assigned trivially, please re-run them with class-conditional training (or per-class training as done for DPSDA) and update the "3.4×" claim accordingly.
  3. Can you provide run-to-run variability (≥3 seeds) for at least Table 3 at ε ∈ {0.5, 1} and for the Table 8 ablation, and bootstrap CIs for the downstream metrics?
  4. Why does Pub2Priv outperform PETS at ε = ∞ on traffic and crypto C-FID but degrade 5–10× at ε ≤ 2? Was its noise multiplier / clipping tuned per budget?
  5. Table 3 reports activity TimeCraft C-FID = 0.33 at ε = 1, while Table 8's deployed-style settings give 0.424 (Nnn = 7) with default 0.514; what is the lookahead degree k used for the non-ECG deployed configuration (F.3 lists k only for ECG), and which configuration produced 0.33?

Suggestions (not affecting the rating): report the private PE-train sizes for all four pairs in the main text; discuss the gap between window-level DP with 50%-overlapping windows and user-level protection more prominently than the ethics statement.

Reviewer_YMS9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PETS, an inference-time framework for differentially private synthetic time-series generation. Instead of training generative models on private data, PETS uses pretrained models and DP nearest-neighbor voting to iteratively refine synthetic samples. The authors explore three implementations using VAE and diffusion-based models. Experiments on four datasets show substantial improvements in fidelity over existing DP baselines, along with better utility on several downstream tasks.

Critical strengths

The paper addresses an important problem. Avoiding DP training on private data is appealing, especially when pretrained models can be reused across domains.

The experimental results are promising. The authors evaluate several privacy budgets, datasets, and downstream tasks. The improvements across multiple fidelity metrics are encouraging.

Critical weaknesses

Privacy guarantee: The preprocessing uses percentile statistics computed from both public and private data (Appendix F.1), while the privacy analysis assumes preprocessed inputs. It is unclear whether the claimed DP guarantee covers the entire pipeline. Similarly, the privacy implications of hyperparameter selection in Appendix F.2 need clarification.

Evaluation protocol: The main results report the best iteration separately for each metric, although the actual released dataset comes from the final iteration. This may overstate practical performance. Appendix J partly addresses this, but final-iteration results should be emphasized in the main evaluation.

Contribution of individual components: Since PETS relies on strong pretrained models, particularly TimeCraft, it is difficult to tell how much of the improvement comes from the proposed framework versus the quality of the public-data priors. More controlled ablations would help.

Generality: The experiments focus on univariate, fixed-length windows. It remains unclear how well the approach extends to multivariate time series or settings with limited relevant public data. The window-level privacy guarantee also does not directly provide user-level privacy.

Questions for authors
  1. Does the DP guarantee cover preprocessing using private-data statistics? How is the privacy cost of hyperparameter selection accounted for?

  2. Can the authors report final-iteration results as the primary evaluation rather than selecting the best iteration for each metric?

  3. Can the authors better isolate the contribution of PETS from that of the pretrained generative models?

  4. How would the method handle multivariate time series and user-level privacy?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.