acceptodds
Under review as a conference paper at ICLR 2027

VIGIL: Variance-Informed Latent Guidance for Private Time-Series Generation

Abstract

Hosting time series diffusion models on third-party servers offers ubiquitous client access. But sharing raw time series with servers raises privacy concerns. Prior art enables clients to work around by sharing latent embeddings instead. But such methods are inapplicable to conditional tasks such as imputation and forecasting, since the constraints are specified in the client's data space, which is inaccessible to the server. We propose VIGIL, a novel framework that enables privacy-preserving time-series diffusion by conditioning entirely in the latent space, with no observable time-series values leaving the client's premises. VIGIL has three phases: (1) latent anchor optimisation finds a plausible latent reference consistent with the observed data; (2) the anchor confidence map identifies the anchor positions corresponding to observed and missing positions in the data space; and (3) uses the anchor and the confidence map to steer generation via latent-guided sampling towards the constraints induced by the observed values. Results on five datasets spanning window lengths from 64 to 168 time steps demonstrate that VIGIL matches the performance of approaches that directly operate on raw data. Despite guiding in the latent space, VIGIL even improves on its data-space counterpart's similarity scores by 1.3% on average across 20 evaluation settings. Code available at https://anonymous.4open.science/r/Synthetic-Time-Series-Generation/

Primary area:
generative models
Keywords:
time-series imputation, diffusion models, latent diffusion, conditional generation, privacy-preserving machine learning

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

This paper suggests VIGIL framework doing imputation and forecasting by using server-hosted time-series diffusion models without prohibiting the client’s data to be exposed. Existing latent-based privacy-preserving methods have limitations that they do not preserve the relative positions between data and representation space. To address this limitation, VIGIL optimizes a latent anchor on the client side by minimizing the reconstruction error on observed entries and constructs a confidence map based on the variance of latent representations induced by changes in missing values. And then, the server uses this anthor and confidence map to adjust the latent at each reverse step on diffusion models and performs conditional generation without any training or decoder access. They achieved good performance on five multivariate time-series datasets and the effect of the confidence map was also confirmed through the ablation study.

Critical strengths
  1. A simple, training-free latent mask for non-local encoders. The variance-based confidence map needs only forward passes of the client’s encoder and works for any encoder.
  2. Empirical results. For five datasets, they achieve competitive performance in the latent space diffusion setting despite reconstruction errors.
  3. The writing is easy to follow.
Critical weaknesses
  1. Novelty comparing with the latent inpainting method. Data Guidance is a combination of normalised gradients, multiple corrections, and dynamic guidance-strength schedule for projectionless MPGD [1]. Therefore, the “(ours)” notation in Table 1 is excessive.
  2. The motivation of the “server” is weak. The authors stated that this method is valid for heavy computation assignments (Appendix H.4). However, if there is a risk of data leakage, wouldn't it be best to use a different method like TEE and simply perform conditional training, as shown in Table 1, to achieve the best performance?
  3. Asymmetry of tuning effort. VIGIL performs 200 corrections per reverse step, while Data Guidance and Spinning Decoder use 10 corrections with "limited pre-tuning." In the data space, the authors say that convergence occurs quickly, and there is no repetition count ablation.

[1] He et al., Manifold Preserving Guided Diffusion, in ICLR 2024.

Questions for authors
  1. In the end, what VIGIL means is a very narrow domain. It is a situation where you cannot use TEEs, they are unreliable, and even running the client yourself is difficult for CSDI, but the autoencoder can still learn. Does such an environment really exist?

  2. In diffusion models, the global structure is determined in the early stage when noise is high, and the later stage is the refinement of details. But, the schedule of the paper imposes the weakest constraints early on when the global structure is defined, and the strongest constraints at the end. Requests η(t) and K(t) ablation.

And please see the weaknesses.

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

Cloud-based time-series forecasting often requires uploading raw data, raising privacy concerns. Latent-space generation can protect raw data, but it is difficult to identify which latent components correspond to observed values, especially when missing values are filled before encoding. To address this, the authors propose VIGIL, which identifies reliable latent components for conditional generation. The method first constructs a latent anchor, then adds random noise to missing positions and uses latent variance to estimate confidence. Low-variance components receive stronger guidance during cloud-based diffusion sampling. Experiments on five datasets across four imputation and forecasting scenarios evaluate CRPS, MSE, different guidance strategies, and computational costs. The results show that VIGIL is generally effective.

Critical strengths

The proposed method has practical significance for addressing real-world user privacy concerns, and it is also relatively easy to implement.

Critical weaknesses

The paper estimates the reliability of latent variables based on their variance after adding random Gaussian noise to the missing positions. However, there is insufficient direct evidence that low-variance latent variables actually contain more reliable information from the observations. In addition, the authors do not systematically investigate how noise intensity, perturbation distributions, and the number of random samples affect confidence estimation and final generation performance. Therefore, the validity and robustness of the proposed method require further verification.

Questions for authors
  1. The paper uses the variance of latent components under random perturbations to estimate their reliability. However, high variance only indicates sensitivity to missing inputs and does not necessarily mean that the information is unreliable. Could the authors provide stronger theoretical justification or direct experimental evidence that low-variance latent components contain more reliable information from the observations?

  2. How sensitive are the confidence maps and final generation performance to the noise intensity, perturbation distribution, and number of random samples? Could the authors provide additional experiments to evaluate the robustness of the proposed method?

  3. Some high-variance latent components may still contain important information from the observations. Could reducing their guidance weights lead to the loss of useful information? In particular, Full Guidance outperforms the proposed method in some settings in Table 2. Could the authors further investigate and explain this phenomenon?

  4. The paper uses linear interpolation to construct the initial latent anchor. Have the authors considered other imputation methods? How would different initialization strategies affect the anchor quality, confidence estimation, and final generation performance?

  5. The current experiments only consider a 50% missing ratio. Could the authors evaluate the method under different missing ratios and forecasting horizons, especially in more challenging settings, to demonstrate its generalization ability?

  6. The experiments use a fixed latent representation of 16×128 and a Transformer autoencoder. How would different latent dimensions, compression ratios, or encoder architectures affect confidence estimation and generation quality? Could the authors provide additional ablation studies?

  7. The method retains only the top 20% of latent components with the highest confidence and strongly downweights the remaining components. What motivates this choice? How sensitive is the generation performance to the retention ratio and the weight attenuation factor?

  8. Could the authors include a simpler baseline that directly fills the missing values, encodes the completed sequence into the latent space, and then decodes it without diffusion-based guidance? This comparison would help clarify the necessity of the additional diffusion and latent guidance mechanisms

Reviewer_SiU1OpenReview ↗
Rating
2(−1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates privacy-preserving conditional time-series generation in a setting where a server-hosted diffusion model cannot access the client's raw observations. Its main contribution is an inference-time mechanism that translates data-space observation constraints into latent-space guidance, avoiding the need for repeated client–server interactions during sampling. Specifically, the method optimizes a latent anchor to preserve observed values and uses perturbation-based variance estimates to distinguish latent components that are relatively insensitive to missing inputs. These estimates determine how strongly different components are constrained during reverse diffusion. The framework is evaluated on five multivariate time-series datasets under four missingness scenarios.The results suggest that effective conditional time-series generation can be achieved through inference-time guidance in a compressed latent space, without explicitly preserving the correspondence between latent components and original observation positions. They also highlight the potential of sensitivity-based guidance to balance observation consistency and generative flexibility, while reducing the computational and communication costs of conditional sampling.

Critical strengths
  1. The paper addresses a employment problem: enabling conditional time-series generation without sharing raw observations with a third-party server. The motivation is clear, and the proposed client–server setting is practically meaningful.

  2. The variance-based guidance is an interesting design choice. Rather than assuming a direct correspondence between data-space and latent-space positions, the method estimates which latent components are less affected by missing values and uses them to guide generation.

Critical weaknesses
  1. The paper is motivated by privacy concerns, yet its evaluation relies entirely on public benchmark datasets, including air quality, traffic, and energy consumption data. While these datasets are suitable for assessing generation quality, they do not fully reflect the privacy-sensitive scenarios motivating the work. Evaluation on healthcare or other sensitive time-series data, ideally accompanied by an empirical privacy analysis, would make the practical significance of the proposed framework more convincing.

  2. All experiments use 50% missingness. Results at other missingness ratios would help establish how robust the method is when very little or nearly all of the sequence is observed.

Questions for authors
  1. Given the paper’s privacy motivation, can the authors provide evidence that VIGIL is effective in privacy-sensitive applications, such as healthcare, and assess whether the transmitted latent anchors and confidence maps could reveal sensitive information under the stated threat model?

  2. How does VIGIL perform at missingness ratios other than 50%, particularly when very few or nearly all values are observed? Can the authors provide results across several ratios to establish whether its advantages over the baselines remain consistent?

  3. Have the authors evaluated how accurately the final generated samples preserve the observed values? Table 6 (a) reports observation reconstruction errors for the latent anchors, but it is unclear whether the same consistency is maintained after the full reverse diffusion process.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.