acceptodds
Under review as a conference paper at ICLR 2027

Privacy-Aware Dirichlet Process Mixture Models: Diffusion-Based Trajectory Protection and Differentially Private Weights

Abstract

Mixture models are widely used to capture cluster structure in modern machine learning systems, particularly in time-varying applications such as smart grids and interacting large language models. However, releasing data from such systems can leak sensitive group information because the temporal trajectories of cluster centroids enable reconstruction of underlying group behavior. Static perturbation methods are insufficient, as repeated noise patterns can be filtered over time. We introduce a dynamic privacy framework based on a Dirichlet process mixture model (DDirPMM), in which centroids evolve through a diffusion process that injects temporally varying randomness to prevent trajectory reconstruction while preserving utility. In addition, our DDirPMM inherently induces static differential privacy for mixture weights, automatically protecting them without additional mechanisms. These two mechanisms address structurally distinct threat models, trajectory reconstruction and membership inference, respectively. Theoretical and empirical results demonstrate greater performance against filtering-based reconstruction and improved privacy-utility trade-offs compared with static trajectory protection methods.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Dirichlet Process, Diffusion Processes, Differential Privacy, Privacy-Preserving 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_ciqJOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces a dynamic Dirichlet process mixture model for privacy-aware synthetic-data generation. It uses time-dependent Ornstein–Uhlenbeck diffusion priors to modify centroid trajectories and derives an approximate differential privacy guarantee for posterior Dirichlet mixture weights. The method is evaluated using a synthetic temporal mixture and a smart-meter dataset.

Critical strengths
  1. The paper identifies a meaningful distinction between group-level temporal information and record-level membership information.

  2. The diffusion-based formulation provides interpretable controls over the temporal behavior of mixture centroids.

  3. The authors provide detailed mathematical derivations and an application to real electricity-consumption data.

Critical weaknesses
  1. Unclear privacy guarantees. The proposed diffusion-privacy definition only characterizes posterior mean deviations and does not establish protection against an adversarial trajectory reconstruction attack. The DP guarantee applies to a restricted randomized weight mechanism, not necessarily to the complete synthetic-data release.

  2. Potential inconsistencies in the theoretical and experimental mechanisms. The posterior predictive construction uses expected mixture weights, whereas the DP theorem analyzes sampled Dirichlet weights. Additionally, the reported ((26.2,0))-DP guarantee in the smart-meter experiment is not adequately justified by the stated theorem, which includes a potentially nonzero failure probability.

  3. Insufficient experimental validation. The evaluation relies on a one-dimensional synthetic example and one smart-meter dataset. The reconstruction attack is limited to clustering and smoothing, without stronger model-aware adversaries. The experiments also use manually designed perturbation targets, making it difficult to isolate the benefit of diffusion itself.

Questions for authors
  1. What formal reconstruction-risk guarantee follows from Us(βj)U_s(\beta_j)-diffusion privacy?

  2. Does Theorem 2 apply to the exact posterior predictive release implemented in the experiments?

  3. How is the reported ((26.2,0))-DP guarantee justified given the positive tail probability in the theorem?

6 comments
AuthorsOpenReview ↗

Response to Reviewer ciqJ (Weakness 1)

Weakness 1: Trajectory protection and scope of DP guarantees

We would like to clarify the two privacy guarantees and their distinct objectives.

1. Trajectory protection and reconstruction risk (Theorem 1 and new Proposition).

Theorem 1 characterizes the displacement of posterior centroid means through Us(βj)\mathcal U_s(\beta_j), allowing the data owner to control perturbation over time, particularly during sensitive periods.

To quantify this protection more explicitly, we have added a new proposition immediately after Definition 1\color{#245A9B}{\text{we have added a new proposition immediately after Definition 1}}, establishing that

{\|\mathbf m_j(s)-\bar{\mathbf X}_j(s)\|} =1-\mathcal U_s(\beta_j).$$ This provides an exact quantitative measure of posterior centroid displacement from the empirical centroid relative to the distance between the empirical centroid and the prior mean trajectory. Unlike conventional differential privacy, which bounds changes in the output distributions of a randomized mechanism under neighboring datasets, our trajectory-protection measure quantifies how strongly the posterior centroid is displaced from its empirical counterpart. Specifically, the measure $1-\mathcal U_s(\beta_j)$ takes values between 0 and 1, representing the proportion of displacement from the empirical centroid toward the prior mean trajectory. For example, a value of 0.95 means that the posterior centroid is displaced by 95% of the distance between the empirical centroid and the prior mean trajectory, indicating substantial perturbation of the original centroid. This provides a direct quantitative characterization of the temporal perturbation achieved by our diffusion mechanism. Regarding adversarial reconstruction, $\color{#245A9B}{\text{Appendix A.3 (Threat Model and Adversary Knowledge)}}$ clarifies that the adversary may know the diffusion schedule. However, the original observations, empirical centroids, instantiated data-dependent target means, and original cluster sizes are not released. These undisclosed quantities prevent direct inversion from knowledge of the schedule alone. We also evaluate reconstruction attacks using repeated $K$-means clustering and Savitzky-Golay filtering. We kindly encourage the reviewer to see our response to Reviewer 5d3Y (Weakness 1, item #1) for further discussion of the quantitative trajectory-protection measure. **2. Motivation and scope of the formal DP guarantee (Theorem 2).** Our paper begins by identifying $\color{#245A9B}{\text{a different privacy risk arising in dynamic clustering applications, where sensitive}}$ $\color{#245A9B}{\text{temporal behaviors of cluster trajectories may matter more than individual observations.}}$ Motivated by reconstruction concerns identified in the literature, our primary objective is to protect these temporal patterns through controlled stochastic perturbation. $\color{#245A9B}{\text{Dirichlet process mixtures (DPMs) provide a natural foundation for addressing this problem.}}$ Unlike conventional clustering methods with deterministic centroids, DPMs model cluster centroids as random variables. Their Bayesian structure is therefore well suited to incorporating diffusion processes, opening a new direction for protecting sensitive temporal behaviors through stochastic centroid trajectories. This motivation led us to develop the proposed diffusion-based protection framework. After adopting the DPM framework, we recognized an additional useful property: $\color{#245A9B}{\text{the Dirichlet distribution governing the mixture weights provides a natural basis for establishing differential privacy for the}}$ $\color{#245A9B}{\text{sampled weights.}}$ We therefore investigated this property and established a formal $(\epsilon(\alpha),\delta(\alpha))$-DP guarantee in Theorem 2 under the stated conditions. $\color{#245A9B}{\text{This weight-level DP result is a complementary theoretical contribution, rather than the original motivation or primary}}$ $\color{#245A9B}{\text{objective of our framework.}}$ Our main contribution is not to establish individual-level DP simultaneously for mixture weights and centroid trajectories. Instead, we introduce and quantify a distinct form of protection for sensitive temporal behavior, while also demonstrating an additional privacy advantage of the Dirichlet weight mechanism. Accordingly, the DP guarantee in Theorem 2 applies to the sampled mixture weights, not to the complete synthetic-data release. For applications requiring individual-level DP guarantees for centroid trajectories which address a different privacy objective, Appendix A.7 (Limitations and Future Work) outlines two possible scenarios for extending the proposed framework in this direction. We kindly encourage the reviewer to see our response to Reviewer 5d3Y (Weakness 1, item #3) for further discussion of our primary motivation, the role of DPMs, and the complementary weight-level DP contribution.
AuthorsOpenReview ↗

Response to Reviewer ciqJ (Weakness 2)

Weakness 2: Consistency between the theoretical and experimental mechanisms and the reported DP guarantee

We would like to clarify that there is no inconsistency between the theoretical and experimental mechanisms\color{#245A9B}{\text{there is no inconsistency between the theoretical and experimental mechanisms}}. Both points raised by the reviewer are consistent with our theoretical framework and numerical results.

1. Sampled Dirichlet weights versus expected mixture weights.

Theorem 2 establishes an (ϵ(α),δ(α))(\epsilon(\alpha),\delta(\alpha))-DP guarantee for the sampled Dirichlet mixture weights. Although the posterior predictive construction uses expected mixture weights, we have already addressed this distinction in Remark 1, immediately following Theorem 2\color{#245A9B}{\text{we have already addressed this distinction in Remark 1, immediately following Theorem 2}}.

Specifically, Remark 1 explains how the privacy analysis of the sampled Dirichlet weights is related to membership leakage through the expected mixture weights used in the posterior predictive distribution. Thus, the use of expected weights in the posterior predictive construction does not represent an inconsistency with the theoretical analysis. Rather, the two results address related aspects of the proposed framework, as explicitly discussed in Remark 1.

2. Interpretation of the reported (26.2,0)(26.2,0)-DP setting.

The reported privacy setting is also consistent with the behavior of our theoretical bounds and numerical results. In Theorem 2, δ(α)\delta(\alpha) depends on the Dirichlet concentration parameter α\alpha. $\color{#245A9B}{\text{Figure 7 in Appendix A.4.3 explicitly illustrates how δ(α)\delta(\alpha) decreases toward zero as α\alpha increases}}$, across different sample sizes.

In particular, our experimental section states:

"Throughout, we set α=500\alpha=500 to impose only a weak (26.2,0)(26.2,0)-DP effect on the weights."

For the experimental sample size n=50n=50, Figure 7 shows that \color{#245A9B}{\text{\delta(500)$ is numerically extremely close to zero}}$, explaining the reported approximation and providing numerical evidence supporting the asymptotic result established in Corollary 3.

Importantly, the general definition of differential privacy permits δ=0\delta=0, corresponding to pure differential privacy. Our theoretical analysis establishes the limiting behavior δ(α)→0\delta(\alpha)\to 0 as α→∞\alpha\to\infty, while the numerical results demonstrate that the failure probability is already negligible at the chosen value α=500\alpha=500.

We therefore emphasize that the experimental setting is consistent with the theoretical trend established in Theorem 2\color{#245A9B}{\text{We therefore emphasize that the experimental setting is consistent with the theoretical trend established in Theorem 2}} and illustrated in Figure 7.\color{#245A9B}{\text{and illustrated in Figure 7.}} To avoid any ambiguity, we will clarify that the reported zero represents a numerical approximation to a very small positive value of δ(500)\delta(500).

AuthorsOpenReview ↗

Response to Reviewer ciqJ (Weakness 3)

Weakness 3: Experimental validation, reconstruction attacks, and perturbation design

We now clarify the scope of our experimental evaluation and address the reviewer's concerns.

1. Real-world dataset and higher-dimensional simulations.

Our real-data experiment uses electricity-consumption records from 7,975 customers, providing a substantial real-world sample for evaluating sensitive group-level temporal behaviors.

As discussed in our response to Reviewer DAUH (Weakness 4, item 3)\color{#245A9B}{\text{Reviewer DAUH (Weakness 4, item 3)}}, we have investigated additional publicly available time-series datasets suitable for time-dependent clustering, where centroid trajectories are meaningful and potentially sensitive. However, the alternative datasets we identified either did not align with our primary research objective or were not publicly available. The GoiEner dataset\color{#245A9B}{\text{GoiEner dataset}} was publicly accessible and provided an appropriate setting for evaluating our framework.

Regarding the one-dimensional simulation, our objective was to clearly illustrate the temporal perturbation behavior induced by diffusion\color{#245A9B}{\text{clearly illustrate the temporal perturbation behavior induced by diffusion}}, including its variation across sensitive and nonsensitive periods. Visualizing these effects becomes more challenging in higher dimensions. We are currently exploring suitable higher-dimensional simulation designs, together with appropriate visualizations and numerical metrics, to extend this evaluation.

2. Reconstruction attacks and adversary knowledge.

Our reconstruction experiment considers an adversary who repeatedly clusters synthetic releases using KK-means and applies Savitzky-Golay filtering to estimate the underlying centroid trajectories. This attack is motivated by the reconstruction vulnerability discussed in the related literature.

Importantly, Appendix A.3 (Threat Model and Adversary Knowledge)\color{#245A9B}{\text{Appendix A.3 (Threat Model and Adversary Knowledge)}} clarifies that the adversary may know the mean-reversion rates βj\beta_j and the functional specification of the diffusion coefficients σj(s)\sigma_j(s). Thus, our protection does not rely on keeping the general mechanism or its diffusion schedule secret.

However, the original observations, empirical centroids, instantiated data-dependent target means, and original cluster sizes are not released. Knowledge of the diffusion schedule alone therefore does not provide sufficient information to directly invert the posterior centroid estimates.

Despite our efforts to identify additional reconstruction attacks in the existing literature, we have not found model-aware attacks directly aligned with our setting. We would greatly appreciate any suggestions from the reviewer if they are aware of suitable methods that could further strengthen our experimental evaluation.

3. Owner-specified perturbation targets and the role of diffusion.

The perturbation targets are intentionally specified by the data owner because the sensitive temporal behaviors requiring protection depend\color{#245A9B}{\text{the sensitive temporal behaviors requiring protection depend}} on the application\color{#245A9B}{\text{on the application}}. This flexibility is an essential feature of our framework rather than an inconsistency in its implementation.

As clarified in Appendix A.3\color{#245A9B}{\text{Appendix A.3}}, the general diffusion parameters βj\beta_j and σj(s)\sigma_j(s) may be known to the adversary, whereas the instantiated data-dependent target means remain internal to the data owner. The sensitive time windows are also not explicitly disclosed, although periods of stronger perturbation may be inferred from the diffusion schedule.

We kindly encourage the reviewer to see our responses to Reviewer DAUH (Weakness 4, item 3)\color{#245A9B}{\text{Reviewer DAUH (Weakness 4, item 3)}} regarding dataset diversity and Reviewer 5d3Y (Weaknesses 2 and 3)\color{#245A9B}{\text{Reviewer 5d3Y (Weaknesses 2 and 3)}} regarding owner-specified perturbation parameters and the experimental baseline.

Reviewer_5d3YOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies a privacy risk in clustering data that changes over time. When centroids are released repeatedly, their paths over time can reveal how sensitive groups behave. The authors also argue that adding fresh random noise at each release does not help much because an attacker can smooth out that noise across many releases. To address the risk, the paper proposes DDirPMM in which centroids follow a random diffusion process. The data owner chooses a target path, and the diffusion pulls each centroid toward it. Separately, the authors show that the model's group weights satisfy differential privacy automatically. The released output is synthetic data generated from the fitted model. The experiments use a one-dimensional toy example and a Spanish smart-meter dataset.

Critical strengths
  1. The paper identifies a relevant and well-motivated problem. Most private clustering methods protect individual records. The authors point out that the collective behavior of groups over time can also be sensitive.

  2. The construction is elegant and gives interpretable controls. The random movement applied to the group averages can be described with a closed-form formula. In addition, the method has two parameters with clear jobs. One controls how much the group averages are shifted away from their true values. The other controls when that shifting happens. As a result, protection can focus on sensitive periods.

  3. The framework handles two different threats within one model. The model adds random movement to the group averages to address the first threat where an attacker rebuilds each group's average. The model also adds randomness to the group sizes to address the second threat where an attacker uses the group sizes to infer whether a specific person is in the data.

Critical weaknesses
  1. The protection for centroid trajectories has no formal guarantee. Definition 1 calls a trajectory "diffusion private" when the posterior mean sits between the observed mean and the owner's target. However, any Bayesian prior pulls estimates toward itself in this way. The definition therefore says nothing about how much an attacker can learn. In addition, Corollary 2(i) shows that the periodic σ(s)\sigma(s) brings centroids back near their true values at the start and end of each cycle. An attacker who knows the schedule could focus on those time points. The released synthetic data is built from both the group sizes and the group averages, but only the group sizes come with a proven privacy guarantee. As a result, the paper cannot state a formal privacy guarantee for the released data as a whole.

  2. The protection depends heavily on hand-tuned choices by the data owner. The owner must pick the sensitive windows, the target jump sizes, β\beta and the σ(s)\sigma(s) schedule for each dataset. The paper's own experiments set these values by hand. Moreover, the target means are built from the true cluster means at the start of each window. The paper does not discuss whether these choices could themselves leak information. Practical guidance on setting them is also limited. Consequently, it is hard to judge how well the method would transfer to new data.

  3. The experiments are too limited to support the paper's claims. The abstract claims better trade-offs than static trajectory protection methods. Yet the only baseline is the same model without diffusion. Noise-based methods from the related work are never compared. The results are mostly visual and include no numeric measure of reconstruction error.

Questions for authors

See weaknesses

3 comments
AuthorsOpenReview ↗

Response to Reviewer 5d3Y (Weakness 1)

Weakness 1: Trajectory protection, reconstruction risk, and DP guarantees

We thank the reviewer for these comments. We address the three concerns below.

1. Quantitative trajectory protection (Theorem 1 and new Proposition).

Our protection is not based solely on the usual shrinkage effect of a Bayesian prior. Theorem 1 characterizes the displacement of posterior centroid means through Us(βj)\mathcal U_s(\beta_j), which depends on the diffusion dynamics and original cluster sizes. This allows the data owner to control the degree of perturbation over time, particularly during sensitive periods.

To make this protection more explicit, we have added a new proposition immediately after Definition 1\color{#245A9B}{\text{we have added a new proposition immediately after Definition 1}}, establishing that

{\|\mathbf m_j(s)-\bar{\mathbf X}_j(s)\|} =1-\mathcal U_s(\beta_j).$$ This gives an exact quantitative measure of displacement from the empirical centroid. \$\color{#245A9B}{\text{For example, if $1-\mathcal U_s(\beta_j)=0.95$}}\$, the \$\color{#245A9B}{\text{posterior mean remains $95$}}\$% $\color{#245A9B}{\text{of the distance from the empirical centroid toward the prior mean trajectory.}}$ Thus, the proposition quantifies the achieved perturbation beyond simply stating that the posterior mean lies between two means. **2. Reconstruction under a known diffusion schedule (Appendix A.3).** Knowing the periodic diffusion schedule does not, by itself, reveal the original centroid trajectories. As clarified in Appendix A.3 (*Threat Model and Adversary Knowledge*): - The adversary may know $\beta_j$ and the functional specification of $\sigma_j(s)$. - The original observations, empirical centroids, instantiated data-dependent target means $m_j$, and original cluster sizes $n_j(s)$ are not released. - The posterior centroid means depend on the undisclosed target means and on $\mathcal U_s(\beta_j)$, which also depends on $n_j(s)$. $\color{#245A9B}{\text{Importantly, focusing on the beginning and end of each cycle does not, by itself, defeat our intended trajectory}}$ $\color{#245A9B}{\text{protection.}}$ Our framework protects sensitive temporal behaviors rather than perturbing every time point equally. The periodic diffusion alternates between stronger perturbation near cycle midpoints and greater data fidelity near boundaries, deliberately balancing privacy and utility. Even if an attacker obtains more accurate estimates at cycle boundaries, these observations do not directly determine the original trajectories during the more strongly perturbed intermediate periods. Direct inversion also requires undisclosed data-dependent target means and original cluster sizes. $\color{#245A9B}{\text{This selective protection is central to our design:}}$ preserving utility at certain time points while protecting sensitive temporal patterns. Moreover, Corollary 2(i) describes limiting behavior, not exact recovery at finite time points. We further evaluate reconstruction from synthetic releases using repeated $K$-means clustering and Savitzky-Golay filtering to assess recovery of the original temporal patterns. **3. Scope of the formal DP guarantee (Theorem 2).** Our primary contribution addresses a different privacy objective: $\color{#245A9B}{\text{protecting sensitive cluster-level temporal behaviors}}$ $\color{#245A9B}{\text{against reconstruction through controlled diffusion.}}$ We defined this risk from the outset and designed our framework accordingly. The absence of an individual-level DP guarantee for trajectories therefore does not invalidate this contribution, characterized by Theorem 1, the new proposition, and reconstruction experiments. $\color{#245A9B}{\text{Dirichlet process mixtures provide a natural framework for incorporating temporal behavior into clustering through}}$ $\color{#245A9B}{\text{diffusion.}}$ During its development, we also identified an additional advantage of the Dirichlet structure: $\color{#245A9B}{\text{a formal}}$ $(\epsilon(\alpha),\delta(\alpha))$-DP $\color{#245A9B}{\text{guarantee for mixture weights}}$, established in Theorem 2. This highlights the potential of Bayesian nonparametric models for privacy-preserving statistical learning. The weight-level guarantee is a complementary contribution, not our original motivation. As emphasized throughout the paper, it does not extend to the complete synthetic-data release. For readers interested in individual-level DP guarantees for centroid trajectories, $\color{#245A9B}{\text{Appendix A.7 (Limitations}}$ $\color{#245A9B}{\text{and Future Work)}}$ outlines two possible scenarios for extending the framework while retaining its cluster-level behavioral protection. We also kindly encourage the reviewer to see our response to $\color{#245A9B}{\text{Reviewer DAUH (Weakness 2)}}$, where we further discuss the motivation, contributions, and practical applications of diffusion-based trajectory protection.
AuthorsOpenReview ↗

Response to Reviewer 5d3Y (Weakness 2)

Weakness 2: Parameter selection, privacy leakage, and generalizability

We thank the reviewer for raising these concerns. We clarify the role of the data owner's choices and their implications for privacy and applicability.

1. The choices of βj\beta_j and σj(s)\sigma_j(s) are not data-dependent and are disclosed to the adversary.

The mean-reversion rate βj\beta_j and the functional form of the diffusion coefficient σj(s)\sigma_j(s) are design choices, rather than quantities that must be estimated from the original dataset. In particular:

  • σj(s)\sigma_j(s): The same diffusion function can be used across different datasets and applications. Its choice determines the temporal pattern of perturbation, from simple periodic behavior to more complex variations. Our toy experiments illustrate both cases, with a simple specification in the main paper and a more complex one in the appendix.
  • βj\beta_j: This parameter determines the rate of mean-reversion and influences the degree of perturbation. Its selection reflects the data owner's desired protection rather than a dataset-specific estimation procedure. The new proposition added immediately after Definition 1 further quantifies the resulting displacement through 1−Us(βj)1-\mathcal U_s(\beta_j).

Although the same choices can be applied to different datasets, the resulting perturbation also depends on internal model quantities. Thus, the framework does not require a new functional form for each dataset, while allowing the owner to adjust the protection level when needed.

2. Target means are intentionally flexible and remain undisclosed.

Our objective is not to prescribe specific target means or jump sizes for every application. Instead, the framework allows the data owner to specify the desired target behavior according to which temporal patterns require protection. Different applications may naturally require different targets (please see the two illustrative scenarios discussed in our response to Weakness 3 of Reviewer DAUH).

As explained in Appendix A.3 (Threat Model and Adversary Knowledge), neither the original dataset nor the instantiated data-dependent target means are released. These quantities remain internal to the data owner.

3. Sensitive windows reflect application-specific privacy needs.

The sensitive time windows are selected by the data owner according to the application and need not be explicitly disclosed. An adversary who knows the diffusion schedule may infer periods of stronger or weaker perturbation. However, our primary objective is to protect the underlying trajectory behavior during sensitive periods, rather than to conceal the locations of those windows. As clarified in Appendix A.3, the posterior centroid estimates also depend on undisclosed data-dependent target means and original cluster sizes, so knowledge of the schedule alone is insufficient for direct reconstruction.

4. Generalizability and additional clarifications.

The framework is not restricted to smart-meter data. Its diffusion functions and mean-reversion rates can be used with other temporal datasets, while the target means and sensitive windows can be adapted to the privacy objectives of each application. We have also expanded Appendix A.7 (Limitations\color{#245A9B}{\text{Appendix A.7 (Limitations}} and Future Work)\color{#245A9B}{\text{and Future Work)}} to clarify the role of these design choices and the scope of the proposed protection.

We kindly encourage the reviewer to see our response to Weakness 3 of Reviewer DAUH, where we further discuss the flexibility of owner-specified target means, the motivation for these choices, and their practical implications.

AuthorsOpenReview ↗

Response to Reviewer 5d3Y (Weakness 3)

Weakness 3: Experimental baselines and quantitative reconstruction evaluation

We would like to address the reviewer's concerns by providing further clarification on our choice of experimental baseline and quantitative evaluation.

1. Why we compare against the non-diffusion baseline.

Our framework is Bayesian, where cluster centroids are inherently modeled as random variables. In the non-diffusion setting, we consider the static Gaussian specification

\sim \mathcal N(\mathbf m_j,\boldsymbol{\Sigma}_j),$$ With time-independent parameters, $\color{#245A9B}{\text{the non-diffusion baseline effectively acts as a static Gaussian noise mechanism for the centroids}}$, incorporating stochastic perturbation without temporal diffusion. This is precisely why we compare the proposed diffusion mechanism against the non-diffusion model as our static Gaussian perturbation baseline. Adding independent Gaussian noise to these already random centroids would merely increase their variance without introducing a fundamentally different form of static perturbation, since Gaussian uncertainty is already incorporated into the centroid model. Therefore, our current comparison allows us to isolate and evaluate the benefit of diffusion in protecting sensitive temporal behaviors, while retaining the same underlying Bayesian mixture framework. We have now added this clarification to the experimental section of the revised manuscript to better explain our choice of baseline. **2. Comparison with existing noise-based methods.** Ravi et al. (2022), *Differentially Private K-Means Clustering Applied to Meter Data Analysis and Synthesis*, consider a related but different privacy objective. Their approach adds noise to deterministic K-means centroids to achieve differential privacy, whereas ours protects sensitive temporal behaviors of centroid trajectories through diffusion. Their Figure 3 in Ravi et al. (2022) illustrates how Savitzky-Golay filtering can help reconstruct underlying centroids from added independent nois to observations. A direct comparison with their method is not straightforward because the privacy objectives differ: individual-level differential privacy versus protection of sensitive cluster-level temporal behavior. Our non-diffusion Bayesian baseline instead provides a controlled comparison that isolates the effect of temporal diffusion. **3. Quantitative reconstruction evaluation using MMD and RMSE.** Our experiments are not limited to visual comparisons. We already employ $\color{#245A9B}{\text{Maximum Mean Discrepancy (MMD)}}$, a well-established measure of distributional discrepancy, to quantify differences between the original and synthetic data distributions. $\color{#245A9B}{\text{Figure 3 (bottom) and Figure 4 (right)}}$ report MMD values with $\color{#245A9B}{\text{95 percent confidence intervals}}$, providing quantitative assessments of distributional differences and uncertainty. We present these results graphically to make trends and privacy–utility trade-offs easier to interpret than in extensive numerical tables. We can also provide the corresponding tables if helpful. To further address the reviewer's concern, $\color{#245A9B}{\text{we will add Root Mean Squared Error (RMSE)}}$ as a direct quantitative measure of trajectory reconstruction error, defined as The additional RMSE analysis will directly assess reconstruction accuracy, complementing MMD's distributional evaluation and strengthening the quantitative comparison between diffusion and non-diffusion mechanisms.
Reviewer_DAUHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a privacy-aware dynamic Dirichlet process mixture model for protecting time-varying cluster information. The method uses a diffusion process to perturb centroid trajectories against reconstruction attacks and provides differential privacy guarantees for mixture weights. The framework is theoretically analyzed and evaluated on a synthetic dataset and a real smart-meter dataset, showing improved resistance to filtering-based reconstruction and favorable privacy-utility trade-offs.

Critical strengths

The paper identifies a meaningful privacy risk in dynamic mixture models by focusing on the reconstruction of time-varying cluster centroids. It also provides a well-motivated framework that combines diffusion-based trajectory protection with formal differential privacy guarantees for mixture weights. These contributions are theoretically supported and empirically validated on both synthetic and real-world data.

Critical weaknesses
  1. The differential privacy guarantee for mixture weights relies on restrictive conditions. The paper should discuss whether these conditions are realistic in practice, particularly for small clusters with near-zero weights, and clarify what happens to the privacy guarantee when the assumptions are violated.
  2. The motivation for adopting the proposed privacy notion is not sufficiently discussed. Although the paper explicitly distinguishes diffusion-based trajectory protection from differential privacy, the motivation for adopting this weaker privacy notion remains insufficiently justified. The paper should more clearly discuss its concrete advantages and limitations relative to stronger privacy guarantees, as well as the application scenarios in which such a threat-model-specific protection is preferable.
  3. The long-term target mean requires substantial manual design and domain knowledge. Since it is specified by the data owner and directly determines the perturbed centroid trajectory, the paper should better discuss how this target can be selected systematically in practice and how sensitive the results are to its choice.
  4. Limited experimental scale and diversity. The evaluation includes only a small synthetic dataset with 50 trajectories and one real-world smart-meter dataset with 7,975 users. Given that the proposed framework targets dynamic mixture models and emphasizes protection against temporal reconstruction attacks, additional large-scale real-world datasets would strengthen the empirical validation.
Questions for authors
  1. What concrete privacy-utility advantages does diffusion-based trajectory protection provide over stronger privacy guarantees such as differential privacy, and under what application assumptions are these advantages sufficient to justify the weaker guarantee?
  2. How realistic are the assumptions required for the differential privacy guarantee on mixture weights, especially for small clusters, and what guarantee remains when these assumptions fail?
  3. How should the long-term target mean be selected in practice, and how sensitive is the method to this choice?
5 comments
AuthorsOpenReview ↗

Response to Reviewer DAUH (Weakness 1)

Weakness 1: Assumptions underlying the differential privacy guarantee

We appreciate the reviewer's concern about Theorem 2, particularly for small or nearly empty components.

The condition πj(s)≥γα\pi_j(s)\geq\gamma_\alpha for all j∈[[N]]j\in[[N]] defines the admissible domain for our (ϵ(α),δ(α))(\epsilon(\alpha),\delta(\alpha))-DP bound. It is sufficient for this bound, but the Dirichlet posterior remains well-defined when nj(s)=0n_j(s)=0, since α/N+nj(s)>0\alpha/N+n_j(s)>0 for α>0\alpha>0.

Dirichlet mechanisms have previously required mechanism-specific conditions. For example, Gohari et al. (2021), Differential Privacy on the Unit Simplex via the Dirichlet Mechanism, derive privacy guarantees under conditions for their simplex-valued mechanism. These do not directly address the data-dependent posterior concentration parameters of Dirichlet process mixtures, motivating our setting-specific analysis.

We acknowledge that our condition can be restrictive for imbalanced mixtures. Our adaptive truncation (Remark 4, formerly Remark 3)\color{#245A9B}{\text{adaptive truncation (Remark 4, formerly Remark 3)}} favors components with meaningful weights. Figure 6 (Appendix A.4.1)\color{#245A9B}{\text{Figure 6 (Appendix A.4.1)}} shows how NN is selected using α\alpha, nn, and ρ\rho; different ρ\rho values can yield different NN.

The condition is equivalent to

\qquad n_j(s)\geq\frac{n[1-(1+\alpha)^{-q}]}N.$$ Thus the minimum integer count is $\lceil n\gamma_\alpha\rceil$. **Scenario 1: Toy simulation ($n=50$, $N=7$, $q=0.1$)** Appendix A.4.1 reports $N=7$ for $n=50$, $\alpha=10^{-5}$, and $\rho=0.05$. With $q=0.1$ (Appendix A.4.2), only $\color{#245A9B}{\text{one observation per component}}$ is required. Holding $N=7$ fixed to illustrate the effect of $\alpha$ gives: | $\alpha$ | $10^{-5}$ | $0.1$ | $1$ | $5$ | $10$ | $40$ | $80$ | $300$ | $1000$ | |---|---|---|---|---|---|---|---|---|---| | $n\gamma_\alpha$ | 0.000007 | 0.0678 | 0.4783 | 1.1717 | 1.5229 | 2.2157 | 2.5400 | 3.1063 | 3.5633 | | Minimum $n_j(s)$ | **1** | **1** | **1** | **2** | **2** | **3** | **3** | **4** | **4** | *Table 1: Minimum counts for $n=50$, $N=7$, $q=0.1$. In practice, $N$ is adaptively selected using $\alpha$, $n$, and $\rho$.* **Scenario 2: Very small sample ($n=10$, $N=6$, $q=0.1$)** | $\alpha$ | $10^{-5}$ | $0.01$ | $0.1$ | $1$ | $10$ | $100$ | $1000$ | |---|---|---|---|---|---|---|---| | $n\gamma_\alpha$ | 0.000002 | 0.0017 | 0.0158 | 0.1116 | 0.3553 | 0.6150 | 0.8314 | | Minimum $n_j(s)$ | **1** | **1** | **1** | **1** | **1** | **1** | **1** | *Table 2: Minimum counts for $n=10$, $N=6$, $q=0.1$; $N$ is fixed for illustration.* Even for 10 observations and 6 components, only $\color{#245A9B}{\text{one observation per component}}$ is required across these $\alpha$ values. The condition is feasible, since six occupied components require six of the ten observations. These calculations show that required component sizes can be small. Adaptive truncation may mitigate the restriction but $\color{#245A9B}{\text{does not guarantee every}}$ $\color{#245A9B}{\text{retained component is occupied}}$; the condition must hold for the neighboring datasets considered in Theorem 2. If it fails, $\color{#245A9B}{\text{the theorem's stated DP}}$ $\color{#245A9B}{\text{bound is not established}}$, though this does not rule out other guarantees. The empirical-proportion condition defines admissible datasets, while $\delta(\alpha)$ bounds the probability that sampled weights lie outside $\Omega_{\gamma_\alpha}^{w}$. **Revision:** We added Remark 3 after Theorem 2's proof, with two numerical tables and clarification that the DP guarantee covers $\color{#245A9B}{\text{sampled Dirichlet}}$ $\color{#245A9B}{\text{weights on the admissible domain}}$, not the entire posterior predictive release. The previous adaptive-truncation Remark 3 is now Remark 4. **Reference:** Gohari et al. (2021), *Differential Privacy on the Unit Simplex via the Dirichlet Mechanism*, IEEE TIFS 16, 2326–2340.
AuthorsOpenReview ↗

Response to Reviewer DAUH (Weakness 2)

Weakness 2: Motivation for diffusion-based trajectory protection

We thank the reviewer for this suggestion. We agree that the advantages, limitations, and suitable applications of diffusion-based trajectory protection relative to DP could be made more explicit. We would like to clarify that the original manuscript already provides motivation for this threat-model-specific protection, as detailed below.

Motivation, advantages, and applications already discussed in the manuscript.

  • Practical motivation and privacy risks (Section 1): We motivate our framework through dynamic mixture models used in smart-grid applications, including energy resource modeling and network equipment analysis (Najafi & Livani, 2025; Li et al., 2025; Gemine et al., 2016). We explain that sequential releases of mixture parameters can reveal sensitive group-level temporal patterns (Xiong et al., 2011; Chicco, 2012; Seo et al., 2025; Dai et al., 2025; Adam et al., 2021; Zhang et al., 2025). In particular, centroid trajectories can expose sensitive aspects of group behavior (Sharma et al., 2025), while independent perturbations may be filtered out across repeated releases, enabling reconstruction of the underlying trajectories. These risks motivate protection specifically against temporal trajectory reconstruction.

  • Practical illustration and distinction from individual-level privacy (Figure 1, Section 1): Figure 1 illustrates a smart-meter scenario in which temporal electricity-consumption patterns reveal sensitive characteristics of customer groups, potentially enabling targeted commercial decisions. This example demonstrates why protecting individual records alone does not necessarily address the disclosure of sensitive group-level characteristics. Our threat model therefore focuses on preventing the reconstruction of sensitive group-level trajectories from repeated releases, rather than solely protecting an individual's participation in the dataset.

  • Advantages and distinction from DP (Sections 1 and 4.2): We explain that diffusion introduces temporally structured perturbations, allowing their magnitude to vary across time. This flexibility enables greater perturbation during sensitive periods while limiting unnecessary utility loss elsewhere. Unlike DP, which provides a formal neighboring-dataset indistinguishability guarantee, our diffusion-based notion targets the reconstruction of sensitive centroid trajectories. This distinction is already explicitly stated immediately after Definition 1 in Section 4.2: "Here, privacy means trajectory protection through controlled perturbation, rather than a DP guarantee."

  • Practical application and experimental evidence (Section 5): We evaluate the framework using the GoiEner smart-meter dataset, demonstrating its relevance to real-world group-level electricity-consumption patterns. Our experiments also investigate filtering-based trajectory reconstruction and temporally varying perturbations. These results illustrate the practical motivation for diffusion-based protection and its ability to concentrate perturbations within sensitive time windows rather than introducing comparable perturbations throughout the entire trajectory.

Additional clarification in the revised manuscript.

To directly address the reviewer's request for a clearer comparison with DP, we have added the following paragraph immediately after Definition 1\color{#245A9B}{\text{immediately after Definition 1}} in Section 4.2\color{#245A9B}{\text{in Section 4.2}}, while retaining the original statement distinguishing diffusion-based protection from DP:

"The main advantage of this approach is its ability to adjust perturbation over time, providing stronger protection during sensitive periods while preserving utility elsewhere. This is particularly useful in applications such as smart meters, where group-level temporal patterns may reveal sensitive information. However, this protection targets trajectory reconstruction and does not provide a formal neighboring-dataset indistinguishability guarantee."

We would like to emphasize that our motivation for adopting diffusion-based protection is not to replace DP or claim a stronger formal privacy guarantee. Rather, as illustrated in Figure 1, our framework addresses a different disclosure risk: the reconstruction of sensitive group-level temporal patterns from repeated releases. Its concrete advantage is temporally adaptive perturbation, which can prioritize sensitive periods while preserving utility elsewhere. Its limitation is the absence of a formal neighboring-dataset indistinguishability guarantee for centroid trajectories. Consequently, this approach is particularly suitable when protecting group-level temporal patterns is the primary objective, whereas applications requiring formal individual-level DP guarantees should not rely on diffusion-based trajectory protection alone.

AuthorsOpenReview ↗

Response to Reviewer DAUH (Weakness 3)

Weakness 3: Selection and sensitivity of the long-term target mean

We thank the reviewer for this comment. We would like to clarify that the contribution of our paper is to provide a statistical tool for data\color{#245A9B}{\text{the contribution of our paper is to provide a statistical tool for data}} owners who wish to conceal sensitive group-level trajectory patterns while sharing synthetic data\color{#245A9B}{\text{owners who wish to conceal sensitive group-level trajectory patterns while sharing synthetic data}}. The purpose is not to prescribe which patterns should be hidden, as these depend on the application and the data owner's privacy concerns. Different data owners may have different preferences, and it is neither necessary nor within the scope of our framework to anticipate every possible scenario. Instead, our method allows the owner to specify the desired alternative behavior through the long-term target mean and incorporate it into the perturbed trajectories during sensitive intervals.

1. Motivation for the data owner's choice.

As illustrated in Figure 1\color{#245A9B}{\text{Figure 1}}, an electricity utility company may share consumption data with external parties, including appliance manufacturers. The company may wish to prevent these parties from identifying sensitive consumption patterns of particular customer groups, which could otherwise be used for targeted or potentially unfair promotions. Importantly, the owner does not need to know the intentions of a particular adversary\color{#245A9B}{\text{the owner does not need to know the intentions of a particular adversary}}. It is sufficient to identify which aspects of the trajectories should be concealed, regardless of how the released data may subsequently be used.

2. Different targets for different privacy concerns.

The choice of the long-term target mean depends on the behavior the owner wishes to hide. We illustrate this with two scenarios:

  • Hiding sensitive changes in trajectory magnitude: This is the scenario considered in Section 5\color{#245A9B}{\text{Section 5}}, where the target means are chosen to make sensitive changes in group trajectories harder to distinguish. In particular, the GoiEner smart-meter experiment considers consumption patterns that may reveal sensitive changes in customer-group behavior. The specific target settings are provided in Section 5.

  • Hiding differences between group trajectories: In other applications, the concern may be distinguishing groups rather than identifying the magnitude of their changes. For example, a healthcare data provider may wish to prevent external users from distinguishing patient groups based on their activity patterns during certain periods. In this case, the owner could choose a common long-term target mean for all components\color{#245A9B}{\text{a common long-term target mean for all components}}, pulling their trajectories toward a similar level during sensitive intervals and making the original differences between groups harder to recover.

These examples illustrate that the same statistical framework can accommodate different privacy objectives without requiring a separate\color{#245A9B}{\text{the same statistical framework can accommodate different privacy objectives without requiring a separate}} mechanism for each application\color{#245A9B}{\text{mechanism for each application}}.

3. Practical selection of the target mean.

When the owner has a specific privacy concern, the target can be chosen to reflect the alternative trajectory behavior they wish to present during sensitive intervals. When this choice is less clear, several candidate targets can be considered, for example, targets that reduce abrupt changes or bring group trajectories closer together. Their effects on reconstruction error and statistical utility can then be compared to guide the selection.

Thus, our framework does not aim to identify an optimal target for every possible dataset\color{#245A9B}{\text{our framework does not aim to identify an optimal target for every possible dataset}}. Rather, it provides the flexibility to incorporate the data owner's privacy preferences into the statistical model.

4. Sensitivity to the target mean and additional experiment.

The effect of the target mean on the perturbed trajectories is intentional: changing the target changes the behavior toward which the centroids are pulled. This allows the owner to control which sensitive patterns are made less apparent while retaining greater fidelity outside sensitive intervals.

To illustrate this flexibility and examine the effect of an alternative target choice, we will add an experiment based on the second scenario\color{#245A9B}{\text{we will add an experiment based on the second scenario}}, where distinct cluster trajectories are pulled toward a common target during sensitive intervals. This will complement our existing experiments on sensitive changes in trajectory magnitude and provide further evidence of how target selection affects the resulting trajectories and statistical utility.

We will clarify these points in the revised manuscript.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Report to the Area Chair and All Reviewers

Report of Major Revisions and Additional Results

We sincerely thank all reviewers for their thoughtful feedback and constructive suggestions. Their comments have encouraged us to examine our framework from different theoretical and experimental perspectives, leading to substantial improvements in the manuscript.

Despite the short rebuttal period, we have worked extremely hard to strengthen the paper, including new theoretical justifications, additional propositions and remarks, extensive numerical and simulation studies, and a new multivariate experiment. We have also expanded the methodological explanations to clarify the assumptions, privacy guarantees, practical choices, and limitations of our framework. All new additions and revisions are highlighted in Blue\color{blue}{\text{Blue}} in the updated manuscript. We summarize the three main areas of improvement below.

1. Stronger theoretical justification and clearer privacy guarantees

Addressing the theoretical concerns of Reviewers DAUH, 5d3Y, and ciqJ, we made several important additions:

  • Quantitative trajectory protection (Proposition 2, Section 4.2): We establish an exact characterization of posterior centroid displacement:

    {\|m_j(s)-\bar{x}_j(s)\|} =1-\mathcal U_s(\beta_j).$$ This quantifies the perturbation achieved by diffusion, strengthening Theorem 1 and Definition 1 (Reviewers 5d3Y and ciqJ).
  • DP assumptions and feasibility (Remark 3, Appendix A.2; Tables 9–10): Addressing Reviewer DAUH's concerns about small clusters, we clarify Theorem 2's admissible domain and numerically demonstrate that its conditions can remain feasible even for small samples, while acknowledging their restrictions.

  • Scope of DP (Remark 1, Theorem 2, Corollary 3, Figure 8): Addressing Reviewer ciqJ, we distinguish sampled from expected mixture weights and clarify that our main contribution targets privacy risks from centroid trajectories, while DP for sampled weights is an additional benefit naturally arising from the Dirichlet process mixture framework, without a separate noise mechanism. We also clarify δ(500)≈0\delta(500)\approx 0, consistent with its large-α\alpha behavior.

  • Threat model (Appendix A.3): Addressing Reviewers 5d3Y and ciqJ, we clarify adversary knowledge and the undisclosed data-dependent quantities, strengthening the interpretation of reconstruction risks.

  • Motivation and practical design (Sections 1 and 4.2): Addressing Reviewers DAUH and 5d3Y, we clarify the advantages and limitations of group-level diffusion protection relative to individual-level DP, along with target selection, sensitive windows, and diffusion parameters.

Together, these additions strengthen the theoretical interpretation, practical relevance, and precision of our privacy claims.

2. Extensive new numerical and simulation studies

Addressing concerns about experimental diversity, reconstruction evaluation, parameter sensitivity, and scale (Reviewers DAUH, 5d3Y, and ciqJ), we substantially expanded Appendix A.4:

  • Reconstruction and utility (Section 5, Figure 3; Appendix A.4.3, Tables 11–12): We added RMSE to the bottom of Figure 3 in the main paper, directly quantifying trajectory reconstruction error. Additional MMD and RMSE results with uncertainty estimates demonstrate stronger protection during sensitive windows while limiting utility loss elsewhere.

  • Diffusion-function sensitivity (Appendix A.4.4; Figures 9–10): A more complex nonlinear periodic diffusion coefficient demonstrates flexibility beyond the original simple specification.

  • Target-mean sensitivity (Appendix A.4.5; Figure 11, Tables 13–14): Directly addressing Reviewer DAUH, we compare three target strategies (between, above, below), quantifying their effects on reconstruction RMSE and MMD and demonstrating flexibility across privacy objectives.

  • Sample-size sensitivity (Appendix A.4.6; Figure 12, Tables 15–16): We extend the experiments to n=10n=10 through 1,0001,000, examining small-sample limitations and protection at larger sample sizes.

  • Adaptive truncation and DP behavior (Appendices A.4.1–A.4.2; Figures 6–8, Tables 9–10): Additional numerical analyses examine component occupancy, truncation, and privacy parameters, supporting the practical interpretation of Theorem 2.

We also clarify the static Gaussian baseline in Section 5, distinguishing MMD (distributional utility) from RMSE (reconstruction accuracy).

These controlled studies isolate different aspects of our framework, providing substantially stronger quantitative evidence and a clearer understanding of its privacy–utility behavior.

3. New multivariate toy example and higher-dimensional extension

Addressing Reviewer ciqJ's concern about one-dimensional experiments, we added a new two-dimensional toy example (Appendix A.4.7; Figures 13–15).

Unlike the original jump-magnitude experiment, this study protects sensitive group locations and movements by drawing distinct centroid trajectories toward a shared target.

  • Figure 13: Shows joint two-dimensional trajectories, illustrating how increasing βj\beta_j reduces group distinguishability during sensitive windows.
  • Figure 14: Compares coordinate-wise reconstruction with and without diffusion, demonstrating the additional protection from temporal perturbation.
  • Figure 15: Quantifies multivariate privacy–utility trade-offs using MMD and reconstruction RMSE, including uncertainty intervals.

Appendix A.4.8 further explains the extension to dimensions d>2d>2.

Together, these new experiments demonstrate that our framework extends beyond one-dimensional trajectories and jump-magnitude protection, providing both visual and quantitative evidence of multivariate trajectory protection.

Closing remarks

New additions in the revised manuscript:

  • Theory: Proposition 2 (quantitative centroid displacement) and Remark 3 (scope and practical feasibility of the DP assumptions).
  • Tables 9–16: Eight new tables covering DP feasibility, reconstruction RMSE, distributional utility (MMD), target-mean sensitivity, and sample-size sensitivity.
  • Figures 7 and 11–15: Six new figures covering component occupancy, target-mean and sample-size sensitivity, and a comprehensive two-dimensional experiment.
  • Figure 3 (updated): A new RMSE panel added to the main-paper toy experiment.
  • Additional clarifications: Expanded discussion of privacy guarantees, the threat model, target selection, and the higher-dimensional extension (Appendix A.4.8).

We would like to emphasize that these revisions are substantive additions to both the theoretical and experimental contributions, rather than merely expanded explanations. The new propositions, numerical tables, controlled sensitivity studies, and multivariate experiments directly respond to the reviewers' concerns and provide a more comprehensive evaluation of the framework.

We are particularly grateful that the reviewers' different perspectives motivated these additional investigations, which have helped us better establish the strengths, practical behavior, and limitations of our approach.

We sincerely hope that the reviewers and Area Chair will consider these substantial improvements when reassessing the manuscript. We greatly appreciate their time, constructive feedback, and careful consideration of our revised work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.