acceptodds
Under review as a conference paper at ICLR 2027

Tight End-to-End Accounting for Differentially Private Synthetic Data

24 views

Abstract

Differentially private (DP) pipelines often combine heterogeneous mechanisms, yet their guarantees are best analyzed using different notions of privacy (e.g., bounded range, zCDP, ADP). When composing them, we have to analyze the guarantees through a reduced common representation that loses accounting precision. This issue is particularly relevant in DP synthetic data generation, where data preprocessing, generation algorithms, and DP-SGD could appear within the same pipeline. We develop novel tools for tight end-to-end accounting of such compositions using privacy loss distributions (PLD), -DP, and Gaussian DP (GDP). In particular, we derive PLD and GDP analyses of bounded range mechanisms such as the Exponential mechanism. Building on these theoretical results, we introduce AIM-GDP, a GDP-native variant of AIM (a state-of-the-art DP synthetic data mechanism) that enables easy composability with complex pipelines via PLD, attack-aware risk calibration, and tight end-to-end auditing while consistently achieving better utility; and MixCal, an algorithm for calibrating adaptive mixed compositions using PLD accounting. Overall, our tools enable tighter end-to-end accounting and thus less noise at the same privacy level in DP synthetic data pipelines.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
differential privacy, PLD, f-DP, GDP, heterogeneous mechanisms, synthetic data generation

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_W2ANOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The private synthetic data generation algorithm, AIM, first applies exponential mechanisms and then applies DP-SGD (multiple Gaussian mechanisms). The standard privacy accounting applies composition theorem of these mechanisms under the same privacy notion. However, these mechanisms are best analyzed under different privacy notions, and converting between them using existing analysis can lead to a looser privacy accounting bound. Specifically, exponential mechanisms are best analyzed with bounded range (BR), while Gaussian mechanisms are better analyzed and composed with GDP or f-DP.

This paper provides a tighter privacy conversion from BR to GDP and f-DP than existing works, enabling a tighter composition of exponential mechanisms and DP-SGD, and hence a tighter privacy accounting for the AIM algorithm.

In addition, realistic data generation pipelines also include pre-processing steps, which consume part of the privacy budget. Hence, this work proposes MixCal to better calibrate the privacy budget assigned to each component in the pipeline, given that each component has an analytical privacy loss distribution (PLD).

Through experiments, they show that their proposed accounting reduces marginal errors in comparison with the original AIM algorithm.

Critical strengths
  • The paper provides a new, tighter privacy conversion analysis from bounded-range mechanisms to f-DP and GDP, which allows for a better composition analysis of exponential and Gaussian mechanisms. This is useful and can improve the privacy analysis of many such pipelines.

  • The paper is very well written. The background section is detailed and clear, and all results and their intuitions are very well explained.

Critical weaknesses

MixCal only calibrates the privacy budget assigned to each component, but its implications for the final utility of the pipeline are not clear. Multiple bb can satisfy the same final privacy guarantee, but MixCal doesn't seem to offer a way to select the best one. Additionally, the method appears to rely on all components having an analytical PLD, which may be difficult for some specific pre-processing algorithms.

Questions for authors

NA

2 comments
AuthorsOpenReview ↗

Response to Reviewer W2AN

Thank you for the review!

MixCal only calibrates the privacy budget assigned to each component, but its implications for the final utility of the pipeline are not clear.

In our experiments, we found that the simple budget allocation rule based on standard zCDP analyses (see Appendix C.1) performs well and outperforms existing composition methods. Exploring the optimal allocation for generic pipelines requires a lot more investigation, including different application settings, as the question of choosing the "dummy" budget notion could be highly application-dependent through the specific choice of component mechanisms. As we mentioned in Section 6, we therefore designate such investigations to future work.

The method relies on all components having an analytical PLD, which may be difficult for some specific pre-processing algorithms.

Indeed, and because our method shows improvements in utility and additional capabilities, we hope our findings motivate further work on deriving PLD-based analyses for more DP algorithms.

Thank you for the response. I agree that finding the optimal budget allocation is a reasonable future direction. Overall, I find the results of this paper interesting and significant, and therefore I'll maintain my positive score and recommend acceptance.

Reviewer_h6YGOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

Novel tools for tighter end-to-end accounting of compositions of heterogeneous mechanisms are developed. These are then used for AIM-GDP, a GDP variant of AIM and MixCal, an algorithm for calibrating adaptive mixed compositions. The tools allow for less noise at the same privacy level in end-to-end DP synthetic data pipelines due to tighter accounting. The contributions are both the theoretical results as well as the very relevant application in DP synthetic data.

Critical strengths
  • The theoretical side is strong. The analysis of bounded range mechanisms that are introduced seem correct and might be of broader interest beyond synthetic data.
  • The new results lead to improved accuracy in certain DP synthetic data algorithms due to tighter composition.
  • The paper is well written and provides an accessible presentation for people who are not experts on DP accounting.
Critical weaknesses

There are certain weaknesses but they are clearly acknowledge by the authors. For example, the fact the accounting improvements might not hold if most of the privacy budget is spent outside of Gaussian mechanisms, among others noted at the end.

One small thing: the title is a bit misleading since "tight" characterizations are for individual bounded-range mechanisms not for the end-to-end DP synthetic data pipelines.

Questions for authors

it remains a bit unclear how to choose the most appropriate accounting strategy for a general heterogeneous pipeline. Evidence was given that, for compositions of Exponential and Gaussian mechanisms, the relative performance of GDP and zCDP depends on the allocation of privacy budget. Is there a more systematic way of choosing between different accounting frameworks?

1 comment
AuthorsOpenReview ↗

Response to Reviewer h6YG

Thank you for the review!

The title is a bit misleading.

We will change the title to "Improved."

Is there a more systematic way of choosing between different accounting frameworks for a general heterogeneous pipeline (GDP/zCDP)?

We are assuming that the reviewer refers to choosing between GDP and zCDP in a composition of Exponential + Gaussian mechanisms as in Fig. 4 (left), modeling AIM (please let us know if this is not the right interpretation).

In Select-Measure-Generate synthetic data generation mechanisms (e.g., MST, AIM, RAP, GEM), it is standard to fix the proportion of the budget for selection (Exponential mechanism) and measurement (Gaussian mechanism), thus we can indeed use this fixed proportion as a systematic way to decide on the accounting mechanism. E.g., GEM uses a 50% split for selection and measurement, which means GDP accounting will not be helpful, whereas AIM or GEM+ use 90%, which means GDP is a better choice.

Select-Measure-Generate Synthetic Data Generation: https://differentialprivacy.org/synth-data-1/

RAP: https://arxiv.org/abs/2103.06641

GEM: https://arxiv.org/abs/2106.07153

GEM+: https://arxiv.org/abs/2511.09672

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

This work considers improvements to privacy accounting for data processing pipelines involving synthetic data generation. In particular, the authors consider pipelines involving the AIM algorithm for producing synthetic data.

The main accounting improvement comes from replacing the zCDP filter employed internally in AIM with a GDP filter. To make this possible, the authors give a new PLD/GDP analysis of bounded range mechansism. This covers the exponential mechanism, which is a key component of AIM. The authors note that the GDP analysis of the selection step of AIM is not universally tighter than the previous zCDP analysis and rely instead on empirical measurements to establish the improvement.

Additionally, the authors provide a heuristic, MixCal, to calibrate privacy budgets allocated to heterogeneous workloads inside a larger processing pipeline. MixCal applies binary search to divide an exact privacy budget according to a user-specified allocation in order to provably satisfy a downstream scalar budget.

The authors provide empirical measurements comparing AIM vs AIM-GDP, which uses a GDP filter and the new GDP analysis of the selection step. In their experiments, AIM-GDP provides consistent but modest utility improvements compared to AIM, with only a small increase in runtime. The authors extend the experiments to cover the performance of AIM-GDP within a larger synthetic data pipeline involving DP-SGD. They show that their accounting improvements lead to a consistent reduction in the amount of noise required during DP-SGD to achieve a given privacy level.

Critical strengths

The authors provide a new exact privacy analysis of bounded range mechansisms and show empirically that it can lead to a practical performance improvement when integrated into a larger synthetic data pipeline. The exact analysis of bounded range mechanisms is new and independently interesting.

The empirical evaluation is generally thorough and covers a range of interesting and practical settings to establish the performance benefits. In particular, it is very nice to see the evaluation of AIM-GDP in isolation as well as when integrated into a practical workflows such as synthetic text generation. Overall, the authors provide a very thoughtful and careful treatment of good accounting practices across the entire synthetic data pipeline.

Critical weaknesses

While it is not a significant weakness, I think the authors may be overstating things somewhat by claiming a "tight end-to-end accounting" for the synthetic data setting. The authors clearly demonstrate exact accounting of an important component of the pipeline (selection) but are unable to retain the exact PLD accounting when integrated into the larger process. Consequently, the overall process still has gaps in its privacy accounting, including the GDP filter/analysis for the selection step. Moreover, a tighter accounting for the entire pipeline compared to the previous approach cannot be established in all cases, as can be seen in Fig 3. Overall, framing this work as something like "improved end-to-end accounting" feels more representative to me than the claimed "tight end-to-end accounting".

Questions for authors

The authors claim that "For fully-adaptive composition, the PLD approach generally does not provide valid bounds... As far as we are aware, the tightest known bounds for general heterogeneous fully-adaptive composition of mechanisms are ρ\rho-zCDP or μ\mu-GDP". The cited Regehr et al. work does on the other hand offer an approach to PLD-based accounting under full adaptivity in terms of residuals. For instance, the filter using GDP residuals always performs at least as well as the GDP filter used in Alg 5. It also offers a large improvement when the PLD has bounded support, which should apply to the bounded range mechanism. Have the authors considered any improvements to the GDP filter used in their algorithm that achieves something closer to exact accounting?

The authors also state that "We provide a method for converting a PLD of an arbitrary mechanism to the PLD for the mechanism wrapped with Poisson subsampling... Previously, only additive mechanisms like Gaussian and Laplace mechanisms were supported." I am a bit confused by this statement as the (later) cited Zhu et al. work does provide a generic PLD amplification result under Poisson subsampling. As far as I can tell, Theorem B.1 is just converting this statement to the language of hockey-stick curves. Can the authors please clarify whether there is any novel contribution for PLD amplification of Poisson subsampling?

2 comments
AuthorsOpenReview ↗

Response to Reviewer NbHX (1/2)

Thank you for the review!

"Tight" is somewhat overstating things, "improved" feels more representative.

We will change the title to "Improved."

The cited Regehr et al. work does offer an approach to PLD-based accounting under full adaptivity in terms of residuals. Have the authors considered it?

This is a great pointer! For this response, we implemented this filter in the setting of our simulations in Fig. 4, and found that the improvement for AIM is less than 1.5% in terms of ε\varepsilon for realistic settings with less than 50 iterations (this is what we typically observe in our empirical results). See our numerical results in the table below (T is the number of compositions of Exponential + Gaussian mechanism pairs).

As this filter in practice adds substantial computational overhead to AIM (it requires a bisection over multiple convolutions of the Exponential PLD with the Gaussian PLD to find the optimal μ\mu), we leave detailed investigations to future work. We will clearly describe this in Section 6:

“In a concurrent work, Regehr et al. (2026) have proposed a PLD-based fully adaptive filter construction. In our experiments, we observed modest improvements in the fully adaptive setting, e.g., under 1.5% improvement in ε\varepsilon values for realistic instantiations of AIM, while at the same time adding computational overhead. Although it is out of scope for our work, we believe that further in-depth investigation of such filters is a potentially fruitful avenue for future work.”

Tε\varepsilon (GDP)ε\varepsilon (Regehr et al.)% improvement
10.510.510.0%
51.241.231.0%
101.821.801.1%
202.682.651.2%
504.534.471.3%

Difference between the PLD amplification result under Poisson subsampling, and the results by Zhu et al.

In terms of the information content, a (infinite-precision) dominating pair of distributions, a privacy loss distribution, and the privacy profile function ϵ(δ)\epsilon(\delta) are all equivalent, and indeed Zhu et al. provides (i) conversions between these representations, and (ii) a mapping from a dominating pair to the dominating pair under Poisson sampling. However, in actual (finite-precision) implementations, the barriers to directly applying their result are: (i) we do not store the dominating pair but the privacy loss distribution, (ii) in practice, the privacy loss distribution and dominating pair representations are unstable, but the ϵ(δ)\epsilon(\delta) function they map to is stable. So, by providing the conversion directly in terms of ϵ(δ)\epsilon(\delta) rather than in terms of dominating pairs, our result effectively makes the results of Zhu et al. applicable to modern accounting libraries in a numerically stable way. We will update the wording in the revision to make this clear (addition italicized here):

"We provide a method for converting a PLD of an arbitrary mechanism to the PLD for the mechanism wrapped with Poisson subsampling by going through the privacy profile, which can easily be converted to/from a privacy loss distribution. This method is numerically stable and compatible with such accounting tools. Previously, Zhu et al. have shown the equivalent results in terms of dominating pairs. However, the dominating pair (and PLD) suffer from numerical stability issues even though the privacy profile they map to remains numerically stable. In turn, modern accounting libraries do not implement the general result of Zhu et al. and often only support subsampling for Gaussian and Laplace mechanisms."

This is one of the theoretical contributions of our work that we present in the Appendix. It constitutes a "quality-of-life" improvement for accountant implementations, rather than being a result that is useful in theoretical applications.

AuthorsOpenReview ↗

Response to Reviewer NbHX (2/2)

In their experiments, AIM-GDP provides consistent but modest utility improvements compared to AIM, with only a small increase in runtime.

We would like to contextualize this more. First, note that our tighter GDP accounting delivers consistent utility improvements across datasets, privacy budgets, and complex/end-to-end pipelines, without changing the underlying mechanism.

Second, we found that the modest improvements are with respect to a strong baseline that we implemented for this work (we did so to cleanly isolate the effect of GDP accounting), and we observe stronger improvements w.r.t. other implementations. For this response, we ran additional experiments and added a benchmark on the Adult dataset with six other open-source AIM implementations. We found that our baseline AIM implementation is on par with the latest version of AIM’s original authors’ implementation (i.e., AIM (MBI latest)), and both of our AIM baseline and AIM-GDP substantially outperform all other existing implementations. We show the results in the table below. The differences are due to updated software dependencies (specifically, the MBI package), and better defaults such as the default marginal degree and the default allocation of budget for 1-way marginal measurements. We will add these details in Section 5.

Therefore, the utility improvements in our end-to-end implementation of AIM-GDP (with the code being a part of a contribution as per Section 1) could be substantial, depending on the reference point.

Algorithm1-way2-way3-wayFullAIM-GDP improvement (%)
AIM (synthcity)0.47120.65740.77690.7339+95.5%
AIM (OpenDP)0.00240.01940.04600.0376+12.8%
AIM (dpmm)0.00360.01900.04460.0366+10.4%
AIM (SmartNoise)0.00370.01860.04380.0359+8.6%
AIM (MBI original)0.00340.01790.04270.0349+6.0%
AIM (MBI latest)0.00320.01710.04140.0338+3.0%
AIM (ours)0.00330.01730.04110.0337+2.7%
AIM-GDP (ours)0.00320.01670.04010.0328

AIM (Synthcity): https://github.com/vanderschaarlab/synthcity/blob/main/src/synthcity/plugins/core/models/aim.py

AIM (OpenDP): https://github.com/opendp/opendp/blob/main/python/src/opendp/extras/mbi/_aim/__init__.py

AIM (dpmm): https://github.com/sassoftware/dpmm/blob/main/src/dpmm/models/aim.py

AIM (SmartNoise): https://github.com/opendp/smartnoise-sdk/blob/677d0bb13ed6f51941739693f3d528eefc6efacf/synth/snsynth/aim/aim.py

AIM (MBI original): https://github.com/ryan112358/mbi/blob/557c077708d3559212a8f65dff3eccd3fd244abb/mechanisms/aim.py

AIM (MBI latest): https://github.com/ryan112358/mbi/blob/master/mechanisms/aim.py

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We genuinely thank all three reviewers for their thoughtful feedback and positive assessment of our work. We are encouraged to see consensus on the paper's clarity and theoretical soundness. In particular, we are happy that all the reviewers highlighted that our new analyses of bounded-range mechanisms are strong theoretical contributions of independent interest, and that our empirical evaluation is thorough and demonstrates consistent utility improvements.

Let us summarize our initial response to the feedback. Based on the suggestions, we will update the manuscript with the following key changes:

  • Change the title from "Tight..." to "Improved...".
  • Add a discussion of the residual-based fully adaptive filter of Regehr et al., 2026, based on our preliminary evaluation.
  • Clarify the goals of our Poisson subsampling conversion via the privacy profile, as it solves the numerical instability of the result based on dominating pairs in Zhu et al.
  • Add a new benchmark comparing our implementation of AIM/AIM-GDP against six open-source AIM versions on the Adult dataset, showing that AIM-GDP achieves the highest utility across all eight implementations.

We provide detailed responses to the individual reviewers below.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.