acceptodds
Under review as a conference paper at ICLR 2027

Can Variable Noise Be Data-Independent? Differentially Private Bimodal SGD

Abstract

DP-SGD serves as the de facto baseline for private deep learning, yet its constant noise scale imposes an unfavorable dilemma: small noise rapidly depletes privacy budgets, while large noise degrades gradient signals. While adaptive noise mitigates this issue, its data dependence violates essential data independence, risking leakage. We propose Differentially Private Bimodal SGD (DP-BiSGD), a strictly data-independent baseline that stochastically alternates between high-probability small perturbations for local exploitation and low-probability large perturbations for global exploration. To translate this concept into a viable baseline, we address two challenges inherent in this bimodal formulation. Specifically, we develop a bypass-scaled update scheme that shields the momentum buffer from extreme-variance corruption, coupled with a dynamic GMM-PLD accountant to overcome loose standard accounting and ensure tight privacy bounds. Extensive experiments demonstrate that DP-BiSGD consistently outperforms DP-SGD in model utility under identical privacy guarantees, preserves strong resilience against inference attacks, and serves as a seamless drop-in replacement.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
deep learning, differential privacy, privacy loss distribution

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_e8yLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DP-BiSGD, which randomly alternates between frequent small-noise updates and occasional large-noise updates using a data-independent coin. Small-noise steps update momentum normally, while large-noise steps bypass the momentum buffer and receive additional scaling. The intended benefit is to preserve useful gradient information during most iterations without allowing occasional large perturbations to contaminate future updates. A Gaussian-mixture privacy-loss accountant is used to determine the allowable training duration. Experiments report accuracy improvements over DP-SGD, particularly on CIFAR-10, but those comparisons depend critically on whether the proposed accountant correctly bounds privacy loss.

Critical strengths
  1. Combining two noise scales with a momentum bypass is an algorithmic idea worth investigating. The ablations help examine how its components affect utility, although comparisons at matched privacy budgets depend on correcting the accountant.
Critical weaknesses
  1. The one-dimensional privacy reduction in Section 2.4 is not justified by isotropy. Because all coordinates share the same randomly selected noise scale, orthogonal coordinates reveal information about that scale. Projecting them away can reduce distinguishability and cannot simply replace the full-dimensional privacy analysis.
  2. The claimed conservative discretization is incorrect. Mapping privacy losses above LL back to LL decreases them, contradicting Appendix A.1.2’s assertion that the mapped loss is always larger. Consequently, the stated upper bound on privacy failure probability is not established.
  3. The convergence proof analyzes an update without the algorithm’s momentum mechanism and assumes an unbiased gradient. Clipping generally introduces bias, which bounded variance alone does not address. The theorem therefore does not establish convergence for the implemented method.
Questions for authors

Can authors respond the the weakness part? I have used AI tool to do the review since I do not have enough time, but it seems the claims are not supported by correct proofs. I will revise my points if the incorrectness claims turns out to be invalid.

Reviewer_SrneOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes DP-BiSGD to improve the privacy utility trade-off. Instead of using constant Gaussian noise, the method applies small noise in most training steps for accurate optimization and occasionally introduces large noise for exploration. It combines data independent noise selection with bypass scaled updates to protect momentum and a GMM-PLD accountant to track privacy loss.

Critical strengths
  • Addresses a fundamental privacy utility trade-off through a data independent noise mechanism.
  • Demonstrates promising empirical utility improvements, particularly under strict privacy budgets
Critical weaknesses
  • Unjustified dimensionality reduction in privacy accounting. Section 2.4 claims that isotropy makes the privacy leakage of a high-dimensional Gaussian mixture equivalent to that of its 1D projection. However, with a shared mixture component and unequal variances, the likelihood ratio generally depends on coordinates orthogonal to the gradient difference. Consequently, the projected PLD can underestimate the true privacy loss, calling into question the validity of the proposed privacy accountant and its theoretical guarantees.

  • Mismatch between the convergence analysis and the proposed optimizer. The convergence proof in Appendix A.2 replaces the momentum-based update in Eq. (6) with a scaled SGD update, omitting the accumulated momentum term. This simplification is not justified, as the momentum state directly affects both the descent direction and the second-order terms in the smoothness analysis. Consequently, Theorem 3 does not directly establish convergence guarantees for the actual DP-BiSGD algorithm.

  • Invalid stochastic dominance argument in privacy accounting. Equation (14) assigns the unbounded right tail of the privacy loss distribution to a finite terminal bin yK−1=Ly_{K-1}=L. However, the proof of Lemma 1 in Appendix A.1.2 claims that this discretization guarantees Y^i≥Yi\widehat{Y}_i \geq Y_i globally. This claim is incorrect because the Gaussian mixture privacy loss is unbounded above, and any realization Yi>LY_i>L is mapped to Y^i=L<Yi\widehat{Y}_i=L<Y_i. Consequently, the stochastic dominance claim in Eq. (23) does not hold, and the asserted conservative upper bound on privacy leakage is not established by the current proof.

Questions for authors
  • Could the authors provide an ablation study isolating the contribution of the bypass-scaled momentum mechanism, along with results across multiple random seeds?

  • Could the authors rigorously justify the equivalence between the high-dimensional GMM privacy loss and its 1D projection?

  • How does Theorem 3 account for the momentum term omitted from the simplified update in Appendix A.2?

  • How does Lemma 1 guarantee Y^i≥Yi\widehat{Y}_i \geq Y_i when Yi>LY_i>L is mapped to the finite terminal bin?

Reviewer_wEsgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

本文提出 DP-BiSGD,试图改进 DP-SGD 的固定噪声机制:训练时随机选择小噪声或大噪声,大噪声步骤绕过动量缓冲区,并用专门的会计器计算隐私预算。论文报告了部分准确率提升。但其隐私会计是否覆盖完整算法,以及“大噪声促进探索”之类的解释,仍需更充分的证明和实验支持。 The paper proposes DP-BiSGD to improve on the fixed noise level used in DP-SGD. During training, it randomly selects either a low or a high noise level. High-noise steps bypass the momentum buffer, and a dedicated accountant calculates the privacy budget. The paper reports some improvements in accuracy. However, whether the privacy accounting covers the full algorithm, and whether explanations such as “high noise promotes exploration” hold, require more rigorous proof and experimental support.

Critical strengths

方法设计具有整体性。 作者不仅提出独立于数据的双档噪声选择,还同时考虑大噪声对动量状态的影响及累计隐私损失的计算,使算法设计思路较完整。 提供了初步的实证支持。 在报告的 CIFAR-10 实验中,DP-BiSGD 在四档隐私预算下均取得高于 DP-SGD 的准确率;将机制用于 DP-SGD-HF 后,也观察到准确率提升。这些结果使方法值得进一步检验 1.The method is designed as a coherent whole. The authors propose a data-independent choice between two noise levels while also considering how high-noise steps affect the momentum state and how to account for cumulative privacy loss. This makes the algorithmic design reasonably complete. 2.The paper provides preliminary empirical support. In the reported CIFAR-10 experiments, DP-BiSGD achieves higher accuracy than DP-SGD at all four privacy budgets. Applying the mechanism to DP-SGD-HF also yields an accuracy improvement. These results make the method worthy of further investigation.

Critical weaknesses

隐私保证的论证尚不充分。 论文用一维 GMM-PLD 计算隐私损失,但噪声分支同时作用于高维梯度,并决定后续更新方式;现有推导尚未说明该会计结果为何覆盖完整算法。有限网格的尾部处理和 FFT 卷积的误差界也需要澄清。这直接影响“相同隐私预算下更优”的结论。 实验未能验证所声称的优化机制。 参数分析与最终准确率不能区分双档噪声、动量旁路及可训练步数各自的作用,也不能直接证明大噪声帮助模型跳出不佳区域。论文缺少针对这些解释的消融实验和训练过程数据。 1.The privacy guarantee is not yet sufficiently justified. The paper uses a one-dimensional GMM-PLD to calculate privacy loss, but the noise branch acts on high-dimensional gradients and also determines how subsequent updates are performed. The current derivation does not establish why this accounting result covers the full algorithm. The treatment of the tails on a finite grid and the error bounds for FFT convolution also need clarification. These issues directly affect the claim of better performance “under the same privacy budget.” 2.The experiments do not verify the claimed optimization mechanism. The parameter analyses and final accuracy results cannot separate the effects of the two noise levels, the momentum bypass, and the number of trainable steps. Nor do they directly show that high noise helps the model escape unfavorable regions. The paper lacks ablation experiments and training-trajectory data to support these explanations.

Questions for authors

能否证明隐私会计器计算的 ϵ\epsilon 是完整高维、分支依赖更新算法的有效上界。 特别是,一维投影、无界右尾的离散化及 FFT 卷积误差分别如何处理? 能否在可比的隐私预算下提供关键消融实验? 例如分别去掉大噪声分支和动量旁路,并报告各设置的实际训练步数、λlarge\lambda_{\mathrm{large}} 及训练轨迹,以检验所提出的机制解释。 文中所说自适应方法的“隐私泄露风险”具体指哪些方法及哪种未经计账的数据使用? 对已经采用私密统计估计和正式隐私会计的方法,DP-BiSGD 的相对优势是什么? 1.Can the authors prove that the ϵ\epsilon computed by the privacy accountant is a valid upper bound for the complete high-dimensional algorithm with branch-dependent updates? In particular, how are the one-dimensional projection, discretization of the unbounded right tail, and FFT convolution error handled? 2.Can the authors provide key ablation experiments under comparable privacy budgets? For example, they could remove the high-noise branch and the momentum bypass separately, and report the actual number of training steps, λlarge\lambda_{\mathrm{large}}, and training trajectories for each setting to test the proposed explanation of the mechanism. 3.What specific methods and what forms of unaccounted-for data use does the paper mean by the “risk of privacy leakage” in adaptive methods? Compared with methods that already use private statistical estimation and formal privacy accounting, what is the relative advantage of DP-BiSGD?

Reviewer_tbT3OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes DP-BiSGD, a data-independent variable-noise variant of DP-SGD that alternates between small- and large-variance Gaussian perturbations using an exogenous Bernoulli variable. Large-noise steps bypass the momentum buffer to avoid contaminating subsequent updates. A Gaussian-mixture privacy-loss-distribution accountant is introduced based on discretization and FFT composition, derive privacy and optimization guarantees, and evaluate classification utility on Fashion-MNIST and CIFAR-10.

Critical strengths

(1) Optimizer mechanism for separating high-noise steps from momentum. The bypass-scaled update directly addresses the persistence of large perturbations in an exponential moving-average buffer. (2) The potentially useful privacy-accounting question can be solved. Treating the noise-scale indicator as hidden and analyzing the resulting mixture mechanism is an interesting direction, provided the full high-dimensional privacy guarantee can be established.

Critical weaknesses

(1) The high-dimensional privacy accounting appears to rely on an unjustified one-dimensional reduction. (2) The convergence proof may not analyze the algorithm that is actually proposed

Questions for authors

(1) For the high-dimensional mechanism, is the same Bernoulli variable shared across all coordinates of each noise vector? (2) The full privacy-loss derivation for the vector-valued mechanism is missing. (3) How does the FFT-based PLD composition prevent circular-convolution wrap-around, and how are finite-grid truncation and tail mass handled conservatively?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.