acceptodds
Under review as a conference paper at ICLR 2027

Final-Model Privacy for Proximal-Anchored DP-SGD

Abstract

Standard R\'enyi differential privacy (RDP) composition charges a record at every access, so its bound grows with the number of fine-tuning updates even when only the final model is released. We study this hidden-state setting for clipped, projected proximal-anchored DP-SGD. Our shifted-R\'enyi PABI certificate separates the discrepancy created by an accessed record from the state difference propagated by later updates. Exact Gaussian smoothing gives dimension-free propagation factors, while a bounded correction at a fixed public reference state reduces the smoothing bias there without changing those factors. For any public access pattern, causal shift allocation computes the certificate in time. If the certified factors contract on every contiguous interval, the final-model RDP bound is independent of and has no separate domain-diameter term. We also give an adaptive anchor rule with a joint bound for its privatized monitor bits and final model. At 500 epochs, the privacy loss is , compared with under step-wise composition.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Differential Privacy, Machine Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_igD4OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies final-model privacy for clipped, projected proximal-anchored DP-SGD.

Critical strengths

The presentation is ok and I beleive with AI assistance the technical proofs should be correct in general.

Critical weaknesses

It is very hard for me to follow the paper, even though I have been working in differential privacy and theory in general for years. I highly suspect the paper is heavily assisted by AI tools, as many steps and description are not standard.

There are many words the authors use are not clear to me. This makes the paper hard to understand. Please refer to the next "Questions For Authors" Section.

Questions for authors

I do not use AI-tools to review, so it is likely I miss some of the details and have grammar mistakes since I am not a native speaker. If so, please just indicate what I miss or where I am wrong. Thank you.

Things that i'd prefer to be clarified:

  • In abstrast, why this is a "hidden-state setting" and what is "proximal-anchored"? Also, for the first time use abbreviations like "PABI", the authors should state what does it mean, even if it is in the abstrast. What are exact Gaussian smoothing and causal shift allocation, are these new techniques by the authors (so that the authors should say they developed new techniques), or existing techniques (so that the authors should say they utilized xxx). The authors say it computes the certificate in O(T) time, what is T?

  • In page 2, what is a "horizon-independent bound"? What is a "diameter"? What is "average factor"? There are other similar unclears things in the first 3 pages, making me unable to follow the paper.

  • What is a "fixed-slot representation"? What is a "propagation factor"?

  • What is a "probe" and "finite probe"?

  • Is this shifted-Renyi divergence in (11) entirely novel? Why is it needed?

  • The authors should not make the fonts in Algorithm 1 substantially smaller than the usual font.

  • What is a "contiguous-interval"?

Overall, the authors used quite a lof of terms that are either undefined or lack of explanation. It is very difficult for me to read the paper.

Moreover, in Ethics Statement the authors say "The proposed watermarking method....." which is clearly not for this paper.

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

The paper consider privacy amplification by iteration (PABI) setting where a convex (possible strongly ocnvex model) is trained via DP gradient descent and privacy guarantees are given only for the final release instead of the whole sequence of updates. In the strongly convex setting, this is intuitive since the one-step map becomes (1−ηλ)(1-\eta \lambda)-contractive, where λ\lambda is the strong convexity parameter (e.g., L2 reg. parameter) and η\eta is the learning rate. This causes forgetting of updates and kind of geometric series argument gives the privacy guarantees for the final model. This paper considers essentially last-layer fine tuning of models such that the L2 regularization is obtained via an anchoring term which pulls the solution to the public features given by the frozed part of the model.

Another contribution of the paper is a certain smoothing of the gradient field. The gradients are evaluated at points θ+τZ\theta + \tau Z, where θ\theta is the model parameter, τ\tau a hyperparameter and Z∼N(0,Id)Z \sim \mathcal{N}(0,I_d). The benefit and aim is to get rid of certain additive terms in the privacy amplification that one has we carrying out clipping of the gradients only (essentially, the random field evaluation is smoothing the gradient field). This obviously comes at the cost of less accurate gradients.

Experimental results illustrate the behavior of the proposed method and its analysis on LLM fine tuning tasks.

Critical strengths
  • The smoothing idea seems interesting as it allows getting rid of certain harmful terms in the existing PABI Rényi DP bounds.
Critical weaknesses
  • I think the main weakness is that there is neither analytical nor experimental evidence for this method leading to improved privacy-utility trade-offs compared to DP-SGD + composition analysis.

  • Appendix C.1 shows what are the accuracies obtained at the initialization with the public features, and those are essentially the same as the test accuracies in Table 3. I.e., in the experiments, there does not seem to be learning happening for this proposed method. Moreover, some of the baseline results in Table 3 have test accuracies that are worse than in the initialization, which makes me question the validity of the experiment altogether.

  • The fact that there does not seem to be any learning happening in the experiments make me suspect that it is very difficult or impossible to adjust the parameters for the method in a way that learning would happen. It is a well-known issue with the existing PABI analyses (see e.g., the best bounds by Bok et al. and their experimental results) that in practice the strong convexity (L2-regularization) has to be so strong for the privacy amplification to kick in that is destroys the learning. The potential benefit is very difficult to attain (again, see experiments by Bok et al.). I suspect what is happening here, is that the condition for L2-regularization parameter λ\lambda (of the anchoring term) given in Eq. (19) forces a condition of the form λ>L\lambda > L, where LL is relatively large, which further forces the solution to stay near the public intialization. At the same time, in high dimensions dd, the random field gradients are evaluated at 2-nrom distance O(d)O(\sqrt{d}) from the solution. This creates some mismatch that prevents learning happening.

Questions for authors
  • Can you comment on the experiments: is it indeed that this method does not improve upon the initialization and why some of the baselines seem to be getting worse?

  • Can you analytically argument, why should there be learning happening and why should the solution leave the neighborhood of the initialization?

  • What is the effect of evaluating the gradients at 2-norm distance O(d)O(\sqrt{d}) from the solution?

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

This paper considers the classical problem of the privacy guarantees of T-step iterative optimization algorithms. They proposed proximal-anchored DP-SGD and derived time independent RDP guarantees of the algorithm. Some light numerical experiments were performed.

Critical strengths

The certification framework is novel and there are technical contributions.

Critical weaknesses

My main concern is the potential application of the algorithm. The implementation uses the approximation (eq 6) instead of the exact expectation, creating a gap between theoretical results and actual algorithm used in practice. Also even from the limited numerical experiments, the proposed algorithm does not seem to show significant advantages over DP-SGD and DP-ADAM.

Questions for authors

The privacy guarantee requires assumptions including (20). How restrictive are these assumptions?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.