acceptodds
Under review as a conference paper at ICLR 2027

Differentially Private Fine-Tuning in Learned Subspaces of LoRA

Abstract

Parameter-efficient fine-tuning (PEFT) has emerged as an effective paradigm for adapting large language models to downstream tasks involving privacy-sensitive data. Differential privacy (DP) mitigates privacy leakage during fine-tuning but often introduces substantial utility degradation. Focusing on low-rank adaptation (LoRA), a representative PEFT method, we observe that although LoRA constrains the rank of weight updates, its parameterization remains highly redundant, which can result in large noise and impaired fine-tuning utility. To address this limitation, we propose a two-stage framework for differentially private fine-tuning within a learned low-dimensional subspace of LoRA parameters. Our method first identifies task-relevant update subspaces by extracting dominant directions from LoRA training trajectories, and then reparameterizes LoRA factors to perform private fine-tuning directly in this compact subspace. By restricting noise-perturbed updates to a low-dimensional, task-aligned space, the proposed method reduces the effective impact of DP noise while preserving the same privacy guarantees. We theoretically characterize how subspace-based updates reduce the effective magnitude of injected DP noise compared to high-dimensional parameter updates. Extensive experiments on multiple benchmark datasets demonstrate that our method consistently outperforms representative differentially private fine-tuning baselines across a broad range of privacy settings. Our code is available at https://anonymous.4open.science/r/8BA2.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Differential privacy, Large language models, Low-rank adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_tfTKOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper proposes a two-stage framework for differentially private fine-tuning in learned subspaces of LoRA parameters. By restricting private optimization to low-dimensional subspaces, their method reduces the effective impact of differential privacy noise while preserving formal privacy guarantees. They provide a theoretical interpretation showing that subspace optimization is equivalent to gradient projection–based optimization and leads to reduced noise magnitude. Extensive experiments on multiple benchmarks demonstrate consistent improvements over existing DP baselines, especially under strict privacy budgets. The authors further show that learned subspaces can transfer across datasets from related task, enabling practical private fine-tuning when target trajectories are unavailable.

Critical strengths

The paper combines efficient training with a focus on differential privacy by proposing a two-stage framework. Private optimization is restricted to low-dimensional subspaces, which is effective in reducing noise introduced by differential privacy. They have proposed a theoretical interpretation of their choice. Empirical results are strong.

Critical weaknesses

The paper is generally hard to read. There is a strong technical contribution aspect that is loosely connected with the motivation. Therefore, the paper is not written for a general audience but for an expert reader in the field.

Questions for authors
  • "Parameter-efficient fine-tuning (PEFT) has emerged as an effective paradigm for adapting large language models to downstream tasks involving privacy-sensitive data." Since when is PEFT proposed for privacy-sensitive data? The abstract does not read well from the beginning. Many problems are introduced in parallel, such as privacy leakage, utility degradation, noise and impaired utility, etc., making us wonder what specific problems the paper is focusing on.
2 comments

We respectfully disagree the questions and would like to clarify the following points.

1. Clarification of PEFT and Privacy

PEFT was not originally proposed specifically for privacy-sensitive data, nor does PEFT itself provide formal privacy guarantees. Our statement is intended to emphasize that PEFT has become a widely adopted approach for adapting large language models to downstream tasks, including scenarios involving sensitive data. In such scenarios, additional privacy-preserving mechanisms, such as DP-SGD, are required to provide formal privacy guarantees. Our work investigates how to improve the effectiveness of PEFT under these differential privacy constraints, rather than treating PEFT itself as a privacy-preserving mechanism.

2. Clarification of the Central Research Problem

The central problem addressed in our paper is the utility degradation caused by differentially private optimization, particularly under strict privacy budgets. Privacy leakage motivates the need for differential privacy, while gradient clipping and noise injection are mechanisms used in DP-SGD that can impair optimization. These are therefore not separate research objectives, but interconnected aspects of the same privacy–utility trade-off.

Our two-stage framework is designed to address this problem directly. In the first stage, we leverage public optimization trajectories of LoRA parameters to identify task-relevant low-dimensional subspaces without consuming any privacy budget. In the second stage, private optimization is restricted to these learned subspaces, reducing the effective dimensionality of the injected noise while maintaining formal differential privacy guarantees. This design aims to improve model utility under a given privacy budget, as supported by our theoretical analysis and experimental results.

Thanks for the clarification. The first line of the abstract still starts with a confusing connection between PEFT and privacy-sensitive data, which is wrong. I would appreciate the authors reflecting the additional clarifications in the paper.

Reviewer_qzgEOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a two-stage framework for differentially private (DP) fine-tuning that restricts LoRA updates to a learned low-dimensional subspace. In the first stage, public data is used to collect LoRA trajectory and compute SVD to extract bases for the AA and BB LoRA matrices. In the second stage, the found bases are used to reparameterize the LoRA matrices and run DP-SGD on the lower dimensional parameters. This dimensionality reduction reduces the effective DP noise used by DPSGD. Experimental results show the effectiveness of the proposed idea.

Critical strengths
  • The paper is clear and well-written with a clear motivation.
  • Experimental results have been proposed to show the effectiveness of the proposed idea.
Critical weaknesses
  • The proposed idea has been proposed previously in the literature, and extensively used [1,2].
  • The SVD decomposition step for extracting the ambient update directions is computationally expensive and runs counter to the LoRA's foundational design goal of maintaining low computational overhead.
  • The utility improvement introduced by the method is marginal.

[1]: Yingxue Zhou, Steven Wu, Arindam Banerjee: Bypassing the Ambient Dimension: Private SGD with Gradient Subspace Identification (ICLR 2021)

[2]: Xin Gu, Gautam Kamath, Zhiwei Steven Wu: Choosing Public Datasets for Private Machine Learning via Gradient Subspace Distance (2025)

Questions for authors
  • How is privacy accounting handled for the split between θA\theta_A and θB\theta_B? Is the mechanism treated as a single joint release with a concatenated clipped vector (sensitivity CC) and a block-diagonal Gaussian covariance (thus a single ϵ\epsilon per step), or as two mechanisms composed per step? Please formalize this to avoid ambiguity.
  • How many snapshots TT do you collect, at what intervals, and how sensitive are results to TT?
  • In order to avoid the SVD, can we simply reduce the adaptation rank of LoRA in order to reduce the number of parameters? For example, if the proposed method reduces the number of parameters by a factor of 4, can we simply reduce the rank from rr to r/4r/4 and compare the results?
4 comments

We thank the reviewer for the comments and address the concerns as follows.

Response to Weakness 1

We thank the reviewer for highlighting these relevant studies. We clarify that our work differs from the cited approaches in both its research objective and optimization mechanism.

Regarding Zhou et al. [1], their Projected DP-SGD estimates gradient subspaces from public data and explicitly projects noisy gradients onto these subspaces. Importantly, their framework periodically recomputes the gradient eigenspace during private training, introducing recurring computational overhead. As discussed in their original paper, reducing the frequency of eigenspace computation can alleviate this cost but may lead to some performance degradation. In contrast, our method constructs task-relevant subspaces from public LoRA optimization trajectories through a one-time SVD before private training. The learned bases remain fixed throughout Stage 2, and DP-SGD directly optimizes the low-dimensional subspace parameters without repeated subspace estimation or explicit projection of full-dimensional LoRA gradients.

Regarding Gu et al. [2], their primary contribution is Gradient Subspace Distance (GSD), which measures the alignment between public and private datasets to guide public dataset selection. Our work addresses a different problem: mitigating the impact of privacy noise arising from redundancy in LoRA parameterization. Rather than evaluating or selecting public datasets, we develop a two-stage fine-tuning framework that identifies task-relevant directions from public LoRA trajectories and restricts subsequent private optimization to the learned parameter subspaces.

In summary, our work differs substantially from both cited studies in its research objective, subspace construction, and private optimization mechanism. We will further clarify these distinctions and provide a more detailed discussion of the related work in the revised manuscript.

Response to Weakness 2

We thank the reviewer for raising this concern. As described in Section 4.1 and Algorithm 1, our method performs SVD only once during public subspace construction, rather than repeatedly during private fine-tuning. Moreover, SVD is applied to the optimization trajectories of LoRA factors rather than the full pretrained weight matrices.

To quantify this cost, we conducted an additional efficiency analysis comparing our method with DP-LoRA under identical experimental settings, using the same backbone, dataset, hardware, and training configuration. The SVD operation takes only 14.6 seconds, accounting for approximately 2.29% of our total end-to-end training time. The peak GPU memory usage over the complete training process is 2222 MiB for our method versus 1938 MiB for DP-LoRA, representing a modest increase of 14.65%. These results indicate that SVD itself is not a major computational bottleneck. We will clarify this computational trade-off in the revised manuscript.

Response to Weakness 3

We would like to emphasize that our method achieves consistent improvements across different model scales, downstream tasks, and privacy budgets, with particularly noticeable gains under stringent privacy constraints.

As shown in Table 1 and Appendix Table 7, under ε=0.5\varepsilon = 0.5, our method achieves average improvements of 1.79% and 2.05% over the strongest DP baseline for each reported metric on RoBERTa-base and RoBERTa-large, respectively. In particular, on RoBERTa-large, our method improves QNLI by 2.74%, MNLI-m by 2.75%, and MNLI-mm by 2.58%. These improvements are consistently observed across the evaluated tasks rather than being limited to isolated cases.

Furthermore, Table 2 demonstrates that these benefits extend to generative tasks. On Qwen2.5-1.5B with SAMSum, our method improves ROUGE-2 from 0.2177 to 0.2289 under ε=0.5\varepsilon = 0.5, compared with the strongest DP baseline.

While the magnitude of improvement varies across tasks, we believe these consistent gains, particularly under strict privacy budgets, demonstrate meaningful improvements in the privacy–utility trade-off.

[1]: Yingxue Zhou, Steven Wu, Arindam Banerjee: Bypassing the Ambient Dimension: Private SGD with Gradient Subspace Identification (ICLR 2021)

[2]: Xin Gu, Gautam Kamath, Zhiwei Steven Wu: Choosing Public Datasets for Private Machine Learning via Gradient Subspace Distance (2025)

Response to Q1

We thank the reviewer for this question. As already described in Section 4.2 (Eqs. (9)–(10)) and Appendix A.2, our method adopts a single joint Gaussian mechanism, rather than treating θA\theta_A and θB\theta_B as two separately composed mechanisms.

Specifically, for each training example, we concatenate the gradients of θA\theta_A and θB\theta_B into a single vector and jointly clip its ℓ2\ell_2 norm to CC. After clipping, Gaussian noise with covariance σ2C2I\sigma^2 C^2 I is added to the aggregated joint gradient, which is equivalent to independent Gaussian noise on the two blocks with a block-diagonal covariance matrix. Thus, each training step is accounted for as one joint DP-SGD mechanism, and the cumulative privacy budget (ε,δ)(\varepsilon,\delta) is tracked using the standard RDP accountant implemented in Opacus.

Response to Q2

As described in Section 4.1 and Algorithm 1 in Appendix B.1, we periodically collect LoRA parameter displacements relative to initialization, ΔMt=Mt−M0\Delta M_t = M_t - M_0, for M∈A,BM \in \\{A,B\\}. The snapshots are recorded at regular intervals of y=⌊T1/T⌋y = \lfloor T_1/T \rfloor optimization steps during public warm-up, where T1T_1 denotes the total number of warm-up steps and TT denotes the target number of snapshots.

To further examine sensitivity to TT, we conducted additional experiments on RoBERTa-base using SST-2 and QNLI under ε=0.5\varepsilon = 0.5, with rsub=32r_{\mathrm{sub}}=32.

DatasetT = 40T = 50T = 60
SST-291.7691.5491.75
QNLI88.1988.3287.89

The results remain relatively stable across different snapshot counts, with maximum variations of only 0.22 and 0.43 percentage point on SST-2 and QNLI, respectively. These findings suggest that our method is relatively insensitive to the number of snapshots within the evaluated range. We will include these additional results in the revised manuscript.

Response to Q3

We thank the reviewer for this question. We would like to clarify that directly reducing the LoRA rank and our proposed subspace reparameterization impose fundamentally different constraints on the optimization space. Moreover, matching the number of trainable parameters through LoRA rank reduction is not feasible under our experimental settings.

Specifically, for a weight matrix of size dout×dind_{\mathrm{out}} \times d_{\mathrm{in}}, standard LoRA requires rLoRA(din+dout)r_{\mathrm{LoRA}}(d_{\mathrm{in}}+d_{\mathrm{out}}) trainable parameters, whereas our method optimizes only 2rsub2r_{\mathrm{sub}} subspace parameters while retaining the original LoRA rank rLoRAr_{\mathrm{LoRA}}.

For example, on RoBERTa-base, with din=dout=768d_{\mathrm{in}}=d_{\mathrm{out}}=768, rLoRA=32r_{\mathrm{LoRA}}=32, and rsub=32r_{\mathrm{sub}}=32, standard DP-LoRA requires 49,152 trainable parameters per adapted projection, while our method requires only 64. Even reducing DP-LoRA to rank 1 still requires 1,536 parameters, making exact parameter matching infeasible through rank reduction.

More importantly, reducing the LoRA rank directly restricts the maximum rank of the weight update, whereas our method preserves the original LoRA rank and constrains optimization to task-relevant directions learned from public trajectories. Thus, the two approaches differ fundamentally in how they reduce the effective optimization space.

Reviewer_qzgEOpenReview ↗

Official comment

Thanks for the responses. I still strongly believe that the idea of computing gradient subspaces from public data and using them to diminish the effect of DP noise has been proposed and used in the literature. A slight difference in the implementation of the idea does not lead to sufficient novelty and originality. Furthermore, the work by Zhou et al. [1], which is very close to the current work in its key components, should have been used as a baseline. Overall, I strongly believe that the proposed idea and its key components already exist in the literature and therefore it lacks novelty and originality. Hence, I keep my score.

We thank the reviewer for the follow-up and would like to clarify one important technical distinction. Our method does not construct gradient subspaces from public data. Instead, it learns a low-dimensional parameter subspace from public LoRA optimization trajectories through a one-time SVD and subsequently performs differentially private optimization through an explicit low-dimensional reparameterization. In contrast, Zhou et al.[1] identify principal gradient directions from public gradient statistics and use them for projected DP-SGD. While both methods share the broader motivation of mitigating DP noise through low-dimensional optimization, they differ in their subspace construction, parameterization, and optimization mechanisms, as detailed in our earlier response. We respectfully hope that these distinctions will be considered in the final assessment

Comments on the paper

2 comments

We thank the reviewer for the comments and address the concerns as follows.

Response to Weakness 1 and Question 1

We would like to clarify several aspects of our experimental protocol, particularly regarding the DP-LoRA baseline.

First, DP-LoRA is already included in our main experiments, as explicitly reported in Table 1 (RoBERTa-base), Table 2 (Qwen2.5-1.5B), and Table 7 (RoBERTa-large). Importantly, all methods have access to the same public auxiliary data and undergo the same non-private warm-up procedure before private fine-tuning. Specifically, all methods perform 3 epochs of public warm-up, after which our method constructs the subspace from the recorded optimization trajectories, while DP-LoRA continues private fine-tuning without subspace constraints. Therefore, the comparison already accounts for the task-specific information and non-private warm-up opportunity raised by the reviewer.

Second, we clarify the experimental configurations as follows:

  • Data partition: Each original training set is randomly shuffled and partitioned into 10% public auxiliary data DauxD_{\mathrm{aux}} and 90% private training data DprivD_{\mathrm{priv}}, with no overlap.
  • Training: All methods use a batch size of 32 and 3 public warm-up epochs. In the private stage, our method trains for 1 epoch, whereas the baselines train for 3 epochs.
  • Snapshots: We record 40 snapshots for rsub=32r_{\mathrm{sub}}=32 and 70 snapshots for rsub=64r_{\mathrm{sub}}=64.
  • Privacy mechanism: Private training employs Poisson subsampling. The noise multiplier is automatically calibrated using the Opacus privacy accountant according to the target privacy budget, sampling rate, and number of private optimization steps.
  • Random seeds: All experiments use seeds 42, 43, and 44.

We acknowledge that these implementation details, particularly the shared public warm-up protocol, could have been stated more explicitly. We will clarify them in the revised manuscript.

Response to Weakness 2 and Questions 2–3

We thank the reviewer for carefully examining our methodology and implementation. We clarify the parameterization, initialization, target modules, and additional empirical evidence as follows.

1. Parameterization and Initialization.

The correct Stage-2 parameterization is M=M0+PMθMM=M_0+P_M\theta_M, as stated in Eq. (8). The omission of M0M_0 in Algorithm 2 and the corresponding reconstruction expressions is a writing error, which we acknowledge and will correct.

The actual initialization first projects the Stage-1 parameter displacement onto the learned subspace:

θ~M=PM⊤flat⁡(Mf−M0),M∈A,B.\tilde{\theta}_M=P_M^\top\operatorname{flat}(M_f-M_0),\quad M\in\\{A,B\\}.

The projected coordinates are then concatenated and jointly normalized:

\theta^{(0)}=10^{-6}\frac{\tilde{\theta}}{\\|\tilde{\theta}\\|_2}.$$ **This initialization is intentional:** it preserves the task-relevant directional information learned during public warm-up while keeping the initial parameter displacement extremely small. Consequently, Stage 2 starts close to the original LoRA initialization rather than directly inheriting the magnitude of the public warm-up updates. This design reduces the influence of inherited parameter displacement while retaining the learned subspace structure. We acknowledge that the normalization step was omitted from the manuscript and will explicitly document the complete initialization procedure. **2. LoRA Target Modules.** Our implementation applies LoRA exclusively to the query, key, value, and attention output projections (Q/K/V/O). Neither feed-forward networks nor other dense layers are included. As already specified in Appendix D.3, *dense* refers specifically to the attention output projection. **3. Noise Analysis and Retained Update Energy.** Figure 3 presents the analytically computed expected noise norm based on $\mathbb{E}[\\|\eta\\|_2]\approx\sigma C\sqrt{m}$, rather than measurements collected during training. To further investigate this trade-off, we evaluate cumulative singular-value energy, projected-update energy retention, and norm-based signal-to-noise ratio on MNLI using RoBERTa-base at $\varepsilon=0.5$. | $r_{\mathrm{sub}}$ | Cumulative SV Energy | Projected-Update Energy | Expected Noise Norm | Norm-Based SNR | |:---:|:---:|:---:|:---:|:---:| | 8 | 99.92% | 97.92% | 1.52 | 2.04 | | 16 | 99.97% | 98.75% | 2.14 | 1.17 | | 32 | 99.99% | 99.01% | 3.04 | 0.79 | The results show that the learned subspaces preserve a substantial proportion of the measured update energy, reaching 99.01% at $r_{\mathrm{sub}}=32$. Meanwhile, the increasing expected noise norm illustrates the trade-off between retaining optimization signal and reducing privacy-induced perturbation. **4. Gradient Clipping Analysis.** We additionally conduct a controlled comparison between DP-LoRA and our method using RoBERTa-base on SST-2. Both methods use LoRA rank 32, logical batch size 32, physical batch size 8, 3 public warm-up epochs, and 1 private training epoch. We measure the average clipping retention as $$\mathrm{Retention}=\frac{1}{N}\sum_{i=1}^{N}\min\left(1,\frac{C}{\\|g_i\\|_2}\right).$$ The measured values are 44.29% for DP-LoRA and 43.05% for our method, differing by only 1.24 percentage points. This indicates comparable average gradient-norm retention under the evaluated conditions, without suggesting that our method benefits from reduced clipping severity. Instead, the learned subspace provides a distinct mechanism for restricting the dimensionality of privacy noise while retaining informative update directions. Finally, non-private Full-FT results are already reported in Tables 1 and 2 as reference performance without privacy constraints, although they are not parameterization-matched zero-noise controls. We will correct the methodological inconsistencies and incorporate these additional empirical analyses and implementation details into the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.