acceptodds
Under review as a conference paper at ICLR 2027

DP-LAIR: Differentially Private LLM Inference with Adaptive Routing and Latent Perturbation

Abstract

Differential privacy (DP) provides a principled framework for protecting sensitive data, where its applications in machine learning have largely focused on training-time privacy. However, Language-model APIs currently can expose information from both protected training records and query time context. Thus, We study the joint protection of these two sources under a practical trusted-server, API-only deployment model, where internal model states remain hidden and private information becomes externally observable only through the released transcript. We propose Differentially Private Latent Adaptive Inference and Routing (DP-LAIR), a unified framework for protecting both training-time and runtime private information at this interface. DP-LAIR introduces a router that directs tokens that can be generated from public information to a public path. For tokens requiring private context, their hidden-state residual is compressed into a latent representation, clipped and perturbed with Gaussian noise, and mapped back to the output space by a decoder. The router and decoder are trained with differential privacy to protect training-time data, while runtime privacy is adaptively accounted for using data-dependent leave-one-out sensitivity. Experiments across multiple datasets and models support DP-LAIR’s protection of both training data and query time context while demonstrating good generation utility.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM inference, Differential Privacy

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_Lb5tOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This research proposes DP-LAIR, a framework for jointly protecting training data and inference-time private context in LLMs. The method combines adaptive public/private routing, latent-space Gaussian perturbation, and differentially private training of the router and decoder. Experiments on TOFU and MUSE demonstrate improved privacy-utility trade-offs compared with independently composed DP mechanisms.

Critical strengths
  1. The paper addresses an important and practical problem by considering both training-time and inference-time privacy within a unified framework.
  2. The adaptive routing mechanism and latent-space perturbation provide an interesting approach to balancing privacy protection and generation utility.
Critical weaknesses
  1. Unclear training-data privacy guarantee. Stage I trains the encoder and covariance using the private training corpus, therefore Stage II should only applies DP training to the router and decoder. Moreover, the private backbone is adapted using downstream data. It is unclear how the overall training-data privacy guarantee holds when these components may already contain information from the protected corpus.
  2. Insufficient analysis of inference overhead. The proposed LOO privacy accounting requires evaluating multiple context variants during generation but lacks of the detailed measurements of inference latency and computational cost.
  3. The experiments primarily rely on perplexity to measure generation utility. Additional task-level metrics, such as QA accuracy or F1, would help demonstrate whether the method preserves the ability to answer questions requiring private information
Questions for authors
  1. How does the training-data DP guarantee account for the Stage-I encoder, learned covariance, and downstream-adapted private backbone?
  2. Does the LOO-based privacy accounting introduce substantial computational overhead for longer private contexts?
1 comment
AuthorsOpenReview ↗

Response to Reviewer Lb5t

Thanks for your question, we would like to answer your question one by one:

Response to W1 and Q1

We would like to clarify that our paper explicitly defines Infer-RDP under the API-only threat model, as emphasized in the Introduction. The adversary observes only the released API transcript, without access to model parameters, hidden states, or internal components. Crucially, Infer-RDP is defined conditional on a shared and fixed deployment state ωf\omega_f, where Phase 1 outputs (e.g., the trained encoder and covariance) are treated as fixed internal components. As specified in Section 3.1 (p. 3, lines 157–159) and formalized in Lemma 3.2, the transcript distribution PSx,SyP_{S_x,S_y} is conditioned on this fixed state. This reflects practical API deployments, where internal components remain fixed and inaccessible to users A similar separation is adopted in Sampling and Aggregation [1], where intermediate computations need not be DP, while the released aggregate is protected by the DP mechanism. We provide a new Figure to highlight threat model difference in introduction in the revised version.

In Stage I, we non-privately train the encoder and covariance using SxS_x. These components, together with the private backbone, are then frozen before Stage II and treated as part of the fixed deployment state ωf\omega_f. Although they remain involved in inference, their state is shared across neighboring Stage-II executions. Importantly, this assumption does not require these components to be independent of their historical training data. In Stage II, we retain the fixed Stage-I components and retrain the router and decoder using the DP-guaranteed DPDT algorithm. Conditioned on the shared fixed state ωf\omega_f, this guarantee transfers to the released API transcript through post-processing, as established in Appendix A.1.

Therefore, Our training-data privacy guarantee protects the Stage-II use of SxS_x, conditional on the shared fixed Stage-I state, rather than its historical non-private use in Stage I. We will make this distinction more explicit in Definition 3.1 and the revised Introduction.

[1] Nissim K, Raskhodnikova S, Smith A. Smooth sensitivity and sampling in private data analysis[C]//Proceedings of the thirty-ninth annual ACM symposium on Theory of computing. 2007: 75-84.

Response to W2 and Q2

Thanks for your suggestion, The inference latency of our method depends on the number of records in the private prompt, as the LOO privacy accounting requires evaluating multiple context variants. We note that additional computational overhead is also common in differentially private next-token generation[1,2], where multiple generations or sampling operations may be required to obtain a DP-guaranteed output. Therefore, such overhead is not unique to our method. We clarify the relationship between the inference overhead and the private prompt size in the revised manuscript. In the revised version, we highlight that the inference latency depends on the size of SyS_y. Equation (8) highlights the privacy computation trade off: obtaining a tighter privacy bound through LOO, rather than using a looser bound based on direct clipping, comes at additional computational cost. Specifically, it requires ∣Sy∣+1|S_y|+1 inference evaluations per token, with the overhead increasing linearly with the size of SyS_y.

[1] Vinod V, Pillutla K, Guha Thakurta A. Invisibleink: High-utility and low-cost text generation with differential privacy[J]. Advances in Neural Information Processing Systems, 2026, 38: 67254-67298.

[2] Thareja R, Nakov P, Vepakomma P, et al. Dp-fusion: Token-level differentially private inference for large language models[C]//International Conference on Learning Representations. 2026, 2026: 151088-151120.

Response to W3:

Thank you for your question. We choose a generative model because the output space of QA tasks is much smaller than that of text generation. In this case, directly applying the exponential mechanism to select an answer would be a more straightforward approach. Therefore, we use perplexity (PPL) to evaluate generation utility, as it is one of the most widely used metrics for differentially private text generation [1,2]. Moreover, we also provide more metrics such as PII exact match and LLM as a judge method to further evaluate the empirical privacy protection performance. Please refer to Response to Reviewer 2C98 for detailed result.

[1] Flemings J, Razaviyayn M, Annavaram M. Differentially private next-token prediction of large language models[C]//Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024: 4390-4404.

[2] Thareja R, Nakov P, Vepakomma P, et al. Dp-fusion: Token-level differentially private inference for large language models[C]//International Conference on Learning Representations. 2026, 2026: 151088-151120.

Reviewer_2C98OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DP-LAIR proposes a two-path API layer for protecting downstream training data S_x and query-time private context S_y. A public-only router handles tokens that do not need private context, while other tokens pass through a clipped, Gaussian-noised latent residual; DPDT trains the router and decoder. Experiments on TOFU and MUSE News compare DP-LAIR with Composition and other baselines, reporting lower reproduction leakage in many settings. However, the privacy proof does not clearly cover every deployed parameter that depends on S_x.

Critical strengths
  1. The paper presents a clear API-only threat model that jointly considers downstream training data and query-time context.

  2. The experiments cover two datasets, four model configurations, a joint baseline, utility and leakage metrics, membership inference, and ablations.

Critical weaknesses
  1. Stage I trains the encoder and covariance from S_x, then Stage II freezes them. However, Lemma 3.2 assumes the fixed state is independent of S_x and that S_x affects the transcript only through the DP-trained router and decoder. The paper gives no DP guarantee for the encoder or covariance, so the stated training-data bound does not hold for the full deployed state.

  2. The runtime proof covers only full-context versus null-record LOO neighbors. It does not establish the claimed bound for replacing one record with an arbitrary record, so the intended adjacency relation must be narrowed or the proof extended.

Questions for authors
  1. In the actual DP-LAIR deployment, are the private backbone, phi_enc, and Sigma trained using S_x? If yes, what mechanism and accountant protect the Stage-I dependence before these parameters are frozen? If not, what data are used instead, and how does that differ from the training examples in Section 3.3.1 and Appendix B.3?
  2. Does the null record exactly replace the intended S_y adjacency, or does it include replacing one record with an arbitrary record or an add-remove adjacency? Please state the formal relation used for the headline guarantee and provide the corresponding sensitivity and accounting if the broader relation is intended.
  3. Could the empirical leakage audit be extended to semantic paraphrases and adaptive API queries beyond the sentence-reproduction metric used in Section 5.1 and Appendix C.5?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 2C98

Thanks for your question, we would like to answer your question one by one

Response to W1 ans Q1

We would like to clarify that our paper explicitly defines Infer-RDP under the API-only threat model, as emphasized in the Introduction. The adversary observes only the released API transcript, without access to model parameters, hidden states, or internal components. Crucially, Infer-RDP is defined conditional on a shared and fixed deployment state ωf\omega_f, where Phase 1 outputs (e.g., the trained encoder and covariance) are treated as fixed internal components. As specified in Section 3.1 (p. 3, lines 157–159) and formalized in Lemma 3.2, the transcript distribution PSx,SyP_{S_x,S_y} is conditioned on this fixed state. This reflects practical API deployments, where internal components remain fixed and inaccessible to users A similar separation is adopted in Sampling and Aggregation [1], where intermediate computations need not be DP, while the released aggregate is protected by the DP mechanism. We provide a new Figure to highlight threat model difference in introduction in the revised version.

In Stage I, we non-privately train the encoder and covariance using SxS_x. These components, together with the private backbone, are then frozen before Stage II and treated as part of the fixed deployment state ωf\omega_f. Although they remain involved in inference, their state is shared across neighboring Stage-II executions. Importantly, this assumption does not require these components to be independent of their historical training data. In Stage II, we retain the fixed Stage-I components and retrain the router and decoder using the DP-guaranteed DPDT algorithm. Conditioned on the shared fixed state ωf\omega_f, this guarantee transfers to the released API transcript through post-processing, as established in Appendix A.1.

Therefore, Our training-data privacy guarantee protects the Stage-II use of SxS_x, conditional on the shared fixed Stage-I state, rather than its historical non-private use in Stage I. We will make this distinction more explicit in Definition 3.1 and the revised Introduction.

[1] Nissim K, Raskhodnikova S, Smith A. Smooth sensitivity and sampling in private data analysis[C]//Proceedings of the thirty-ninth annual ACM symposium on Theory of computing. 2007: 75-84.

Response to W2 and Q2

Thanks for your question, we have emphasized that our paper considers replacing one record with a null record. (line 185, line 245247, line 317319). Meanwhile, replacing one record with a null record turns to be a default setting in many differential private papers.[1,2]

[1] Abadi M, Chu A, Goodfellow I, et al. Deep learning with differential privacy[C]//Proceedings of the 2016 ACM SIGSAC conference on computer and communications security. 2016: 308-318.

[2] Zhu Y, Wang Y X. Poission subsampled rényi differential privacy[C]//International Conference on Machine Learning. PMLR, 2019: 7634-7642.

Response to Q3

We report the PII exact match result and LLM as a judge result as follow. For PII Exact Match, we extract PII from the private prompt and check whether it appears in the generated response. For LLM-as-a-Judge, we use GPT-6 to assess whether the response leaks private information. The empirical result shows that both Composition and DP-LAIR provide DP protection. However, as we reported in our paper, DP-LAIR provide better PPL and smaller average privacy loss.

PII exact match:

DatasetModelNo protectionPublicCompositionDP-LAIR
TOFULlama 3B74.5%0%0%0%
TOFULlama 8B95.0%0%0%0%
TOFUQwen 3B26.0%0%0%0%
TOFUQwen 7B35.0%0%0%0%
MUSE NewsLlama 3B98.5%0.5%1.0%3.0%
MUSE NewsLlama 8B99.5%1.0%0.5%5.0%
MUSE NewsQwen 3B91.0%1.0%1.5%1.5%
MUSE NewsQwen 7B94.0%1.0%1.0%5.0%

LLM-as-a-Judge

DatasetModelNo protectionPublicCompositionDP-LAIR
TOFULlama 3B88.5%2.5%4.5%6.0%
TOFULlama 8B97.0%7.0%4.0%4.0%
TOFUQwen 3B76.0%5.0%5.0%6.5%
TOFUQwen 7B78.0%3.5%5.0%6.5%
MUSE NewsLlama 3B97.5%13.0%10.0%12.5%
MUSE NewsLlama 8B98.0%19.5%14.0%20.0%
MUSE NewsQwen 3B96.0%9.0%7.0%8.5%
MUSE NewsQwen 7B96.0%9.5%7.0%10.0%
AuthorsOpenReview ↗

[Author Follow-up] Request for Feedback on Rebuttal Responses

Dear Reviewer

Thank you again for your time and thoughtful feedback. As the discussion period approaches its close, we would greatly appreciate any thoughts on our responses and the additional evidence provided.

In particular, we would be grateful to know whether any of your concerns remain unresolved, and we would be happy to provide further clarification. If our responses have addressed your main concerns, we would also appreciate your reconsideration of the overall assessment.

Thank you for your time and consideration.

Best regards,

Authors

AuthorsOpenReview ↗

Request for Feedback Before Discussion Deadline

Dear Reviewer,

With less than 24 hours remaining before the discussion deadline, we would sincerely appreciate your feedback on our rebuttal responses and additional evidence.

We have made substantial efforts to address your concerns, and we hope to have the opportunity to clarify any remaining issues before the discussion closes. Your feedback at this stage would be particularly valuable, as it would allow us to address any unresolved concerns and provide further evidence that may help you reconsider your assessment.

If our responses have sufficiently addressed your main concerns, we would be very grateful if you could reconsider your overall rating before the deadline.

Thank you again for your time and consideration.

Best regards, Authors

Reviewer_CX1dOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper proposes DP-LAIR to protect both the fine tuning corpus SxS_x and the query time context SyS_y of an LLM served through an API. Each token comes from either a public path or a private path, where the residual between private and public hidden states passes through a clipped and noised low dimensional latent, and a router using only public features together with a runtime privacy budget decides which path is used. The router and decoder are retrained with DP, and on TOFU and MUSE News the method shows lower leakage and perplexity than composing DP-LoRA with InvisibleInk at the same maximum privacy budget.

Critical strengths

Strengths

  1. The paper addresses a practical deployment setting with a new design. In this setting the fine tuning corpus and the query time context both shape the same API response, and the paper protects them within one pipeline. The private path lets the residual between private and public hidden states reach the output only through a clipped and noised low dimensional latent, and a router on public features sends tokens that the public model can produce to a path with no privacy cost. This differs from runtime methods that clip the change in logits caused by each private record [1].

  2. The private path can support a standard DP guarantee. Because each latent coordinate is clipped to a fixed range, every private token has a cost that does not depend on the data, given by 2αC2tr(Σ−1)2\alpha C^2 \mathrm{tr}(\Sigma^{-1}) on the right of Equation (8). Charging this fixed amount for each accepted private token would give a guarantee that holds for every pair of neighboring contexts.

  3. The evaluation separates the two privacy risks from general utility. Section 5.1 queries training records without context to test the fine tuning corpus, provides private context to test the query time context, and uses a public QA task to check general utility. These tasks are run on four models and two datasets.

References:

[1] Vishnu Vinod, Krishna Pillutla, and Abhradeep Guha Thakurta. Invisibleink: High-utility and low-cost text generation with differential privacy. Advances in Neural Information Processing Systems, 38:67254–67298, 2026.

Critical weaknesses

Weaknesses

  1. (Solved) The privacy guarantee for the training corpus is incomplete. Section 3.3 states that DPDT protects only the Stage II use of SxS_x, which covers the router and decoder. The encoder and noise covariance are learned from SxS_x in Stage I without DP, and Section 5.1 and Figure 1 describe the private backbone as a model trained on SxS_x. Definition 3.1 treats these parts as fixed, and Lemma 3.2 assumes that SxS_x reaches the output only through the router and decoder. Each private path step therefore carries information from SxS_x that no budget accounts for.

  2. (Solved) The runtime accounting depends on the private context, so the runtime bound does not follow as stated. The paper charges ρt\rho_t computed from the actual context instead of the fixed bound in Equation (8). If the input were a neighbor Sy(−j)S_y^{(-j)}, the server would compute the charge around that neighbor, and the budget gate could make a different decision. Section 4.1 only compares the actual context with its own neighbors, while the filter of [1] requires each charge to depend only on earlier outputs. Since public steps release fixed logits and private steps add noise, a different gate decision changes the output by more than ρt\rho_t allows.

  3. (Solved) The joint protection experiment cannot separate utility from leakage. The question asks for the record placed in the private context, and Section 5.1 uses the answer sentence of that record both as the leakage reference and as the target for perplexity. DP-LAIR reaches lower perplexity than Public, for example 6.65 against 7.88 for Llama 3B on TOFU in Table 1, which shows that information from the record reaches the output. Leakage is scored by ROUGE-L over the full sentence, which in TOFU often restates the question, so a match on the extracted names would measure leakage more directly.

References:

[1] Vitaly Feldman and Tijana Zrnic. Individual privacy accounting via a renyi filter. Advances in Neural Information Processing Systems, 34:28080–28091, 2021.

Questions for authors
  1. Is the private backbone the LoRA model trained on SxS_x without DP? If so, how is SxS_x protected when its hidden states enter the private path at every query, and how is ρt\rho_t defined in the training data protection task of Section 5.3, where no private context is given?
6 comments
AuthorsOpenReview ↗

Response to Reviewer CX1d

W1: non-DP in Phase1

We would like to clarify that our paper explicitly defines Infer-RDP under the API-only threat model, as emphasized in the Introduction. The adversary observes only the released API transcript, without access to model parameters, hidden states, or internal components. Crucially, Infer-RDP is defined conditional on a shared and fixed deployment state ωf\omega_f, where Phase 1 outputs (e.g., the trained encoder and covariance) are treated as fixed internal components. As specified in Section 3.1 (p. 3, lines 157–159) and formalized in Lemma 3.2, the transcript distribution PSx,SyP_{S_x,S_y} is conditioned on this fixed state. This reflects practical API deployments, where internal components remain fixed and inaccessible to users A similar separation is adopted in Sampling and Aggregation [1], where intermediate computations need not be DP, while the released aggregate is protected by the DP mechanism. We provide a new Figure to highlight threat model difference in introduction in the revised version.

In Stage I, we non-privately train the encoder and covariance using SxS_x. These components, together with the private backbone, are then frozen before Stage II and treated as part of the fixed deployment state ωf\omega_f. Although they remain involved in inference, their state is shared across neighboring Stage-II executions. Importantly, this assumption does not require these components to be independent of their historical training data. In Stage II, we retain the fixed Stage-I components and retrain the router and decoder using the DP-guaranteed DPDT algorithm. Conditioned on the shared fixed state ωf\omega_f, this guarantee transfers to the released API transcript through post-processing, as established in Appendix A.1.

Therefore, Our training-data privacy guarantee protects the Stage-II use of SxS_x, conditional on the shared fixed Stage-I state, rather than its historical non-private use in Stage I. We will make this distinction more explicit in Definition 3.1 and the revised Introduction.

[1] Nissim K, Raskhodnikova S, Smith A. Smooth sensitivity and sampling in private data analysis[C]//Proceedings of the thirty-ninth annual ACM symposium on Theory of computing. 2007: 75-84.

W2: runtime privacy accounting

Thanks for your question, the routing strategy can be divided into two parts:

  1. The router we use, which is independent of the private prompt, to determine whether or not to enter the private path. Thus, the private prompt does not affect the routing decision.
  2. If we choose to enter the private path, we use equation (8) to find the worst case among all neighbors. The section 4.1 highlights that we evaluate privacy with respect to the records in its all adjacent context for each query, as the adversary observes only the transcript released for the current API call. Due to the maximum among all adjacent, each charge thus depends only on the previous token.

Importantly, the router and decoder are trained with DP guarantees, while the router's input is independent of the private prompt. Thus, this charge can depend on the previously released prefix, but it is evaluated over the same fixed query family throughout the privacy comparison. Thus, it depends only on the fixed family and the previous transcript, satisfying the predictability requirement of the adaptive Rényi privacy filter in [1]. We also provide detailed analysis in Appendix A.

[1] Vitaly Feldman and Tijana Zrnic. Individual privacy accounting via a renyi filter. Advances in Neural Information Processing Systems, 34:28080–28091, 2021.

AuthorsOpenReview ↗

Response to Reviewer CX1d(Cont.)

Q3: Experiment result explanation:

Evaluating DP methods using utility metrics such as accuracy and perplexity (PPL) is standard practice[1,2]. As for your example that the DP-LAIR reaches lower perplexity than Public, for example 6.65 against 7.88 for Llama 3B on TOFU in Table 1. DP-guaranteed algorithms aim to preserve common statistical knowledge while protecting individual data. In contrast, the Public baseline does not use any private data and therefore incurs zero privacy loss. DP-guaranteed algorithms allow non-zero privacy loss to retain useful information from private data, naturally achieving higher utility. Thus, the results for Llama 3B on TOFU (6.65 vs. 7.88 in Table 1) are reasonable. As an additional empirical audit of training-data exposure, we also report membership-inference results in Section 5.3 and Appendix D.

We also report the PII exact match result and LLM as a judge result as follow. For PII Exact Match, we extract PII from the private prompt and check whether it appears in the generated response. For LLM-as-a-Judge, we use GPT-6 to assess whether the response leaks private information.

PII exact match:

DatasetModelNo protectionPublicCompositionDP-LAIR
TOFULlama 3B74.5%0%0%0%
TOFULlama 8B95.0%0%0%0%
TOFUQwen 3B26.0%0%0%0%
TOFUQwen 7B35.0%0%0%0%
MUSE NewsLlama 3B98.5%0.5%1.0%3.0%
MUSE NewsLlama 8B99.5%1.0%0.5%5.0%
MUSE NewsQwen 3B91.0%1.0%1.5%1.5%
MUSE NewsQwen 7B94.0%1.0%1.0%5.0%

LLM-as-a-Judge

DatasetModelNo protectionPublicCompositionDP-LAIR
TOFULlama 3B88.5%2.5%4.5%6.0%
TOFULlama 8B97.0%7.0%4.0%4.0%
TOFUQwen 3B76.0%5.0%5.0%6.5%
TOFUQwen 7B78.0%3.5%5.0%6.5%
MUSE NewsLlama 3B97.5%13.0%10.0%12.5%
MUSE NewsLlama 8B98.0%19.5%14.0%20.0%
MUSE NewsQwen 3B96.0%9.0%7.0%8.5%
MUSE NewsQwen 7B96.0%9.5%7.0%10.0%

[1] Flemings J, Razaviyayn M, Annavaram M. Differentially private next-token prediction of large language models[C]//Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024: 4390-4404.

[2] Thareja R, Nakov P, Vepakomma P, et al. Dp-fusion: Token-level differentially private inference for large language models[C]//International Conference on Learning Representations. 2026, 2026: 151088-151120.

Response to Q1:

Yes. The private backbone is a non-DP LoRA model trained on SxS_x. Stage I learns the encoder and Σ\Sigma, Stage II freezes these components and DP-trains the router and decoder. The guarantee protects the Stage-II use of SxS_x under this fixed mapping, as explicitly stated in Section 3.3, lines 204–205. Within that comparison, API outputs are post-processing of the DP-trained interface. For more detailed explanation, please refer to answer on W1: non-DP in Phase1 part.

AuthorsOpenReview ↗

[Author Follow-up] Request for Feedback on Rebuttal Responses

Dear Reviewer

Thank you again for your time and thoughtful feedback. As the discussion period approaches its close, we would greatly appreciate any thoughts on our responses and the additional evidence provided.

In particular, we would be grateful to know whether any of your concerns remain unresolved, and we would be happy to provide further clarification. If our responses have addressed your main concerns, we would also appreciate your reconsideration of the overall assessment.

Thank you for your time and consideration.

Best regards,

Authors

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

The paper claims to provide a unified framework for protecting both training time and runtime private information while interacting with LLM inference api's.

Critical strengths

the claims seem to be wrong.

Critical weaknesses

What is the use of reporting average privacy loss? is it to emphasize better suitability for privacy budget accounting/composition? Privacy level s are to be generally reported from a worst-case standpoint.

Questions for authors

1.) How is infer-RDP different than RDP from the perspective of the privacy definition, and not from the perspective of the chosen application? If that cannot be substantiated, then it must just be called RDP.

  1. Are, θ_priv ​ ,ϕ_enc independent of Sx?

3.) What about Σ as well?

4.) what does downstream mean in "downstream training corpus"?

minor: this sentence does not read well and needs to be rephrased: The router and decoder are trained with differential privacy to protect training time data, while runtime privacy is adaptively accounted.

3 comments
AuthorsOpenReview ↗

Response to Reviewer qXeE

Thanks for your comments, we would like to answer your question one by one

1. Critical weakness for average privacy loss:

Thanks for your question. We employ a privacy filter [1] that allows generation to continue as long as the cumulative privacy loss remains within a user-specified budget, stopping when the budget would be exceeded. This allows users to adjust the privacy budget according to task requirements. Please refer to Section 4.2 for a more detailed explanation. Moreover, the rigorous privacy analysis is provided in Appendix A.

For comparison, we report results using the largest privacy budget, together with the average realized privacy loss. The latter highlights the variation in privacy costs across queries: some can operate with smaller budgets, while others may require larger ones. In practice, service providers can adjust different privacy budgets for different datasets.

[1] Vitaly Feldman and Tijana Zrnic. Individual privacy accounting via a renyi filter. Advances in Neural Information Processing Systems, 34:28080–28091, 2021.

2. Response to Q2

Thanks for your question. RDP is commonly used to characterize the Rényi divergence between output distributions of training mechanisms (e.g., trained model parameters). In contrast, our work focuses on the Rényi divergence between API output distributions under an API-only threat model. We therefore use the term Infer-RDP to distinguish this inference-time setting, where both the training dataset SxS_x and the prompt dataset SyS_y can contribute to the released API outputs.

2. Response to Q3 and Q4

θpriv\theta_{\mathrm{priv}}, ϕenc\phi_{\mathrm{enc}}, and Σ\Sigma depend on the SxS_x. Moreover, although θpriv\theta_{\mathrm{priv}}, ϕenc\phi_{\mathrm{enc}}, and Σ\Sigma depend on SxS_x and are trained without privacy guarantees in Phase I, they remain fixed in Phase II. The components interfacing with the adversary, namely the router ψc\psi_c and decoder ϕdec\phi_{\mathrm{dec}}, are trained with DP guarantees in Phase II, thereby providing privacy protection for the released outputs. Please see general response for more details.

3. Response to Q4:

The downstream training corpus is introduced in lines 35–36 and 45–46 and used for Stage-II training. Before Stage II, the private backbone θpriv\theta_{priv} has been adapted, and the encoder and covariance have been learned. These components are then frozen and treated as a fixed mapping within the API-only deployment. Stage II uses the downstream training corpus to train the router and decoder with DPDT, while keeping this mapping unchanged.

AuthorsOpenReview ↗

[Author Follow-up] Request for Feedback on Rebuttal Responses

Dear Reviewer

Thank you again for your time and thoughtful feedback. As the discussion period approaches its close, we would greatly appreciate any thoughts on our responses and the additional evidence provided.

In particular, we would be grateful to know whether any of your concerns remain unresolved, and we would be happy to provide further clarification. If our responses have addressed your main concerns, we would also appreciate your reconsideration of the overall assessment.

Thank you for your time and consideration.

Best regards,

Authors

AuthorsOpenReview ↗

Request for Feedback Before Discussion Deadline

Dear Reviewer,

With less than 24 hours remaining before the discussion deadline, we would sincerely appreciate your feedback on our rebuttal responses and additional evidence.

We have made substantial efforts to address your concerns, and we hope to have the opportunity to clarify any remaining issues before the discussion closes. Your feedback at this stage would be particularly valuable, as it would allow us to address any unresolved concerns and provide further evidence that may help you reconsider your assessment.

If our responses have sufficiently addressed your main concerns, we would be very grateful if you could reconsider your overall rating before the deadline.

Thank you again for your time and consideration.

Best regards, Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to Common Question

Thanks to all reviewers for their efforts in improving this paper. During the review process, we noticed that all reviewers raised a common concern regarding the privacy guarantee: How can DP-LAIR provide training-data privacy when the same corpus is used non-privately in Stage I and privately in Stage II, while the Stage-I components remain involved in inference?

We would like to clarify that our paper explicitly defines Infer-RDP under the API-only threat model, as emphasized in the Introduction. The adversary observes only the released API transcript, without access to model parameters, hidden states, or internal components. Crucially, Infer-RDP is defined conditional on a shared and fixed deployment state ωf\omega_f, where Phase 1 outputs (e.g., the trained encoder and covariance) are treated as fixed internal components. As specified in Section 3.1 (p. 3, lines 157–159) and formalized in Lemma 3.2, the transcript distribution PSx,SyP_{S_x,S_y} is conditioned on this fixed state. This reflects practical API deployments, where internal components remain fixed and inaccessible to users A similar threat model is adopted in Sampling and Aggregation [1], where intermediate computations need not be DP, while the released aggregate is protected by the DP mechanism. We provide a new Figure to highlight threat model difference in introduction in the revised version.

In Stage I, we non-privately train the encoder and covariance using SxS_x. These components, together with the private backbone, are then frozen before Stage II and treated as part of the fixed deployment state ωf\omega_f. Although they remain involved in inference, their state is shared across neighboring Stage-II executions. Importantly, this assumption does not require these components to be independent of their historical training data. In Stage II, we retain the fixed Stage-I components and retrain the router and decoder using the DP-guaranteed DPDT algorithm. Conditioned on the shared fixed state ωf\omega_f, this guarantee transfers to the released API transcript through post-processing, as established in Appendix A.1.

Therefore, Our training-data privacy guarantee protects the Stage-II use of SxS_x, conditional on the shared fixed Stage-I state, rather than its historical non-private use in Stage I. By explicitly accounting for the practical threat model, DP-LAIR designs a tailored DP mechanism that achieves a better privacy utility trade off.

We provide a new Figure in revised version to make this distinction more explicit, we also provide more explicit explanation in Definition 3.1 and the revised Introduction. We sincerely hope the reviewers evaluate our method in the context of the stated threat model. We welcome any further questions or concerns and would be happy to provide additional clarification.

[1] Nissim K, Raskhodnikova S, Smith A. Smooth sensitivity and sampling in private data analysis[C]//Proceedings of the thirty-ninth annual ACM symposium on Theory of computing. 2007: 75-84.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.