Differential Privacy
Robustness and Sensitivity of Public-Data-Assisted Differential Privacy Under Distribution Shift and Proxy Misspecification
Generated automatically from the limitations stated in 4 papers (AISTATS, ICML, NeurIPS), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Differentially private algorithms increasingly rely on auxiliary public data or pre-set structural parameters—such as covariance proxies, compact feature bounds, and distributional similarity—to mitigate noise. However, in realistic deployments, public datasets rarely match private target distributions exactly, and assuming tight compact bounds or positive-definite covariance proxies can fail. When these auxiliary assumptions degrade, algorithms either suffer severe utility loss or must spend substantial privacy budget to estimate these quantities from private data. Without a systematic evaluation of how public-assisted DP methods perform across varying degrees of proxy error, bound truncation, and distribution shift, practitioners cannot determine when auxiliary public data provides a net benefit over purely private baselines.
Why it matters
Practitioners can evaluate whether a candidate public dataset will improve or degrade private learning utility before committing privacy budget, and algorithm designers gain concrete empirical baselines for DP algorithms robust to auxiliary distribution shifts.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Benchmark existing public-data-assisted DP algorithms (e.g., private regression with covariance proxies, public-feature learning, and DP fine-tuning) across controlled synthetic and semi-synthetic distribution shifts (covariate shift, concept shift, label shift). Measure downstream task utility and privacy budget overhead across varying levels of similarity ($\gamma$) and shift magnitude.
- 2
Quantify the sensitivity of DP algorithms to covariance proxy misspecification (condition number, rank deficiency) and domain bound truncation (clipping threshold error vs. private bound estimation cost), measuring the exact point where auxiliary data assistance ceases to outperform standard DP-SGD or unassisted private baselines.
- 3
Construct a suite of diagnostic tests that estimate whether an available public auxiliary dataset is sufficiently aligned to improve private utility before allocating privacy budget to joint estimation.
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
If the community shifts entirely away from auxiliary public data assistance toward foundation-model zero-shot private adaptation, or if simple DP pre-processing heuristics (such as adaptive quantile clipping) already neutralize most sensitivity to distribution shift in practice.
Evidence
Each paper's own statement of the limitation, verbatim.
- Differentially Private Linear Regression and Synthetic Data Generation with Statistical GuaranteesAISTATS 2026
Requires prespecified compact bounds for covariates and responses, which may cause utility loss or require additional privacy budget if estimated privately.
- Decoupling Regularization and Privacy in Differentially Private Ridge Regression and ERMICML 2026
Requires a positive definite covariance proxy $\Sigma$ from public/auxiliary data or must expend additional privacy budget to estimate it.
- Private Zeroth-Order Optimization with Public DataNeurIPS 2025
Requires public data sufficiently similar to the private distribution; the theory's γ-similarity assumption and empirical gains degrade under distribution shifts between public and private data.
- Certified private data release for sparse Lipschitz functionsAISTATS 2024
The practical experiments require a public dataset, and utility stagnates once the public-data approximation error matches the noise level, so no further gains from more public data
Nearest existing work
- Effectively Using Public Data in Privacy Preserving Machine LearningICML 2023
- In-distribution Public Data Synthesis with Diffusion Models for Differentially Private Image ClassificationCVPR 2024
- Private Estimation with Public DataNeurIPS 2022
- Private Distribution Learning with Public Data: The View from Sample CompressionNeurIPS 2023
- Better Private Distribution Testing by Leveraging Unverified Auxiliary DataCOLT 2025
- Public Data-Assisted Mirror Descent for Private Model TrainingICML 2022
- Lower Bounds for Public-Private Learning under Distribution ShiftAISTATS 2026
- On the Benefits of Public Representations for Private Transfer Learning under Distribution ShiftNeurIPS 2024
- Learning-augmented private algorithms for multiple quantile releaseICML 2023
- Optimal Differentially Private Model Training with Public DataICML 2024
- DP-Mix: Mixup-based Data Augmentation for Differentially Private LearningNeurIPS 2023
- Differentially Private Domain Adaptation with Theoretical GuaranteesICML 2024
- Private Set Generation with Discriminative InformationNeurIPS 2022
- Bayesian Differential Privacy for Machine LearningICML 2020
- Improving the Privacy and Practicality of Objective Perturbation for Differentially Private Linear LearnersNeurIPS 2023